Despite mammoth advanced in artificial intelligence and computer vision, and a slew of humanoid robots either being piloted or hitting parts of the market, there’s still one major challenge that machines need to overcome.

It’s tactility, sensation, knowing how to hold an object. Picking up a strawberry, for example, a robot might not know how much pressure to apply, whether it’s secure, slipping, or about to be crushed. It’s something that we humans can feel through our skin, but it’s much harder to replicate on a machine level.

But it’s a work in progress, because there are companies out there like Touchlab who are working to give these skin-like characteristics to robots in a bid to make them more useful around the home, in care or in manufacturing. And it’s a lot more complex than we could ever imagine.

In Touchlab’s instance, it actually relies on quantum tunnelling technology to measure pressure, force and direction through a material that’s thinner than human skin, and with response times of less than a millisecond.

E-skin technology exists, but deployment is a challenge

Although the solution is laid out on the table, moving from pilot to production is a different ball game. Durability, reliability, affordability, adaptability… they’re all considerations that stand between theory and real-world use cases.

Touchlab is addressing some of these concerns with off-the-shelf components that already exist for established platforms, using bespoke engineering only where necessary.

It’s also seeking to reduce latency by bringing processing to the edge so that the robot only transmits the data necessary for the task, being that the e-skin can detect and collect huge amounts of micro-data.

With robots now rolling out across workplaces, care environments and, likely soon, homes, the ability to sense contact is becoming as important as the ability to see.

I speak with Touchlab’s team to understand how we can go from pilot to production in a realistic way.

• Why do robots need the sense of touch, and how is Touchlab trying to achieve this?

Computer vision allows a robot to predict, but touch is what confirms. Without tactile feedback, robotics remains trapped in a classic paradox: a machine can calculate complex equations in milliseconds, yet struggle to pick up a strawberry, handle soft medical instruments, or turn a valve without either crushing the object or dropping it. Without tactile sensing, a robot is functionally numb.

Touchlab solves this by developing biomimetic electronic skin (e-skin) that is actually thinner than human skin. Powered by quantum tunneling technology, our e-skin measures 3D forces, direction, and pressure in real-time. Crucially, it detects incipient slip—meaning the robot senses that an object is beginning to slip before it actually drops, enabling instantaneous, sub-millisecond grip adjustments just like the human nervous system.

• I tried the technology at Future Lab... I tried to control a robotic arm with haptic feedback and it felt a bit clunky with cables and restrictions. Why is it so hard to make it smoother?

Teleoperation is a classic engineering balancing act! First, there is a natural learning curve—controlling a high-degree-of-freedom robotic arm through a haptic glove isn't always intuitive on the very first attempt. It usually takes a bit of practice time and muscle memory to get the hang of the subtle force dynamics.

Second, running live demonstrations at high-traffic public events like Future Lab introduces specific hardware constraints. To ensure hundreds of visitors with completely different hand sizes can safely participate, setup requires universal-fit, heavy-duty tethered rigs and robust cabling.

Longer-term, our roadmap moves away from bulky, tethered operator suits entirely. By embedding our e-skin directly onto the robot and pairing it with onboard AI, the robot handles slip detection and dexterity autonomously—removing the need for an operator to wear restrictive physical gear.

• Right now, the focus is on your proprietary robot fingertip tech. What's the plan going forward? Do you plan to license it or sell it as a finished product? How will you deal with various robotic architectures?

Every robot platform is unique—whether it’s a humanoid deployed into hazardous industrial environments or a cobot assisting nurses in hospital wards. Touchlab is uniquely positioned to design, develop, and manufacture bespoke tactile devices tailored to specific OEM architectures.

At the same time, we offer scalable, off-the-shelf tactile components that are pre-integrated (both electronically and mechanically) with a wide variety of popular robot end-effectors. We provide volume-based discounts alongside strategic hardware loan and collaboration programs for robotics companies aligned with our vision. Over the past seven years, we have become extremely adept at rapidly building reliable integrations to suit diverse customer requirements.

• What are currently the biggest technological barriers Touchlab is encountering? What does the end game look like, and will we ever digitize our largest organ, the skin?

Our ultimate end game is for Touchlab’s tactile technology to become as ubiquitous as MEMS sensors— integrated into every commercial robot interacting with objects and humans. We envision full "finger-to-toe" tactile coverage powered by specialized edge computing and tactile orchestration software that mirrors the human somatosensory system.

Regarding barriers: the challenge isn't just technical; it's navigating western market adoption timelines for deeply transformative hardware. While digitizing human skin at scale presents manufacturing and durability hurdles, Touchlab can already deliver sensors matching or exceeding human-level spatial resolution, readout frequency, and sensitivity. The real task is packaging this capability into rugged, cost-effective formats ready for mass market deployment.

• How is Touchlab designing the user interface to reduce operator fatigue during long shifts?

We focus on shared autonomy. Rather than forcing a human operator to manually sense and adjust every micro-force vector, our e-skin handles low-level reflexes (like slip prevention) locally on the robot. The human operator directs higher-level tasks, while the robot's tactile system autonomously maintains optimal grip— dramatically reducing both physical strain and cognitive fatigue.

• High-density e-skin generates massive amounts of data. How do you handle bandwidth and network transfer without overloading systems?

High-density sensing does not automatically equal richer functional data, and often comes at the cost of hardware robustness. For our commercial products, we optimize spatial and temporal sensor density specifically for manipulation tasks, keeping data payloads lightweight.

For advanced R&D, we utilize distributed edge computing and are pioneering event-based asynchronous sampling in tactile sensing (similar to event-based vision cameras). Because our e-skin does not rely on heavy image-based data streams, it operates with extremely low latency and modest network bandwidth— making long-distance remote teleoperation surprisingly efficient.

• Because e-skin can identify objects and textures, what protocols protect tactile data security and privacy?

Tactile data on its own does not contain personally identifiable information. However, when integrated with video, audio, and LiDAR streams, multi-modal data security becomes critical.

For example, during our deployment with the Valkky medical robot in a Finnish hospital, all data transferred across hospital networks was strictly encrypted. Identifiable multi-modal data remained entirely on-premises on air-gapped storage devices and was fully anonymized before being processed for research or system analytics.

• E-skin is often seen as a high-end luxury. What is your roadmap for driving down production costs for commercial robotics?

Much like the early days of automobiles, groundbreaking technology often enters the market at a premium before scaling. While Touchlab is capable of producing full human-spec density today, deploying that level of complexity everywhere would make commercial robots cost-prohibitive.

Our strategy focuses on pragmatic engineering: delivering touch sensing tuned precisely to real-world task requirements at an accessible price point. We are actively collaborating with industry bodies to establish standardized performance metrics for tactile data, driving down manufacturing costs and making e-skin a standard baseline for all commercial robotics.

I've been intrigued by AI photo editors for a while now. Some work well, some not so well. Photoroom falls into the former camp. By and large, I found the AI tools to be very impressive and intuitive to use. There are some caveats to that which I'll get into, but overall, Photoroom delivered what it promises.

But what does it promise, exactly? Now, compared to a lot of AI photo editors, things here are a little different. This isn't a platform for touching up your holiday snaps or brightening up social media posts.

The emphasis here is on business use. Effectively, if you run an online store or similar, then you use Photoroom to transform staid, dull product shots into more enticing and engaging imagery that pops off the page.

Keen to see how it's developed since we last reviewed it three years ago - AI has come a long way since then - I tried out the platform to see what I could create.

Photoroom: Pricing & plans

• Online service via browser, iOS and Android app
• Aimed at businesses large and small
• Reasonably priced subscriptions and free trial

You can check out Photoroom by clicking here.

The platform is available online through a dedicated app for your iPhone or Android phone, or if you’re working on a computer, via your favourite web browser. Unlike many I've tested, you're not restricted to either Chrome or Edge here.

As with any online service, you need to subscribe to it to use it with a monthly cost billed annually. Photoroom comes in three tiers.

‘Pro', recommended for ‘solopreneurs’, will cost you $13 per month, billed annually. This level grants you 4,250 AI credits monthly, allows you to perform 1,000 exports, and comes with Photoroom’s standard set of tools.

The middle tier is ‘Max’ and is designed for small businesses. For $35 a month (again, billed annually - all tiers are), your credits are upped to 12,000, your exports to 3,000, and the advanced ‘Video Generator’ tool is now available.

Finally, there’s 'Ultra', for “growing brands”. You’re given 20,000 AI credits,10,000 exports, and can create videos and images in 4K, all for $100 a month.

If you’re looking for a more bespoke option, there’s a tailor-made Enterprise tier, with a tailor-made cost to match.

I recommend checking prices as Photoroom often offer discounts on those standard prices. Plus, you can try any of these tiers free for seven days, although don’t forget to cancel your trial before that should you decide not to carry on with the service and you’re needlessly charged a hefty sum for your forgetfulness.

Photoroom was generous enough to grant me access to their Ultra tier, so let’s see what it’s capable of, and what you can do with all those AI credits.

• Pricing & plans: 4 / 5

Photoroom: Interface

• Simple, well designed, and clear interface
• Easy ways to import images and work on them
• For an online service, I was impressed

The interface is simple and well organised. Once you’ve logged in to your account, you’ll find a sidebar to the left which contains various tools you need.

Thiss includes all the AI tools in your arsenal, templates, a direct link to your Shopify account if you have one, a batch tool so you can work on up to 250 images at once, a Brand Kit where you keep all your assets, and a Design section where your products’ photos are stored.

All the AI tools, which are essentially all your editing tools, are easy to use: you select one, upload a new photo or select one already in your account, apply the effect and you’re done.

There is one thing that I found surprising: Photoroom supports HEIC, the High Efficiency Image Container. It’s a highly compressed image format that’s been around since 2015 and has been increasingly prevalent since Apple made it the default image format for their iPhones in 2017.

The reason I make a point of mentioning this is because far too many services still do not support it, forcing any image or photo saved as an HEIC to be converted into either a JPEG or PNG prior to import. There is no such issue here, saving you time and frustration.

• Interface: 4.5 / 5

Photoroom: Tools

• On the whole, the AI tools do an absolutely excellent job
• Impressive ways to transform passable shots into something potential customers will love
• But... not all tools are created equal

Such a service succeeds on the strength of its tools, and Photoroom iss betting the farm on AI for this. But AI is a very generic term these days. Is it really artificial intelligence to remove an image’s background?

I’d put it more in the “complex algorithm does a very good job with what it’s given” category. Same with removing objects or people from a photograph and replacing it with an artificially generated background that blends with the existing one.

Such tools already existed before “AI” became an all-encompassing marketing brand. Or, taking inspiration from Monty Python's The Life of Brian, “it’s not AI, it’s a very clever tool”.

But whether it’s AI or not is less important than how these tools fare: I found the achieved results to be excellent. More often than not in apps like Photoshop or Pixelmator Pro, I end up having to touch up the output of similar tools, but here, it was essentially a one-click wonder.

Tools that really impressed me were definitely more AI based. Say you want to sell a shirt, which you’ve photographed on a surface, with as few wrinkles as you can manage… except you know how it is: such layouts are never really good enough.

Well fret no more: Photoroom has a tool that digitally irons your shirt for you, transforming your amateurish-looking photo into a meticulously smooth out product. Couple that with the background remover tool, and you’re well on your way to have a professional looking photograph.

But why stop there? Photoroom has a bunch of AI-generated models who can wear your outfit for you. There’s even what they call a ‘ghost mannequin’ to puff up your shirt so customers can see what it would look like on someone, without being distracted by an actual (or virtual) person wearing it.

I found all this pretty incredible, and that’s just scratching the surface. It’s not all sunshine and rainbows though: for instance, Photoroom has a Video Generator.

Select a photo, browse through all available templates, maybe even type in a prompt if you need, and the AI will get to work… this is where I encountered the familiar “AI tries way too hard yet fails miserably” syndrome.

Some templates removed my test products from the usual confines of physical reality, sometimes in a most disturbing way. Or they just didn’t work at all as expected. I trust the team working hard on this aspect of the service (what I was able to use bore no relation to their intro video about it, so things are moving fast in this arena), but for now, maybe stick with still images, where the results are truly impressive.

Photoroom: AI credits

• Set number of monthly credits, so use them wisely
• Many common AI tools don't cost any credits at all
• However, sometimes, the requested action won’t work, so try, try again

As you’d expect, most AI tools will cost you credits. A video generation goes for 500 (money not well spent for now), Exporting an image at 2K or 4K will cost you 135 and 240 credits respectively, adding an AI background to an image is worth 5 credits, and so forth.

Thankfully though, some tools you’ll use most often, like removing the background of a shot, retouching a photo or performing some image enhancements won’t cost you a thing (beyond your subscription).

You’ll just have to balance what you are given each month with what you hope to achieve in that time period, with the relief that when an action doesn’t work, you won’t have lost any credits - yes, sometimes an error occurs.

However, I found that if I request the same action again - and again - it finally does what I asked it to. Teething problems, I’m sure, and at least it does work impressively well eventually.

Should I buy Photoroom?

Buy it if...

You’re looking to sell a ton of product and want the images advertising your wares to look their best, using a service that is well designed, and is easy to use (when it works)

Don't buy it if...

You don’t want to encourage the proliferation of AI, know your way around Photoshop anyway, and can’t justify yet another subscription, no matter how great some effects are.

How I tested Photoroom

I tested this service on my trusty Apple MacBook Pro. Generally I do my best to check how software works on a variety of devices, both Macs and PCs and portable devices, when needed.

But as this was an online service, whose tools are dependent on servers the main bottleneck would be my fibre connection. I experienced no issues on that score.

Everything was extremely responsive. The free tools were nearly instantaneous, and depending on the complexity of what I was asking the AI to do, it took gradually longer.

Video creation took the longest amount of time, clocking in at around 4 minutes for me. Still, that didn't feel like a long time, when I got it working.

For more creative tools, we tested the best photo editors.

• Ukraine's new antenna weighs little enough for compact unmanned platforms
• BlueBird Tech CLOVER can serve drones, monitoring equipment, and ground robots
• The antenna covers frequencies widely used by modern FPV systems

Ukrainian defense technology firm BlueBird Tech has introduced a compact antenna for boosted battlefield communications systems.

The CLOVER 5.8 GHz weighs only 25 grams, yet operates across a broad frequency range spanning 4.9 to 6.1 GHz for reliable connectivity.

Its clover-shaped design gives the antenna circular polarization and an omnidirectional radiation pattern suited to fast-changing combat environments across contested zones.

A compact design with wide compatibility

The antenna is fitted with an SMA male connector and an RG402 cable, features that simplify its integration into existing platforms currently in use.

BlueBird Tech representatives told RBC-Ukraine that the CLOVER 5.8 GHz can be installed on FPV drones, UAV detection equipment, and radio-monitoring systems.

The company also noted that measurement devices and ground-based robotic systems can carry the antenna, not only aerial FPV platforms.

Its operating range overlaps closely with the 5.8 GHz band that most FPV drone operators already rely on for video links.

Its lightweight construction minimizes the additional burden on FPV drones, helping preserve flight characteristics during normal operational use.

Compatibility with major systems already on the market means operators may avoid redesigning their equipment before adopting the new component.

Whether such compatibility claims hold up under sustained, real-world combat conditions is something only continued field use can eventually confirm.

Ukraine expands its unmanned technology

The antenna arrives alongside other Ukrainian developments involving unmanned systems designed for military applications.

Ukrainian defense technology company WINFLY has introduced the WINFLY Spider, a reusable interceptor drone designed to capture aerial targets with a net before returning toward its launch location.

The Spider is based on the WINFLY 10 Hunter platform and reportedly reaches speeds of up to 160 km/h during operation.

Its stated operating ceiling reaches 5,000 meters, giving the system capabilities that differ substantially from the communications role associated with the CLOVER antenna.

The two systems therefore represent separate parts of Ukraine's growing unmanned technology sector, rather than components of one disclosed weapon system.

CLOVER is described as communications hardware, while Spider is designed specifically for intercepting aerial targets.

BlueBird Tech has not disclosed that the antenna itself contains electronic warfare, interception, encryption, or autonomous flight capabilities.

Its documented function remains transmitting and receiving radio signals across the specified frequency range, with integration depending on the host platform and communications equipment.

The antenna could therefore become useful within systems requiring lightweight radio hardware, although its actual battlefield effectiveness would depend heavily on the equipment and conditions surrounding its deployment.

Via United24Defense

Tech companies, including those in Silicon Valley and others outside of the set like Microsoft, have been through various guises through the years. Plenty of the old tech giants, like IBM and Toshiba, continue to enjoy success in the 21st century due to a chameleon-like quality that keeps them innovating and bringing in new revenue streams.

Evolve or die

The Microsoft CEO Satya Nadella wrote these words in his first email to company employees when he took the reins from Steve Ballmer.

In this opening salvo, he set out his vision that the next decade of computing would be dominated by the proliferation of devices and the rise of more accessible "intelligence" – hinting at the rise of AI in some form.

The company that he was to lead would embark on a strategy that would "harness the power of software" and deliver this through a network of connected devices for every organization and individual. The innovation that he referenced largely came in the form of a move away from a desktop-oriented mindset and one that was more embracing of the cloud.

Pivoting for AI

As the decade since this email wore on, Microsoft stayed true to that mantra. And, in the last few years, Nadella led the company through another significant shift in light of the rise of AI.

When ChatGPT burst onto the scene, Microsoft rushed to forge ties with the company and has aggressively added AI into its core products and services – from Bing to Windows.

While Azure continues to scale at a tremendous pace, surpassing $100 billion in revenue, the company has seen a mixed picture on the AI front. That's not due to a lack of demand, but to a shortage of physical space and energy to power these services and to expand capabilities in model training, deployment, and usage.

• USS Theodore Roosevelt is now the Navy’s first carrier ready for MQ-25 operations
• US Navy spent years developing infrastructure needed to operate unmanned aircraft aboard carriers
• Ronald Reagan is expected to join the MQ-25 carrier fleet in August 2026

The US Navy has completed a major step toward operating unmanned aircraft from carriers with dedicated facilities for controlling them.

The USS Theodore Roosevelt is now the service’s first carrier with a fully operational Unmanned Air Warfare Center supporting MQ-25 operations.

Another carrier, the USS Ronald Reagan, is expected to have its own center ready for operational deployment by August 2026.

Theodore Roosevelt leads the carrier integration effort

The Unmanned Air Warfare Center combines equipment required to control and manage unmanned aircraft operating from an aircraft carrier.

The Navy says its systems include the Unmanned Carrier Aviation Mission Control System and MD-5E Ground Control Station.

Those systems provide the command-and-control functions required for operating Boeing’s MQ-25A Stingray unmanned aerial refueling aircraft.

The USS George H.W. Bush became the first carrier to receive and test the center during August 2024.

However, Theodore Roosevelt is the first carrier where the capability has reached fully operational and deployable status.

According to a Pentagon acquisition document dated April 21, 2026, the Theodore Roosevelt “became the first fully operational/deployable MQ-25 capable aircraft carrier.”

The development gives the carrier force a dedicated infrastructure for unmanned aviation rather than treating drones as temporary additions.

That distinction matters because the MQ-25 requires specialized equipment alongside existing systems used for conventional carrier aviation.

The Navy expects the Ronald Reagan facility to become operational and deployable toward the end of August 2026.

MQ-25 moves closer to carrier deployment

The carrier infrastructure comes as the MQ-25 program advances after several years of development and testing.

In May 2026, the Stingray reached milestone C, allowing the program to move from development toward production and deployment.

The milestone followed nearly eight years of work since the Navy awarded Boeing an $805 million contract in 2018.

That contract covered four aircraft for design, development, fabrication, testing and delivery under the original program.

At that time, Navy estimates put the eventual cost near $13 billion for a planned fleet of 72 aircraft.

The program nevertheless experienced delays, including a first flight that occurred later than earlier schedules.

On April 25 2026, the Stingray completed an approximately two-hour flight near Mascoutah, Illinois, advancing its testing program.

The aircraft is intended primarily to provide aerial refueling, reducing the burden currently carried by F/A-18E/F Super Hornets.

That change could allow more Super Hornets to remain available for strike missions rather than supporting carrier refueling requirements.

The MQ-25 therefore depends on both the aircraft itself and carrier infrastructure capable of controlling unmanned operations.

With another carrier approaching readiness, the U.S. Navy is now moving from isolated testing toward a broader carrier-based unmanned aviation capability.

Via Defense News

In today's social media age, celebrity brand ambassadors are a common sight, helping endorse or promote a brand's products or tools, giving that extra bit of sparkle or star power.

This is especially true in the consumer market, but we're now increasingly seeing it in the B2B tech world, where what 10 years ago would have been niche businesses now have the power to hire top talent to promote their brand - I'm talking Idris Elba promoting ServiceNow, Keanu Reeves bigging up Palo Alto Networks, and Matthew McConaghey being the voice of Salesforce.

But as a massive Formula 1 fan, I was intrigued to read a press release announcing reigning world champion Lando Norris as a "brand ambassador" for a data company I'd never heard of - so who (or what) exactly are Starburst?

All about Starburst

Founded in 2017, and boasting that organizations in more than 60 countries rely on it to unify their data, "from startups to Fortune 500 enterprises" - Starburst's initial pitch is strong.

In its press release, Starburst describes itself as, "the enterprise intelligence platform for governed data and AI" - a pretty big claim, and one that the likes of Salesforce and Oracle might raise eyebrows at.

Continuing, Starburst says its platform will allow customers to, "run AI directly on distributed data in place, without moving or replatforming it, across cloud, on-premises, and hybrid environments while providing consistent business context to queries, models, or agents regardless of where it is stored or processed."

Overall, this should help companies develop and run faster analytics, "trusted" AI, and "better decisions at enterprise scale."

So what does this have to do with Formula 1? Over to Lando to explain.

“People often see motor racing at the highest level as being about reactions, but most of the performance comes from preparation and understanding the details," he says (in the same press release).

"I’m constantly asking questions, looking for information, and trying to build the clearest possible picture before making decisions. Whether that’s racing, running a business, or working with a team, having access to the right information at the right time gives you confidence to act. That’s what stood out to me about Starburst and why I’m excited to partner with them."

Not much clearer then - but it seems the two-year partnership will launch Starburst’s “Ask. Know. Go.” campaign, which the company says will be "connecting the speed, precision, and data-driven decision-making of Formula 1 racing with the enterprise challenge of turning distributed data into trusted action."

The three parts of this, unsurprisingly, are around getting access to data via AI agents, making sure this data is accurate and trusted, and supporting it with fast and powerful systems.

“Lando represents what happens when preparation, data, and split-second decision-making come together,” said Justin Borgman, Founder and Chief Executive Officer, Starburst.

“That's the challenge enterprises face with AI: ask the right questions, access the complete context behind the answer, and act before the opportunity passes. Starburst makes that possible by giving teams and AI agents governed access to data wherever it lives.”

So there you have it - another stitch in the fabric joining enterprises and Formula 1 - and one that will hopefully lead to a more successful second half of the 2026 season and beyond for Lando.

• OVHcloud will raise dedicated server prices from September 2026, with latest-generation Game range going up by as much as 87%
• Current-generation hardware will see increases of 51%, while older, 2024-gen hardware will see prices go up by 28% on average
• Move is linked to AI-related surge in memory and storage prices, with OVHcloud CEO stating the former has increased sixfold and could potentially double again by early 2027

French cloud platform OVHcloud is sounding the alarm for customers who use its dedicated gaming servers about a price increase that kicks in starting in September 2026 for server configurations, including its 'Game' offerings, that currently start at $357 on the US storefront.

In a post on X, CEO Octave Klaba highlighted the projected price increases as a percentage over prior commitments, stating that while 2024-built machines would not see a price increase, Gen 2026 would see a mammoth 87% increase in cost for consumers.

This marks the steepest increase across the board; otherwise, Gen 2026 servers saw a +51% increase in cost, even as OVHCloud states that existing options would see a lower overall increase in cost and that existing commitments would see "nothing change".

A sign of troubling times for private gaming servers?

Many gamers host private servers for their gaming needs, ranging from a small group of friends and family running a persistent Minecraft instance to servers that cater to hundreds, if not thousands, of players simultaneously, often with commercial caveats.

An increase in price affects both, though possibly the latter disproportionately, as monetization or donations generally sustain them, and they can already run into hundreds of dollars a month in upkeep.

OVHcloud's CEO has highlighted why the move, which seems to be focused on newer gaming servers, which saw the largest increase across the board, is necessary: its memory costs have increased sixfold and could rise to as much as twelvefold by early 2027.

The underlying reason is one that is easy to point at: AI server demand is not only dictating storage and memory market prices, but it could also be weighing purchase decisions at cloud providers, many who can envisage growing and consistent demand for such server instances, therefore raising the opportunity cost of other deployments; OVHcloud's 2026 'Game' servers for example, come with AMD's Ryzen X3D CPUs in two trims currently.

While these CPUs are great for games, they might not be cost-effective for datacenter needs due to higher per-core costs and lower clock speeds than comparable mainstream options, placing them a distant third in a list that has core-heavy server-grade CPUs at the top.

The writing hasn't been on the wall exactly, and it makes for an interesting comparison of how far numbers have traveled in less than a year: in November 2025, Klaba anticipated cloud product prices rising 5 to 10% between April and September 2026. In February 2026, announcing the year's first increase, he described an impact on cloud deployments between 2021 and 2025 of 2 to 6%, depending on hardware age.

He conceded at the time that asking existing customers to subsidize new ones was a little unfair, and argued it was the only way to keep cloud accessible over the following two years.

OVHcloud is an unusual case here: it builds its own servers, which is rare in an industry that often offloads that end to a third-party provider, and while that vertical integration has historically been the source of its price advantage, it has not protected the company here because OVHcloud still buys the same DIMMs and NVMe drives as everyone else.

The Register spoke to the founder of a mid-sized managed service provider who was entirely unsurprised by OVHcloud's announcement and expects AWS and Azure to follow.

The hyperscalers have deeper balance sheets, longer supply agreements, and more room to absorb component costs quietly, so any move from them will likely be slower and less legible than a founder posting percentages on X. It will, however, arguably not be smaller and will be considerably more far-reaching thanks to how deeply integrated some of their products are across the world.

• Volume and spend on tokens for Anthropic’s AI solutions made it more revenue than any of its competitors in July 2026, cementing its position as the top LLM provider
• This success is despite Anthropic’s tokens being more expensive than its competitors
• The data was collated by Vercel’s AI Gateway, an API service that routes LLMs for easy AI provider swapping

Monthly analysis of the AI industry has revealed a number of surprises for July 2026, not least the continued growth of Anthropic both in terms of revenue and token use.

This is despite the company charging a higher rate for tokens, AI’s computational data currency.

The data was compiled by agentic infrastructure company Vercel, whose AI Gateway is used by many major companies and developers, and found the price-per-token average across all analysed platforms fell following two months of growth in May and June. Despite the overall 37% increase in spend across the industry, the cost per token rate dropped.

Anthropic growth

Vercel’s analysis of transactions through its APIs has repeatedly demonstrated a dominance for Anthropic. Its reporting on gateway spend started in December 2025 and in every month since, Anthropic has had over 60% of the spend share, and in July that figure sat at 65.1% with 30% of the total volume (of all tokens sent through the Vercel AI Gateway).

Yet the average price per token was on 4.4 times higher than the overall average across all AI labs. Token prices have rarely proved to be steady, with Vercel’s data reporting an average price drop of 13.6% after a 20% rise in May and little change in June.

(For reference, the report’s definition of the price per token is the total spend divided by the total volume.)

Despite a higher price per token, however, Anthropic’s price per token was lower in July, a change attributed to the return to public access of Claude Fable 5 after its temporary suspension.

Monthly data collection

Vercel’s monthly look at the data that flows through its API paints a fascinating picture not just of Anthropic’s growth, but of the performance (or otherwise) of its competitors.

For example, it charts Chinese AI company DeepSeek’s surprise growth from May, through June; the report notes that “we said an open-weight lab would soon be second by volume. In July, DeepSeek surpassed Google to take that place.”

DeepSeek’s growth over Google has proved to be a surprise, one that puts it in second place on volume. The real story is Anthropic’s gradual cementing of its strong position over the rest of the LLM makers. Its larger revenue pull despite higher token price indicates that the recent iterations of Claude and anticipation over the gradual release of Mythos 5 are giving the company a strong reputation – much like Apple.

As many attendees of this year’s DEF CON hacker conference departed Las Vegas recently, many unsuspecting passengers on Delta Flight 591 attempted to access an onboard Wi-Fi network.

What they didn’t know was that ‘Delta WiFi Fast’ was actually a fake network, allegedly set up by a fellow passenger intended to mimic the actual onboard Wi-Fi network and scam other users.

The unknown passenger was able to disable the legitimate Wi-Fi networks for 30 minutes while they launched the attack, and in doing so, may have violated United States federal law.

How did the attack take place?

According to Aircraft Communications Addressing and Reporting System (ACARS) messages, the situation was first brought to light by the crew of the flight, who shared the following message:

“HEY ALERT CORP SECURITY WE HAVE A PAX [passenger] ON THAT HAS CREATED A SCAM WIFI CALLED DELTA WIFI FAST WE BELIEVE THEY ARE TRYING TO SCAM THE OTH PAX”

Another message read:

“NO INFO AS OF NOW WE HAVE A BUNCH OF PAX THAT WERE AT A CYBER CONFRENCE IN LAS THE WERE ABLE TO JAM OUR WIFI AND BRODCAST THERE SIGNIAL”

The actual details of what happened on the flight outside of these messages isn’t clear, but according to Monika Hathaway, head of press for DEF CON, similar attacks happened in Las Vegas: “Our conference this year also suffered from multiple similar ‘deauthorization’ Wi-Fi attacks and it impacted some of our operations.”

Delta airlines confirmed that no aircraft operating systems were affected and flight safety was never in question.

Wi-Fi deauthorization attacks can be launched with cheap, widely available ‘deauth boards’ which are small, battery powered devices that send deauthentication frames to devices within their range. On board a plane, these could easily reach most devices.

Once the legitimate Wi-Fi has been jammed and the other users booted from the network, the attacker can then set up an ‘evil twin’ network that users will attempt to connect to, which can be used to snoop on their internet traffic, steal credentials, and perform other malicious activities.

Expert perspectives on the Delta Wi-Fi attack

‘Evil twin’ attacks and the risks of connecting

Aras Nazarovas, Senior Information Security Researcher at Cybernews:

An evil twin attack is when hackers create fake Wi-Fi networks with the goal of stealing sensitive information from people, or exploiting known vulnerabilities present on victim devices. The fake networks often have a very similar (or identical) name to the legitimate network, which was the case here.

Once a person connects to the hacker’s Wi-Fi network, the hacker may be able to see what the victim is doing online and what data they transfer. However, since most websites have HTTPS/TLS encryption, much of what the user does, even on the rogue network, is private.

The risk here is that the hacker may attempt to redirect the victim to a phishing website – for instance, in this case, it may have been a fake Delta login page asking for personal data like name, email, address, etc. Or, the hacker may even go further and provide fake login pages for banks, social media, and try to extract login details from the victims.

Connecting to such a network comes with some risk in itself. Connecting to a network controlled by a threat actor allows them to probe your device for potential vulnerabilities and maliciously redirect your internet traffic to their own servers.

If a person entered credentials into a Wi-Fi login page, noticed security warnings popping up after visiting a website, downloaded something, or entered payment information into an unfamiliar page, then they may have had their data stolen. In that case, the victim should immediately change any passwords that were transmitted, do a thorough scan of their device for malware, and if bank details were transmitted, freeze the bank account until new credentials are received.

However, if a user just connected and disconnected to the Wi-Fi without entering any details or clicking suspicious links, they should be fine.

Who would launch the attack?

Seemant Sehgal, Founder & CEO, BreachLock:

Flying out of Vegas after Black Hat myself just a few days before this incident, I can tell you the security conference crowd that passes through that airport is unlike any other, and the crew on Flight 591 made the right call with the information they had in front of them.

Rogue access points impersonating a legitimate network are one of the oldest tricks in the book, and doing it on an aircraft to scam passengers is a federal crime regardless of the sophistication involved.

The people most likely to pull something like this on a DEF CON departure flight are the ones who know exactly where that line is, which makes crossing it a choice rather than a mistake. Disabling the Wi-Fi and investigating was exactly the right instinct.

Denis Calderone, CTO, Suzu Labs:

Hackers will hack. I go to DEF CON most years, and it's pretty common to have a terrible wifi experience on those flights because everyone is playing with their WiFi Pineapples and whatnot. That said, my flight home this year had no rogue SSIDs that I could see, and although, as usual, the wifi was shoddy, I never took the time to analyze the radio signals in the cabin, but if a few deauths were flying around, I wouldn't have been too surprised. It is concerning to hear about attempted credential harvesting on the flight though, and I feel that that's taking the expected hijinks way too far.

The deauthentication and evil twin combination used on Flight 591 is a well-documented attack that the security community has been demonstrating for a good two decades. These sorts of wifi threats are very common. DEF CON still displays their famed Wall of Sheep which displays the sniffing clear text credentials on the conference network, and every year the WiFi Pineapples have been selling out at the Hak5 booth.

But there's a significant difference between demonstrating a technique at a conference and deploying it against 199 unsuspecting passengers on a commercial aircraft. Last November, an Australian man was sentenced to seven years and four months in prison for running the exact same attack on domestic flights using a WiFi Pineapple and now the FBI is already involved in this case. There is definitely a legal exposure here.

For anyone who travels for work, in-flight WiFi should be treated as an untrusted network, period. The enterprise advice is encrypted DNS through your MDM and always-on VPN with captive portal remediation configured. But honestly, a VPN is something every traveler should be using, not just corporate road warriors. I make sure mine is on whenever I travel, and my family does the same.

Beyond that, if a WiFi network on a plane doesn't match what the crew announced or what's printed on the seat card, don't connect to it. If a network asks you to log in with your Google account or email credentials to get WiFi access, that's not how airline WiFi works. Airline captive portals ask for a credit card or a loyalty account, not your personal email password. If you're being asked for something that doesn't make sense for the context, you're probably not on the real network.

Reputational harm for the cybersecurity industry

Jacob Warner, Director of IT, Xcape, Inc.:

While a rogue Wi-Fi access point on a commercial airliner poses zero direct risk to air-gapped flight safety controls, it creates a serious enterprise security hazard for business travelers relying on inflight networks.

Dismissing an onboard network impersonation as a harmless prank ignores the reality of man-in-the-middle attacks, credential harvesting, and fake authentication portals targeting captive passengers connecting to the Internet.

Given that the flight departed Las Vegas immediately following DEF CON, it requires little imagination to conclude an attendee deployed the unauthorized access point.

This juvenile behavior is precisely why hackers suffer such a poor reputation among non-technical audiences and why security professionals struggle to build mainstream trust. Enterprise security teams must mandate always-on virtual private networks or zero-trust network access, disable automatic connections to open SSIDs on corporate endpoints, and instruct travelers to treat cabin wireless environments as untrusted networks.

Setting up an evil twin at 30,000 feet does not make you a clever researcher; it just proves why we cannot have nice things.

John Strand, Owner, Black Hills Information Security, Inc.:

This one hits differently because this is my community. These are my people. When security professionals engage in this kind of behavior, they’re betraying the very community they’re claim to represent.

There’s nothing impressive about it. It doesn’t make you look clever, and it certainly doesn’t make you an elite hacker. In most cases, these attacks aren’t even technically sophisticated. They’re simply people with enough technical knowledge taking advantage of others who don’t have the experience to recognize what’s happening. That isn’t skill. It’s bullying.

I hope the people responsible are held accountable. This isn’t funny, it isn’t clever, and it doesn’t demonstrate technical excellence. It’s just people abusing their knowledge to prey on those who are at a disadvantage. That’s not what this profession should stand for.

Networking never fails to surprise me when it comes to dependencies and its live nature – it’s a world that’s continually changing, and yes upgrades designed to improve performance or security are often the ones that cause outages.

Even carefully planned updates can have consequences that might initially be difficult to see during early pilots and test, due to a highly dependent mix of switches, firewalls, clouds and more – not to mention the challenges that come with interoperability in a world of mixed vendors.

Until now, it’s mostly been a firefighting exercise of responding to incidents and outages as they happen, admitting that they’re just an inevitable outcome of virtually any change.

Autonomous networking needs a safety-first design

Forward (formerly Forward Networking) wants to challenge this in its entirety, though, and the work it and companies like it are doing could fundamentally revolutionize network stability and uptime, significantly reducing costs.

Its solution is a mathematically accurate digital twin of the production environment, giving companies space to make changes and observe unintended impacts without actually affecting the real deal.

The principle isn’t anything new at all, and it’s been used in software development and other areas for decades, but where it could really shine is in an upcoming age of automated networking. Forward argues that providing a place for systems to autonomously test changes pre-deployment is a necessary stage to true automation.

I wanted to know whether technology like this could have prevented some major high-profile incidents lately, including those we’ve seen relating to Amazon, Cloudflare and more, and spoke to Forward’s co-founder and Chief AI Officer, Nikhil Handigol, on some of the technology’s biggest challenges, including how the digital twin could stay current and why pre-deployment proof could become essential in the world of autonomy.

• Forward proposes to prevent 100% of network outages caused by configuration errors. Are there any strings attached? And what does a “mathematically accurate digital twin of the entire production network” actually mean?

There are no strings, but let’s be precise on what that rests on. Forward Predict eliminates network risks through routing validation, end-to-end path analysis, firewall and ACL verification, segmentation and compliance verification, and automated regression testing.

The closest parallel is what software development did with continuous integration. No code goes to production without being tested first. Forward Predict brings that same discipline to networking.

Every proposed change is tested against the full production-equivalent model before it touches a live environment – routing validation, end-to-end path analysis, firewall and ACL verification, segmentation and compliance verification, and automated regression testing.

That model isn’t a sample or a simplification. It's built from the configuration and operational state of every device in the network — every vendor, every OS, every layer from L2 to L7, on-premises and in the cloud — using header space analysis to trace every possible path any packet can take.

The platform regularly collects state from the live network, and you can trigger a fresh snapshot on demand immediately before a push, so the "digital twin" you're testing against is always current. That continuous synchronization is what makes the math trustworthy: a proof is only as good as its inputs, and this is how we keep the inputs honest.

Run a change through that model and you get a deterministic outcome with evidence, not a guess. If it breaks something, if it opens a security hole, if it creates an unintended path, you see exactly why.

And because the answer comes from mathematical path analysis rather than a probabilistic risk score, it doesn't degrade as the network grows more complex.

Pair that with Forward AI and the platform does not just flag the problem, it iterates on its own until it lands on a change that is verified to deliver the intended behavior. Change windows that once took weeks now take minutes. There is no anxiety pushing a change live, because you already know how it ends. Weekends are for time off, not change windows and war-rooms.

• One point I wanted to raise is that you're essentially pegging it against a snapshot. Sysadmins will tell you that networks change. Firmware get updated, access points get refreshed, switches swapped and servers decommissioned. How do you make sure that Forward remains relevant throughout the lifecycle of the production network?

Forward Enterprise regularly collects configuration and state from every device in the estate, and you can trigger an on-demand snapshot immediately before pushing a change for absolute certainty.

When firmware gets patched, when an access point gets refreshed, when a switch is swapped, when a server is decommissioned, any change to the network is updated in the digital twin, it is a mathematically accurate model of the entire production network.

This is exactly why the underlying model holds up under scrutiny. Our engineering team validates the software against real behavior on live test hardware from every supported vendor, so the model's fidelity is proven before it ever reaches a customer network.

Then in production, that same discipline continues against the customer's own, constantly changing estate. Sysadmins are right that networks never sit still. Neither does the network digital twin. That is the only way proof stays proof instead of becoming a guess based on past behavior.

• Are there other companies that compete with Forward? How does your proposition differ from theirs?

The competitive threat isn't another vendor. It's a set of processes built over decades, in an era before a network could be modeled at all, that include change windows, war rooms, and the belief that the only place to truly test a network change is the production network itself.

Every other industry running mission critical infrastructure solved this problem years ago. Airlines train pilots on flight simulators instead of real aircraft. Pharmaceutical companies model outcomes before anything reaches a patient.

Manufacturing validates changes on a digital twin before touching the factory floor. Software development has run on continuous integration for two decades, where no code ships without being tested first.

Networking is the industry that never got its equivalent, not because nobody wanted one, but because building a mathematically accurate network digital twin is genuinely hard.

Against observability and monitoring vendors, the difference is which question gets answered. Those tools are essential for understanding what already happened on the network. Forward Predict answers a different question entirely, what will happen if this change is made, and it answers it with a deterministic proof rather than an educated guess.

Against other platforms using the words “digital twin,” the difference is a decade of foundational work building a model with header space analysis across all major vendors, every device, and every layer, rather than a partial view stitched together after the fact. That foundation cannot be a shortcut, and it is the part everyone else in the category is still missing.

• How do you see Forward evolving in terms of features and roadmap? Could this pre-emptive, predictive approach find a receptive audience in other sectors?

The roadmap moves in one direction, widening the aperture of what gets proven before it goes live. Today Forward Predict proves the outcome of changes to connectivity and security. Next is broader impact analysis.

We are not predicting what may happen, we are proving what a network will do, because a network's behavior can actually be modeled with mathematical certainty in a way human intent cannot.

That distinction is exactly why this belongs anywhere teams run complex, interconnected infrastructure with the same stakes, critical infrastructure operators running their own networks underneath everything else.

Wherever teams are changing complex systems without a way to prove the outcome first, the same problem exists, and the same fix applies.

Where this is headed fastest, though, is autonomous networking. As agents start proposing and executing changes at machine speed, a mistake stops being a single event and starts propagating at scale.

Proof is not an added feature there. It becomes the precondition for letting an agent operate at all.

The Creality Sermoon P1 is a handheld 3D scanner and follows similar use lines to the superb Sermoon S1 that I looked at earlier in the year. However, unlike the S1, it can be used in isolation: no wires, no need for a computer; everything needed from the scan to the display is held within one small, portable box.

This includes the initial capture, point-cloud fusion, mesh creation, editing, and multi-scan merging, all of which can be handled on the device, which is impressive considering the amount of data being captured. Inside is an eight-core Qualcomm processor with 24GB of RAM and 256GB of storage, and the scans appear on a six-inch Full HD touchscreen.

The fact that everything is contained makes a difference to the user compared with wired models. When I reviewed the Sermon S1 earlier this year, the biggest limitation I had was the Y-cable connecting the S1 to the computer, especially when I knocked it mid-scan. As the P1 essentially has an onboard computer and no wires, this isn’t an issue.

Of course, this does mean that the P1 is heavier at 830g compared to the S1's 508g, but shaped a little like an oversized game controller, the shape and extra weight really aren’t an issue when you have that freedom to walk around an object without watching out for the cable catching.

Compared with the Sermon S1, the P1 is a complete device with a scanner and screen that lets you see the scanning process without needing to be constantly connected to a computer or a power source.

There are essentially two scanning modes: blue laser with lines that sweep across the object to be scanned and require markers to be stuck to the object, and then IR for objects where markers can’t be stuck; the scanning quality is also a small step up from the S1, although through the test I’d say they were actually pretty much on a par.

Physically, the large game-controller design of the handset, with its two substantial grips, makes it easy to handle, and at 830g, it’s thankfully not to weighty. As the device is completely standalone, it’s powered by a swappable 3,300mAh battery, with the kit including two; each charges to 80% in 50 minutes.

As you scan, everything is shown on the touchscreen, which is bright and responsive. While the screen is glossy and does show reflections, they weren’t too much and remained easy to see throughout scanning.

Like the Sermon S1, the scanning technique was still the sticking point; it was a test. Again, preparation is important to give yourself a good chance of a decent scan, and pumping light into the screen is also key.

Using the blue lasers to scan, after the initial calibration, marker placement, lighting, and placing marker cubes around the object helped with getting a successful scan. So did elevating on transparent blocks, all helped with the success of the scan.

While having everything in the handset is a step forward, the most significant feature here, as with all scanners, is a slow, methodical approach.

Creality Sermoon P1: Price and Availability

The Sermoon P1 is available direct from Creality in the US for $3299, and in the UK, the Creality store has it in stock for £2799.

You can also pick up the 3D scanner along with bundled items such as an automatic turntable and scanning markers. Depending on your use-case and workflow, they're well worth checking out.

Creality Sermoon P1: Design

The P1 features a compact, game controller-like body with a grip on either side and a large touchscreen. The weight is actually pretty modest at 830g, and it feels solid enough with its rubberised plastic exterior, which suggests it was designed for use out in the field as well as in the studio.

The six-inch touchscreen is where things really differ from the other handheld 3D scanners I’ve used, as it lets you control object scanning and editing without a computer connection. However, a data connection is still available if needed.

Handling the device around an object is easier than with the wired 3D scanner, and the grips make it easy to handle and manoeuvre, although I still find that using a turntable is the best solution.

One of the issues with a wireless handheld device is that the batteries will inevitably drain. Thankfully, Creality has designed this so the battery can be swapped out, with two pinch clips embedded in it that make it relatively easy to remove when needed. More importantly, they lock firmly into place so there is no possibility of the battery coming out at any point.

The feature that I was pleased to see, despite this being a handheld device, is the 1/-inch thread on the base. This means it will fit a rig if you have one and will actually make the whole scanning process far less painful for many objects. I also found that using a turntable is a must for higher-quality scans.

• Design: 5/5

Creality Sermoon P1: Features

A big feature of this and several other Creality 3D scanners is the dual light source, which enables different scanning technologies to be used depending on what you’re scanning.

The blue lasers, which are my preferred method of scanning, handle precise geometry. At the same time, dual NIR VCSEL projectors cover larger or more organic subjects, and the scanner switches between them rather than just using one.

As with the S1, there are three blue-laser modes that you can use to cover different types of jobs. The most common object, and the one to start with, is the 22 crossed lines that enable you to scan-sweep large surfaces quickly. Then there are the 7 parallel lines that capture more detailed features. Finally, there’s the single line, which can be used to probe deep into scanning holes, gaps, and recesses. Creality quotes a depth-to-diameter ratio of up to 5:1, tested on a 29.08mm pipe at a depth of 152.31mm; in the test, this was enough to reach the bottom of the coffee cup without issue.

Two features change from the Creality Sermoon S1 I reviewed. The fast mode drops from 34 crossed lines to 22, while volumetric accuracy improves from 0.02mm + 0.08mm/m to 0.02mm + 0.06mm/m. Fewer lines with better-stated accuracy should mean that the wider field of view is doing the work

The fact that the P1 has been designed as a standalone device means that all 3D scanning and modelling is processed internally.

When I first got started with the P1, I had to run through a similar calibration process that takes between 5 and 10 minutes, and during the test I found that it was best to recalibrate at each new location.

In use, its small internal eight-core Qualcomm processor, Adreno 740 graphics with FPGA acceleration, 24GB of memory, and 256GB of storage enable capture, point-cloud fusion, mesh editing, and multi-scan merging, all running on the device.

If you do want to connect to a PC, you still can with three options: wireless or wired PC, with Wi-Fi 7 handling the wireless link.

Alongside the 3D scanning, surface texture is captured using a global-shutter RGB camera and 12 white LEDs, which together record surface colour.

If you have used 3D scanners before, you’ll be used to scanning spray. Here, Creality states that blue-laser modes capture dark and metallic surfaces without requiring a spray scan. However, some reflective surfaces will still benefit.

• Features: 4.5/5

Creality Sermoon P1: Performance

The Sermoon P1 arrives in a semi-hard case rather than the plastic waterproof cases that I’ve seen with previous Creality 3D scanners, which is a bit of a shame in some respects. However, it still keeps everything neatly packed in one case and, of course, offers protection during transport.

Before getting started, the 3D scanner needs to be calibrated. For some, this process will be fast and straightforward, while for others, adjusting the angle and moving the scanner up and down to align with the points on the 6-inch screen can be surprisingly challenging.

For the glass calibration chart, place it flat on the table and follow the on-screen instructions, such as aligning the markers by repositioning the handset. The first run of this takes some time and can be tricky; however, as I found during the test, regular recalibration is a good idea, and the more often you calibrate, the quicker it becomes.

Once calibration is complete, you can start scanning. I've found that starting with small objects is always the best place to begin, then gradually increase as your familiarity with the scanning process improves. I’ll also point out that 3D scanning is a skill, not something that you can just do out of the box. The first 10 attempts will more than likely be a disaster.

From experience, I know that a methodical approach works, and diving in headfirst almost always ends up in disaster. However, diving in headfirst is exactly where I decided to start, and with the first scan of whatever came to hand, I selected Infrared. I also just left it on the desk next to me and started the process.

Usually, this would be an instant fail; however, I was surprised that, on the first pass, the quality was usable, though not perfect. With a bit of mesh fixing and filling the hole in the bottom of the mesh, I’d managed quite a decent scan that would be perfectly printable on the K2.

I then noted at the bottom of the screen the “Colour Mapping”; clicking this, the texture from the mug was applied, which was even more impressive. What the scan and application of the mapping showed was that the light levels were a little low, so the graphics of the print looked duller than I would have liked - a good place to start, with the Original PRUSA INDX waiting in the wings for a full test.

I switched to the blue lasers, then grabbed the Chilli water bottle and covered it with markers. This didn’t go so well, and try as I might, I just couldn’t get a scan. This was also surprising as it’s usually the other way round.

However, this is where the methodical approach comes in: placing marker blocks around the bottle, elevating it onto transparent blocks, and placing it on a turntable with decent lighting suddenly showed what was actually possible.

I then selected a range of smaller objects and used the Infrared and Blue lasers to scan, switching between the line mode in the Blue Laser and the Object type in the Infrared.

The first point about the scanning process is that this is far easier to handle than the Sermoon S1 that I had looked at previously, and while it still takes some skill, I’m at about 50% after a few months' practice; this design definitely makes things easier, although still not quite perfect.

The screen's scan quality looked excellent, and what I really liked was the in-Studio workflow. Essentially, after scanning a few objects, I sat back in the studio, ready to download the models, which I expected would be a bit of a process.

However, using the provided USB cable, I plugged it into my MacBook Pro M1 Max, clicked through the usual authorisations, and loaded CrealityScan. When I plugged it in, I got the option to choose between Data Sharing and Scan Connection. You can then connect and share the files to the desktop software.

During the test on the Mac, the wired connection didn’t seem to work at all; however, it worked smoothly on the PC, whereas the Wi-Fi connection was straightforward, although you do need to switch your computer's Wi-Fi connection to the Wi-Fi hotspot of the Sermoon P1 even if both devices are connected to the same wireless network.

Checking over the scan, for the majority of objects, while there is some smoothing of finer textures, the detail is surprisingly good. What really stands out is the dimensions of the objects scanned; they remain one-to-one, so you get an exact scale match to whatever you're scanning.

I also really like the texture mapping; with the filament-mixing features of the latest multi-tool 3D printers, it suddenly becomes especially interesting. Although I haven’t as yet had much success, the potential here is massive.

Once again, the main issue I had was with the tracking, and despite having plenty of markers around and over the object, I had to be focused and methodical with the scans to get the results I expected. Again, practice is needed.

There were also times when the blue light scan would just lose the object or report not enough markers, even when things were covered.

What I really liked about the Yuse, however, was the fact that everything is in one place and in a far more portable form. The fact that you can take two scans of an object and blend them on-site quickly makes a massive difference.

Again, the frame rate and scanning techniques really come into play with many objects, and the auto-switching of scanning modes didn’t tend to deliver exactly what I expected; it was only my prior knowledge that enabled me to get the decent scans I wanted for this review.

What I will say is that, while Creality has worked on the user guide, it’s still a little clunky; a better getting-started guide is definitely needed, as the potential of these machines is now massive.

At the start of the test one of my big worries was that the internal processing was going to be slow, but actually from the scanning to editing the speed is actually pretty decent, there are some pauses in the scanning, which I get from the desktop version as well, but the process is quite smooth and this is a big step forward for the technology but also the ease of scanning.

• Performance: 4.5 / 5

Creality Sermoon P1: Final verdict

The Sermoon P1 is a step forward in 3D scanning technology, and after some time spent mainly practising the techniques, I am impressed by its capabilities. The initial scans showed a flip in the ease of scanning wth trhe different techniques. Infrared, which I usually fail to scan with any real success, worked well, whereas I had initial issues with the Fine blue-laser modes.

However, stepping back and being more methodical about the process again yielded the results I expected, and what impressed me was the speed of the onboard processing and the scanner's flexibility, which didn’t require a direct connection to a machine. This also made it easier to scan in various places.

As I looked through the scans, the quality can be excellent, but this really depends on you and the time you put into it. When I reviewed the Creality Sermoon S1, the verdict came down to a steep learning curve, and the P1 does circumvent some of the issues that I had with that scanner.

The cable and the laptop are essentially gone, unless you need them. When it comes to the calibration, marker placement, lighting and coverage, all remain much the same.

The Sermoon P1 really does make sense if you’re moving between the workshop and a client's site where setting up a tethered rig takes more time than the scanning itself. The purchase price once again keeps it out of hobbyists' hands, but as the technology advances, this will, of course, change.

As it is, if you put in the time and take things step by step, this has huge potential.

Should you buy the Creality Sermoon P1?

Buy it if...

Don't buy it if...

For more crafting essentials, I've also tested out the best 3D printers and best laser engravers

• RBI Governor Sanjay Malhotra tells banks that AI trained on alternative data such as cash flows, GST filings, and utility payments can extend credit to borrowers that manual underwriting cannot currently assess
• The pitch came bundled with caution, insisting that banks take safety and visibility measures including complete AI inventories, red-teaming before deployment, and the capacity to explain lending and fraud decisions
• The RBI insists, however, that accountability stays human; banks, not their algorithms, will be answerable to customers, auditors, and the Reserve Bank when AI-driven decisions go wrong

The governor of the Reserve Bank of India (RBI) has told the country's banks to put artificial intelligence at the heart of their lending patterns.

The approach is based on the assumption that models trained on unconventional data can identify and leverage data points that manual underwriters would struggle to spot or justify easily.

This would, as per RBI governor Sanjay Malhotra, both strengthen the credit market and enable financial inclusion by adding a new class of borrowers previously overlooked by conventional methods.

Cautious AI optimism from one of the world's largest central banks

The Reserve Bank of India is not a small central bank by any measure, and remarks by its governor therefore can often shape not only domestic but global markets.

His position on unlocking credit markets for users with little or no financial history by adding raw compute that considers other signals such as cash flow, GST tax filings, utility payments, and one's digital footprint considerably changes the landscape in a part of the world where banks are traditionally more conservative than their global peers when it comes to lending.

Malhotra also noted that this is exactly the data that a first-time borrower, a gig worker, or a small enterprise without formal books lacks, even as AI enables better monitoring of debtors' financials.

“Predictive models can identify borrowers on the cusp of default early enough to counsel rather than merely recover,” he added. “Used well, AI may be the most powerful accelerator to financial inclusion.”

It is important to point out Malhotra wanted AI to be used as an assistant here - not that AI should overrule loan officers on applications they have already rejected.

His claim is narrower and arguably more persuasive: entire categories of borrowers are effectively invisible to conventional credit assessment because the paperwork it requires does not exist, and machine learning, over other data, can make those borrowers assessable to banks and vice versa.

Malhotra also cited concerns that most modern AI models are black boxes and do not always explain their reasoning fully or adequately, and that borrowers are generally entitled to know why they were turned down. This could help keep potential algorithmic bias in check and rationalize such decisions at a time when AI models are known to be vulnerable to data poisoning attacks.

This would require a human touch, especially when it comes to accountability for such decisions, rather than the complete hands-off approach that AI enthusiasts sometimes suggest is inevitable.

To this end, the conclusion of his speech may be the most telling of what India's central bank feels about an increased AI footprint in the banking industry: Malhotra said that the winners of the AI era will not be the fastest or heaviest adopters but the institutions that best understand what they deploy, own its outcomes, and keep customer trust, something he equated to the enduring capital of Indian banking.

• Ukrainian drone forces have proved so successful against US troops in exercises that “destroyed” equipment had to be reintroduced into the simulated combat
• Over the course of the two-week-long exercise, US forces improved at camouflage, dispersing, and using electronic warfare against the aerial attacks
• The Combined Resolve military exercise took place in Germany earlier this year

US forces engaged in military exercises were stunned to be defeated quickly by Ukraine drone teams, so much so that the units were allowed to be re-added into the scenario.

Taking place in Germany earlier in 2026, the event was one of several exercises intended to simulate real drone warfare and develop responses and countermeasures to attacks. Ukraine's drone warfare has proved vital in its ongoing war with Russia, with its successes in the conflict defining the country as a pioneer of Unmanned Systems Forces and electronic warfare integration.

The US isn’t the only country to have engaged in wargames with Ukraine, with the UK, Estonia, and Sweden all recently outmatched

Recording a “kill”

Despite the US incorporating long-range drones into its arsenal during the War on Terror, it has had mixed results in defending against drone attacks, both in combat and in exercises. The simulated conflict with Ukraine took place in April and May 2026, and while the Fort Hood-based US forces may have come away from the event with some pride, the speed and decisiveness of the Ukraine drone forces clearly took them by surprise.

Rather than using live ammunition in these exercises, an observer is employed to determine whether a “kill” has been made. This is usually achieved by the drone controller hovering the craft above a vehicle or other combat equipment long enough for a potential kill to have been made.

In this particular incident, the Ukraine drones were recording kills so quickly that vehicles had to be sent back as new units, like a video game respawn.

However, as the exercises progressed, the US forces were able to improve their performance. Their responses to drone attacks relied on faster camouflage and concealing, effective dispersal, and employing electronic countermeasures to block the drones.

Ukrainian drones dominate

As part of the training, Ukrainian drone controllers also switched sides to assist the US troops in making effective offensive use of drones.

The Wall Street Journal reports that Air Force General Alexus Grynkewich, NATO’s top military leader, stated that “You can’t get training like this in the United States,” but observed a noted improvement in NATO forces after training: “The change is remarkable.”

It isn’t just the Americans that have struggled with Ukrainian drone warfare in simulated scenarios. Other events in 2026 have revealed shortcomings in handling drone attacks in exercises involving the UK, Estonia, some US troops, and other NATO forces. In addition, a military exercise involving Swedish-led NATO troops and Ukraine's drone warfare resulted in the NATO side being outmatched.

Humanoid robots are on the rise, with a slate of companies building several different prototypes. Whether they're designed to be on the production line or engineered to roam around your house performing various chores, different kinds of machines are on the cusp of mass production.

Setting expectations

Australian roboticist Brooks remarked on the appearance of robots in the first of his own three laws of robotics, inspired by Isaac Asimov, in a blog post that he self-published.

The co-founder of robotics company iRobot, which created the Roomba robot vacuum, Brooks explained that the anthropomorphization of humanoid robots – both by engineers and by consumers – could be counterintuitive.

While the form is necessary to gain pubic acceptance, engineers must deliver on performance and capabilities. Adopting a similar form to that of humans may imply a level of humanlike intelligence or competence in navigating the real world. Should robotics companies fail on that count, people will see straight through the ruse.

Robots of mass production

Brooks' laws focus on the competency and capabilities of robots and the need for creators to ensure that they are trained, that they're capable, and that they don't remove human agency.

Plenty of highly polished demos, recorded promotional videos, and on-stage appearances have done well to highlight the fact that these machines are not ready for prime time – despite impressive capabilities in highly controlled environments. That's evidenced by a series of 'robot fails' when they're introduced into real-world settings.

Companies, however, say they're improving the AI technology powering these machines, as well as the dexterity of appendages. They plan on ramping up production toward the end of this year, with the first machines set to be deployed in commercial settings by 2027.

• Quad-Level-Cell (QLC) 3D flash memory sets the industry benchmark for performance, bit density, and power efficiency
• It is the first QLC 3D flash to reach a 4.8 GB/s interface, ensuring super-fast data transfer between storage and processor
• This is the breakthrough that could deliver 1 PB (petabyte) storage to data centers

SanDisk has unveiled the chip that could revolutionize data centers and push cloud storage and AI to the next level with its QLC 3D flash memory technology, developed with Kioxia. Employing the companies’ CMOS directly Bonded to Array (CBA) technology, the 10th generation Quad-Level-Cell 3D flash memory has set a new industry standard, representing the baseline for competing memory manufacturers.

The chip has a 60% increase in bit density compared to the 8th generation 3D flash memory, also setting the benchmark for performance and power efficiency.

With a 37 Gb/mm² (gigabits per square millimeter) density, the QLC 3D flash can be used to assemble the first 1 petabyte (that’s 1,000 terabytes) solid-state drives for data center servers.

Wafer bonding

The CMOS directly Bonded to Array technology, in which CMOS wafers are bonded with cell array wafers after separate manufacturing, is at the heart of the QLC 3D flash memory development. Using “high-precision wafer-to-wafer alignment,” the new chips are the first to hit a 4.8 Gb/s interface (under test conditions) using the QLD 3D technology.

The die capacity and packaging density for a potential 1 PB SSD based on the QLC 3D flash memory are not practical (both in cost and space) for home systems but can be adapted for ultra-dense 1U and 2U rack servers in modern purpose-built data centers.

The size of the chips makes other uses impractical and requires advanced 3D packaging in 64- or 128- die stacks to hit the 1 PB SSD target, using the ruler-like, 318 mm long EDSFF E1.L and E3.L form factor PCBs, as well as the recently defined E2 form factor (measuring 200 mm x 76 mm). This was specifically designed to handle 1 PB drives.

Next-generation infrastructure

The march towards implementing AI across the connected world requires increasingly sophisticated storage so that agent-based, physical AI, and generative AI can be efficiently delivered. A chip that is dense enough to deliver a 1 PB SSD is key to this next step.

“By redefining the performance and efficiency envelope of QLC NAND, our 10th-generation QLC 3D flash memory delivers simultaneous gains in density, bandwidth, and energy efficiency and establishes a new paradigm for high‑capacity flash storage to provide a scalable foundation for next‑generation infrastructure applications," noted Sandisk’s Chief Technology Officer Alper Ilkbahar.

Kioxia’s CTO Hideshi Miyajima adds, “QLC technology enables the efficient storage of rapidly expanding data volumes, helping deliver greater performance and scalability for AI systems.”

For AI data centers and cloud providers, this is a long-awaited development, and places Sandisk and Kioxia ahead of the main competitor in this area, Samsung, which is approaching the challenge using more layers (1,000+ in ultra-tall stacks, compared with Sandisk/Kioxia’s 332-layers). Additionally, this announcement suggests that a 1 PB Sandisk/Kioxia SSD could ship in 2027, ahead of any forthcoming 1 PB Samsung SSD.

• iVANKY FusionDock Ultra packs 26 ports into a desktop workstation dock
• Four external 6K displays can run natively on compatible Apple Silicon Macs
• Up to 120Gbps bandwidth supports demanding display-focused professional workflows simultaneously

The iVANKY FusionDock Ultra is a 26-port desktop dock designed specifically for Apple Silicon Macs and demanding professional workstations.

The company says its dock supports native quad 6K monitors, high-speed connectivity, and charging capabilities for professional users.

Unlike many universal docking stations, this model does not support Windows PCs, Intel Macs, or Chromebooks, limiting its audience considerably.

Built around high-end Apple Silicon workstations

The dock uses a dual-chip architecture intended to keep multiple displays, storage devices, networking equipment, and peripherals operating simultaneously under sustained professional workloads.

It supports up to 80Gbps bidirectional bandwidth, while display-heavy configurations can reach 120Gbps through asymmetric data allocation across connected devices.

An included 80Gbps active cable is designed to support the connection speeds required by demanding storage and multi-display workflows during professional production sessions.

On compatible Apple Silicon systems, the dock can drive four external 6K displays without relying on DisplayLink or third-party display drivers.

However, the number of supported screens, resolutions, and refresh rates still depends on the specific Mac chip and connected monitors.

That native approach may appeal to creators who want additional displays without introducing another software layer into their macOS workstation.

Apple's hardware limitations are not bypassed, meaning buyers still need a compatible Mac capable of supporting their intended display configuration.

The dock also provides 10Gb Ethernet, giving users a high-speed wired connection for network storage, shared files, and other demanding workflows.

Its 26 ports include USB-C and USB-A connections, DisplayPort 2.1, HDMI 2.0, UHS-II card readers, and optical audio.

It features front and rear 3.5mm audio connections, allowing microphones, headphones, speakers, and other audio equipment to remain connected.

The magnetically articulated dual-cable connectors are designed to accommodate different Apple Silicon computers, including MacBook Pro, Mac mini, and Mac Studio.

Cooling and power for extended workloads

Power delivery is another major part of the dock, with up to 140W available for compatible MacBook Pro systems during demanding workloads.

A separate 45W USB-C power-delivery port can charge devices such as an iPad, iPhone, or another compatible accessory.

The cooling system combines passive heat dissipation with two adaptive fans that increase cooling activity as workloads become heavier.

iVANKY claims noise levels of 44 to 46 dBA at one centimeter under controlled conditions, with lower noise expected farther away.

That cooling design is intended to maintain stable connections and performance during extended sessions involving displays, storage, networking, and other peripherals.

At a promotional price of $649.99, reduced from an RRP of $749.99, the FusionDock Ultra enters an expensive docking market.

The company says the current price represents a 13% discount, although buyers should consider their actual requirements before paying for its extensive connectivity.

If you are not running an Apple Silicon Mac, this particular dock was never really built with you in mind.

Windows users, Intel Mac owners, and Chromebook owners are shut out entirely, regardless of how much they might want in.

• Study finds young people are increasingly pessimistic about AI
• Nearly half say AI will have a negative impact on their careers
• They're also suspicious of leading AI figures and executives

Younger Americans are becoming increasingly distrustful not only of AI, but also the executives and evangelists who promote the technology, a new study has claimed.

A CNBC/Generation Labs survey of 1,088 Americans between the ages of 18 and 34 found the vast majority saying they distrusted the leaders of major companies investing in AI, including the likes of Elon Musk, Sam Altman and Mark Zuckerberg.

The reasons behind this distrust were fairly clear to spot - with nearly half (45%) of respondents saying AI would have a negative impact on their careers, compared to just 10% who believed the technology will help them.

AI villains

The study covered a range of wider concerns young people may have around the future, with worries around the economy and jobs among the biggest pressures, however the threat of AI was seemingly ever-present, with the technology increasingly influencing all areas of everyday life.

This expands to the figures becoming the poster children for AI - respondents were given nine key people in the AI industry and asked whether they trusted each of them to act responsibly on AI moving forward.

Perhaps unsurprisingly, every figure was marked negatively - with Palantir CEO Alex Karp scoring the worst, as 81% of respondents saying they do not trust him. Karp was followed by Palantir Chairman Peter Thiel (79%), Alphabet CEO Sundar Pichai (75%) and Anthropic CEO Dario Amodei (75%).

Other notable figures, including OpenAI CEO Sam Altman, Meta CEO Mark Zuckerberg, Nvidia CEO Jensen Huang, and SpaceX CEO Elon Musk all scored about the same 70% figure - but there was 'better' news for Microsoft CEO Satya Nadella, who recorded the "highest" score, with just 65% of respondents saying they don’t trust him.

Looking at technology's place in the wider world, 40% of Americans said they believed “the federal government” should establish rules for AI, with just over a third (36%) believing “an independent expert body” should do so - and just 8% said they think AI “shouldn’t be regulated.”

Keeping with the spirit of protests across the nation, nearly two-thirds (60%) of Americans believed data center construction “must be slowed,” with just 15% saying they wanted to be “speeding it up.”

Looking forward, the when asked about how they would rate the American economy, nearly 80% said they had a negative sentiment - and when asked about the future of the economy, 50% believed it “will get worse,” compared with just 22% who said it “will get better.”

• Researchers uncovered “Download more RAM” flaw in consumer DDR4/DDR5 memory
• Attack bypassed Windows VBS and HVCI, disabling antivirus and protections
• Microsoft patched CVE‑2026‑23670; tools now help enable memory write protection

Microsoft has recently fixed a vulnerability that allowed threat actors to bypass advanced security measures, disable antivirus software, and expose the target device to full system takeover.

All of this, it seems, could have been possible with a very simple script, and a single click from the victim’s side.

Luckily, the vulnerability was discovered by white hat hackers, reported to Microsoft and remedied before falling into the wrong hands.

Download more RAM

During the 2026 USENIX Security Symposium, security researchers from the University of Birmingham and Durham University presented a discovery they called “Download more RAM”.

According to the researchers, some consumer memory chips (DDR4 and DDR5 DIMM) allowed software to alter the configuration reports it sends to the motherboard. In practice, it means that a threat actor could instruct the RAM chip to tell the computer it was bigger than it actually was, making the device “think” it has twice as much RAM memory as it actually has.

The researchers then established that this phantom extra memory can serve as an alias for real memory locations, granting them the ability to both read, and modify, memory allocations that should be under the processor’s, and Windows’ protection.

This window let them work around both Virtualization-based Security (VBS) and Hypervisor-Enforced Code Integrity (HVCI).

VBS, first introduced with Windows 10, is a security feature that uses hardware virtualization to isolate critical security functions from the OS, while HVCI uses virtualization to make sure only trusted and verified code can run in the Windows kernel.

The researchers also used the flaw to disable both antivirus and endpoint detection and response (EDR) software, re-introduce older, vulnerable drivers, compromise corporate systems under lockdown, and bypass kernel-level game anti-cheat protections.

The worst part is that this entire process can be chained together into a one-click script. In theory, if a victim is served this script as a file and they run it, they would trigger a chain of events that includes creating memory aliases, rebooting the computer, and disabling security protections.

"Our work exploits the fact that all processes share the same memory to bypass Windows' strongest security guarantees,” said Professor Tom Chothia, from the University of Birmingham. “Previous attacks of this kind needed a screwdriver and physical access to the machine. This one just needs a script. That changes who can carry it out and how far it can spread."

Who is vulnerable and how to stay safe

The attack surface is rather large, as well. The researchers analyzed the market and found three major manufacturers (Corsair, G.Skill, and ADATA) shipping at least one consumer memory product line in which the configuration chip was left entirely unprotected. Together, these vendors make up more than half (55%) of the high-performance consumer memory market and more than 70% of the gaming market (this doesn’t mean that 55% of the high-performance market is vulnerable - many devices are running other modules, too).

Modules from Crucial, Kingston and HyperX, and some G.Skill lines, were found to use partial write protection, but still enough to keep the device secure.

Before publishing their work, the researchers disclosed their findings to Microsoft, who quickly addressed it. The bug is now tracked as CVE-2026-23670, and is described on the National Vulnerability Database (NVD) as an “untrusted pointer dereference in Windows Virtualization-Based Security (VBS) Enclave” which “allows an authorized attacker to bypass a security feature locally.”

The bug was given a severity score of 5.7/10 (medium), and was fixed as part of the April 2026 Patch Tuesday cumulative update. Therefore, systems with Secure Boot running should be protected against this vulnerability.

Corsair added a feature to its iCue tools that lets users retroactively enable write protection on their memory modules. There is also a free tool called HWinfo with the same functionality, the researchers said, stressing that it mitigates the issue on non-Corsair models. Also, some motherboards offer a BIOS setting to block writes to memory configuration chips, which users can enable as an interim measure.

“The ‘Download More RAM’ attack demonstrates once more the importance of understanding systems, especially in terms of security guarantees. If a lower layer can become compromised, it puts the full system at risk,” said Dr Marius Muench, from the University of Birmingham.

“Windows makes a strong promise: that even an attacker with administrator rights can't touch the secure kernel. We found that promise rests on the assumption that your memory is telling the truth about itself - on a lot of the memory people actually buy, it doesn't have to."

• Huntress uncovers Shopify refund scam using fake orders and app notifications
• Attackers embed contact details in shipping addresses to trick victims into paying
• Users advised to ignore suspicious info, verify refunds, and report fake stores

Hackers are targeting businesses and individuals running Shopify stores with a highly sophisticated fake refund scam, experts have warned.

Security researchers at Huntress outlined how the fake refund scam works: first, a victim gets a notification that they received a refund. It could be for a returned iPhone, or a canceled service or order. The “refund” can be anywhere from a few hundred, to a few thousand dollars. Soon after, the scammers call (or mail) the victim, say they work at the company that gave the erroneous refund, and convince the victim to return the funds.

If the victim complies, they are actually sending their own money to the victims, since the “refund” part never happened.

Abusing Shopify's infrastructure

There are a couple of ways to pull this attack off: sometimes the scammers really make the initial transaction, but are able to cancel it and return the funds; in other scenarios, they create spoofed pages showing the transactions, tricking those slightly more gullible.

In most cases, fake refund scams can be spotted relatively easily, which is why they are not that popular nowadays. However, this new campaign comes with a sinister twist that will make even hardened veterans wince.

Huntress’ report notes the attackers start by creating a Shopify store of their own (or use a compromised one). The one the researchers observed was called “My Store” and was later deleted before it could be further scrutinized. Then, the attackers make a fake order themselves, setting their targets as the recipients using their phone numbers, or email addresses.

This type of information isn’t that difficult to come by these days. There are hundreds of email and phone number databases leaked on the dark web, which can be picked up for free (or for a handful of dollars). When they submit the purchase order, a notification appears in the victim’s Shop apps.

Yes, that’s right. In the Shopify app itself. Coming through Shopify infrastructure. As such, it can easily be confused for an authentic notification. It is even worse for users that enabled push notifications on their mobile phones, since they’ll see a notification with the Shopify logo, next to all the other notifications on their phone.

But the attackers still need to pull off the hardest part - getting the victim to “return” the money. In this case, instead of calling or messaging them, they leave their contact information in the shipping address, hoping victims would panic and reach out themselves.

In one shared example, the shipping address was listed as: Owen Nolan “2856 If You Didnt Place This Order Call Us at 1_888_690_3420-”, Albany NY United States 1_888_690_3420.”

For those treading carefully through the internet’s wastelands, the message included in the shipping address is an immediate red flag. Grammar mistakes, all letters capitalized, and a phone number completely out of place should be quite an obvious sign of an attempted fraud. However, since these kinds of scams bet on people being fast, reckless, overworked, and afraid, it might just work.

Huntress did not say if the scam made any meaningful impact among the Shopify’s community, or if it targeted a specific subgroup of users.

Defending against fake refund scams

To protect against such attacks, the researchers advise users never interact with phone numbers, email addresses, or links contained in an order that aren’t recognizable. They also advise users concerned about the security of their SHop account or personal data to contact support, and stress users should check their bank accounts to confirm whether they were actually changed. If they weren’t, they can report the order as “Not my order” in the shop app.

Finally, when purchasing from a store on Shop in general, users should check the store and its product reviews to learn about other customers’ experiences. If users are concerned that a product or store could be fake, it can easily be reported. Many of the shops in this scam were brand-new, with some using a "coming soon" description, as well.

The global stage of innovation is once again shifting, and where Silicon Valley once dominated over all, other equally powerful regions are rising up to challenge it.

Sadiq Khan recently spoke out on the dynamic politics in the US and claimed it was a key reason why tech talent and VC is being driven to the UK in competition against Silicon Valley.

As more investment and talent are drawn to the UK, the volume of R&D activity carried out here will only grow, and with it, the number of businesses turning to R&D tax relief to fund that innovation.

But under the watchful eyes of HMRC, whose growing scrutiny is felt by all across the industry, the need for robust claims is more important than ever.

R&D activity must be properly captured and evidenced, yet while the underlying innovation is real and substantive, the claims can run into trouble.

If the documentation was incomplete, the evidence was gathered retrospectively in a rush, and the finance and technical teams had never properly aligned on what needed to be captured or when, the claim is left operationally weak.

This is where strong, solid R&D activity is undermined by the operations and processes behind the claim.

The R&D disconnect

In my experience, the most common point of failure is rarely businesses’ technical knowledge. There’s often a disconnect between those individuals conducting R&D and the people preparing the claim. If we think about it, technical and engineering teams do not naturally think in the language of tax legislation – and why should they?

It’s not in their wheelhouse, and frankly, it isn’t a requirement of their day-to-day role. Equally, finance and tax functions often do not have the depth of technical understanding needed to translate innovation activity accurately into a compliant claim. This gap can leave the whole process vulnerable.

The challenge that businesses face is compounded by poor documentation and reactive evidence gathering, which remain some of the biggest operational weaknesses across the market. HMRC has been increasingly clear that it expects evidence captured at the time the R&D activity occurs, not reconstructed after the financial year has closed, as this leaves it open to incorrect recall and inaccuracies.

My advice to any business approaching its R&D claim in 2026 would be listen to HMRC, take it seriously and be proactive. Think about your R&D activities and how you are recording them throughout the year, not just at year end.

The pressure created by increased HMRC scrutiny

The R&D tax ecosystem has undergone substantial change in recent years, with the introduction of the merged scheme, changes to rates, new compliance requirements, and all under the shadow of HMRC’s growing scrutiny.

Its approach is far more rigorous, more targeted and it’s more likely to release enquiries into claims today than it was five years ago. The merged scheme is designed to bring greater consistency and clarity for businesses making claims, but in the short term, navigating HMRC’s expectations requires more robust operational processes.

The critical word is quality. For today’s claimants, this means ensuring claims are compliant with legislation and carry minimal risk of enquiry. It also means the quality of service, so making it as easy as possible for teams to provide the information needed for the claim, ensuring nothing is missed and maximizing the claim from a compliance perspective.

Those two things should work in tandem to successfully create a robust claim that doesn’t demand unnecessary workload from those involved and ultimately produces a better outcome.

Establishing the conditions to claim with confidence

As a means of making the claims process easier, businesses have turned to AI. There are well-documented, tangible benefits to deploying this technology across the claims process, but only when it’s tightly cornered off with guardrails. Yes, AI tools can make the compilation of claims faster and smoother, but when it’s used to write the narratives and technical descriptions, there’s a greater risk of inaccuracies, and it’s something that HMRC is cracking down on. The technology is only as good as the underlying data and the human judgement applied to it.

But we’re clearly heading in the right direction. HMRC’s new levels of scrutiny are ultimately there to set a better standard for R&D claims and filter out those with false claims. We exist in a volume-driven era, where the focus has previously been to build and submit claims quickly and at scale, rather than prioritizing quality. The UK has so much to offer when it comes to innovation, and the shifts taking place reflect a broader maturing of the market as we cement our place on the global stage.

Any business destined to become the foundation of the country’s global success is investing their own R&D processes, giving their innovation the backing it deserves to enable cyclical investment from R&D tax relief. We know the operational gap is where good claims break down, so closing it is where the real competitive differentiation now lies.

We list the best UK tax software.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

• Trump levies huge tariffs on 'sensitive' drones, such as those heavier than 25kg
• Tariffs have also been applied to a range of other countries, such as the EU, UK, Japan, and Taiwan
• Trump says the tariffs will help strengthen national security while also bolstering the domestic drone market

President Trump has levied a new set of tariffs against imported drones and their components for national security reasons.

For ‘particularly sensitive’ drones, such as those over 25kg, those with docking stations, or those with thermal imaging capabilities, the tariffs are 100%.

“President Trump’s drone tariff program will protect the national security of the United States and its defense and defense-adjacent industrial base, supporting and creating American jobs,” the White House fact sheet said.

Strengthening US industry through national security

Drones below the 25kg threshold will still be subject to a 25% tariff, with drones imported from the European Union, Japan, Liechtenstein, Republic of Korea, Switzerland, and Taiwan subject to a 15% tariff. The United Kingdom has a lesser tariff of only 10%.

The intended consequence of these tariffs is to strengthen the US domestic drone construction industry and the wider domestic drone market.

Outside of outright upping the cost of drones in the US as vendors pass on the costs of drones to consumers, the tariffs will also likely increase the costs of an already constrained supply of specialized drones in the US market.

Many drones imported from abroad, particularly those from China, are used for specialized purposes such as search and rescue and power line inspection. As the US does not have the dedicated industries and resources to quickly begin manufacturing these drones, there will likely be a huge increase of demand for a very limited supply.

The new tariffs also close a loophole allowing Chinese parts to be assembled into a drone in the US. The fact sheet states that drones for commercial and US military purposes “rely on foreign sources for critical Unmanned Aircraft System (UAS) components, which poses significant risks to U.S. national security and creates cybersecurity vulnerabilities.”

The tariffs are due to kick in on September 3 2026, giving the US drone market just 21 days to adjust or pay massively inflated import costs.

A small business owner receives a call from a customer.

An email that appeared to come from the business led the customer to a fraudulent website, and their personal information may have been compromised.

What makes situations like this so damaging is that the owner never knew the risk existed.

The domain used in the attack had been registered for a campaign years earlier and left quietly active, sitting outside anyone's management, until someone else found a use for it.

Situations like this rarely begin with a major security breach. More often they start with something small: a forgotten domain, an outdated email configuration, or a digital asset nobody realized was still active.

AI makes finding those unnoticed weaknesses all too easy for attackers. According to KnowBe4's 2025 Phishing Threat Trends Report, 82.6% of phishing emails now show some use of AI, a 53.5% increase year-over-year. Attackers are adopting the same automation businesses use to improve efficiency, and the speed advantage has shifted. The good news is that businesses can use the same advances in AI to identify many of these risks before attackers do.

But the speed of detection is only part of the challenge. The harder problem is visibility. You can't secure what you don't know you own.

More and more, small businesses manage domain names, websites, email systems, social channels, and third-party tools, and as those layers expand, the gaps between ownership and oversight become easier to miss.

Forgotten domains are a common example. Domains created for promotions, campaigns, or discontinued services often stay active long after their purpose disappears. Left unmanaged, they become blind spots that attackers exploit through phishing, impersonation, and brand abuse, and they can quietly erode credibility well before any breach, since a domain that no longer resolves correctly signals neglect to customers and machines alike.

Simply put, many business owners no longer have a complete view of the digital assets they own or the vulnerabilities that come with them.

Security needs to be embedded, not bolted on

Small business owners are focused on serving customers, growing revenue, and running their businesses. They are not thinking about DNS records, certificate renewals, or dormant subdomains during their day. Nor should they have to.

But many do not have the option. According to VikingCloud's 2026 research, 84% of SMB owners manage cybersecurity themselves, often without dedicated training or expertise.

The most effective security strategies are built into the infrastructure which businesses depend on every day, rather than added after problems arise. There are three layers where this matters most: the domain, which serves as a business' identity online; the website, where customers form opinions about credibility and trustworthiness; and email, which remains one of the most important channels for customer communication and one of the most common targets for impersonation and fraud.

Embedding security into the solutions businesses already use helps them maintain visibility and confidence without constant manual oversight of all of those moving parts.

Trust is now measured by both people and machines

Today’s cybersecurity conversation always comes back to trust. That’s what SMBs want to win by securing their online business. It is the engine of their success.

We know that trust is one of the most important competitive advantages a business can have: according to McKinsey reporting on digital trust, 40% of consumers have completely pulled their business from a company after discovering the organization was reckless with customer data, and 10% of consumers will cut ties with a brand immediately upon learning of a data breach, regardless of whether their own personal information was actually compromised or stolen.

For years, trust online was primarily a human judgment. Customers visited a website, received an email, or interacted with a brand, and decided whether it appeared credible. Today they still make those decisions, but they are no longer the only ones making them.

Search engines, AI assistants, and automated systems increasingly evaluate trust signals on behalf of users. Roughly two-thirds of Google searches now end without a click, according to Similarweb clickstream data analyzed by SparkToro. Trust is no longer just a customer's perception; it is becoming part of how businesses get discovered.

Credibility is no longer determined solely by what customers see. Domain resolution, certificate validity, email authentication records, and the consistency of businesses’ online presence all feed into the assessments that influence search rankings, AI-generated recommendations, and discovery across the platforms customers use every day. A business that doesn’t deliver on these fronts may be overlooked long before a customer ever decides whether to trust it.

Trust and security are now competitive infrastructure

For decades, businesses viewed security as a defensive function, meant to reduce risk and respond to threats. That perspective is changing.

Trust and security now influence customer acquisition, retention, reputation, and long-term growth. In the AI era, trust is no longer just a security outcome. It is a business strategy. As AI accelerates both innovation and risk, customers have become more selective about who they engage with and where they share their information. Credibility is difficult to earn and almost impossible to buy back once lost.

At Network Solutions, we have spent decades helping businesses establish and protect their digital identities. One lesson remains consistent: investing in trust early creates advantages competitors struggle to replicate.

Businesses that stand out in the years ahead won't simply adopt more AI. They'll build trust into every layer of their digital presence. The business owner who took that call from a customer deserved a better security infrastructure, not a better incident response after the fact.

Technology will continue to evolve, and so will the threats. Trust will only become more valuable. The businesses that thrive won't simply adopt more AI; they'll build stronger foundations for trust. That's where our industry needs to go next.

We've listed the best free website builders.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

The recent release of Anthropic Mythos is a wake-up call for the tech industry – and the fact that Anthropic themselves chose not to release it publicly speaks volumes about the level of risk we have now reached. AI agents have evolved from chatbots with upgraded capabilities to effective employees with database access, API keys, and system privileges.

However, the security protecting them is built on the same strategy that failed to stop ChatGPT jailbreaks in 2023. And this time, there’s no human to review an agent’s output, just an autonomous agent carrying out commands in a silo.

AI agents are reshaping enterprise systems and the way work gets done. Securing them requires an equally fundamental shift in thinking. Ultimately, now that agents act independently, resilience must be rooted in foundational controls, including hardware-level and lower-stack security, to be ready when the higher-level safeguards fail.

How AI agents expand the attack surface

Before agentic AI, the biggest AI risks were bad recommendations, inappropriate responses, and conversational data exposure. Human oversight acted as a safeguard for every action, and AI systems operated without direct access to sensitive information. The primary concern was reputational damage rather than risks to underlying infrastructure.

When Anthropic released the Model Context Protocol (MCP) in November 2024, it established a standardized framework that allows AI agents to connect to databases, file systems, and enterprise tools. But within eight months, a critical vulnerability emerged (CVE-2025-49596, CVSS9.4), triggering emergency security responses across the industry.

The risk came from four factors working together. Autonomy means agents can decide and act without human review. Privileged access gives them credentials, tokens and file system permissions. Machine-speed execution leaves little time for human intervention. And cross-system reach means one compromised agent can move across connected environments.

Together, these factors expanded the attack surface far beyond what traditional security controls – even AI-enabled ones – were built to manage.

Why software-only defences keep falling short

The industry is moving quickly to secure AI agents, but the response largely mirrors a familiar approach: adding more layers of software. Most companies are focusing on two main layers: input guardrails – implementing more software tools designed to stop malicious instructions from ever reaching AI agents, and permissions and monitoring – limiting what compromised agents can access.

It’s the same strategy the industry had relied on for decades: deploy quickly, remain agile, and address vulnerabilities as they emerge. Both methods operate inside the software trust boundary.

But history shows this approach often ends the same way: with the need for hardware-layer protections. In the 1990s and 2000s, network security responded to software exploits by deploying additional software layers. Breaches persisted until organizations eventually adopted hardware-enforced network segmentation.

The same pattern played out with endpoint security in the 2000s and 2010s. As malware evolved to bypass detection, the response was behavioral analysis, sandboxing, and endpoint detection and response. Yet more software. Breaches continued until TPM (Trust Platform Module) chips and hardware-enforced secure boot became widely adopted. Cloud security, in the 2010s and 2020s, followed a similar path.

A common lesson runs through each of these domains: when the software trust boundary is compromised, the hardware layer – where data actually lives – must be secured too.

The case for hardware-level security

This time, we cannot afford to learn slowly. Agents are already being connected to the systems that business rely on for their daily operations. Incidents like the MCP critical vulnerability and recent reports of a data leak caused by a Meta AI agent show how quickly the risks can become real.

Guardrails, permissions, and monitoring are necessary, but they are insufficient, and they represent the security layers that history shows will eventually be bypassed. Effective defense requires a third layer – one that exists beyond the software trust boundary and provides oversight at the hardware level, where sensitive data is ultimately stored and processed.

Hardware Root of Trust serves as the final security barrier, helping contain breaches before they escalate into a full system compromise. As the number of companies using AI agents continues to grow, security needs to move deeper than the application layer.

The industry has already learned that software alone cannot secure complex systems – it should not wait for a major compromise to learn the same lesson again.

We've featured the best endpoint protection software.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

In today’s supply chain, a week is not merely a long time. It can be the difference between protecting margins and absorbing avoidable cost; maintaining availability and losing sales; or preserving customer trust and explaining why another commitment has been missed.

Geopolitical tensions, tariff changes, economic uncertainty and supplier instability are no longer occasional interruptions to an otherwise predictable operating environment.

They increasingly overlap, interact and move faster than traditional planning cycles.

The external risk picture supports that conclusion. The World Economic Forum’s Global Risks Report 2026 identifies geoeconomic confrontation as the leading risk for both 2026 and the period to 2028. Half of the experts surveyed expect the global outlook over the next two years to be turbulent or stormy.

This is creating a decision-speed challenge at the heart of global commerce.

Research with supply chain leaders, found that only 20% can develop and deploy a response to a geopolitical disruption within 24 hours. A further 38% require more than a week.

During those seven days, transport costs can change, capacity can disappear, inventory can become stranded and competitors can secure alternative sources of supply. By the time a response has passed through every functional review and approval, the original assumptions may already be obsolete.

The critical question is not whether an organization can see disruption. It is whether it can convert that signal into an executable enterprise decision while there is still time to influence the outcome.

Visibility is not the same as readiness

For many years, supply chain transformation focused on improving forecasts, optimizing individual functions and reducing cost. Those disciplines remain important, but many of the operating models surrounding them were designed for a more predictable world.

Today’s disruptions cut across procurement, manufacturing, logistics, inventory, commercial planning and finance simultaneously. Yet many organizations still manage them through separate systems, functional metrics and sequential decisions.

Consider a tariff change. It may require the business to reassess sourcing, production, inventory deployment, pricing, customer allocation and margin exposure. If each function develops its own answer before the enterprise reconciles the trade-offs, valuable time is lost.

A dashboard may reveal the disruption quickly, but it does not determine which customers to prioritize, what financial exposure is acceptable or who has the authority to act.

Visibility without decision rights is simply faster awareness of the same problem.

Confidence is becoming polarized

The research also reveals a decline in overall confidence. Only 66% of leaders now consider their supply chains ready for the future, compared with 73% in 2025.

46% describe themselves as highly optimistic about their supply chains. These leaders are also more likely to report shared cross-functional KPIs, integrated data and stronger end-to-end visibility.

Optimism itself does not create better performance. Rather, confidence appears to reflect the capabilities these organizations have already built.

Less optimistic organizations are 2.3 times more likely to experience slow data sharing and integration, 2.5 times more likely to operate in functional silos and four times more likely to describe their supply chains as disjointed.

The divide is therefore not simply technological. It is organizational and operational. Connected businesses are better positioned to develop a common understanding of events, evaluate enterprise-wide consequences and align teams around one response.

From unified data to unified decisions

Unified data platforms are now the most widely adopted new technology in the survey, deployed by 51% of organizations. Adoption rises to 64% among the more optimistic cohort, compared with 40% among less optimistic leaders.

That foundation is essential, but it is only the beginning.

The real value emerges when common data supports connected decisions across planning and execution. This enables organizations not only to identify disruption, but also to assess its end-to-end impact, evaluate alternative responses and orchestrate action across the network.

Technology provides the foundation, but achieving decision speed also requires shared enterprise outcomes, clearly defined decision ownership, agreed intervention thresholds, pre-modelled scenarios and the authority to act.

European businesses are already adapting. European Commission survey evidence found that 27% of EU firms had changed, or planned to change, their strategies in response to tensions, disruption or policy changes in foreign markets. Among those adapting, 38% were changing sourcing or destination countries, while 22% were increasing inventory buffers.

Resilience is no longer an abstract ambition. Businesses are actively reconfiguring supply networks—but every adjustment carries consequences for cost, working capital, service and risk. The faster those trade-offs can be understood across the enterprise, the more choices leaders retain.

In grocery and consumer goods, the response window is narrower still because shelf life, availability and promotional commitments leave little room for delay.

AI can accelerate decisions - but foundations determine value

AI will play an increasingly important role in reducing the time between signal, insight and action. Machine learning and predictive AI are already used by 45% of respondents, with another 29% implementing them. Generative AI adoption has doubled from 12% to 24%, while agentic AI remains at an earlier stage, with 8% reporting live deployment.

AI can continuously monitor conditions, identify exceptions, evaluate scenarios and recommend responses at a scale human teams cannot match. When connected to planning and execution, it can also help translate those decisions into coordinated action across the network.

However, autonomy is not a shortcut around unresolved transformation challenges. AI operating across fragmented data, conflicting objectives and unclear decision rights may accelerate activity without improving outcomes.

Before giving AI agents greater scope to act, leaders must be confident in the data, guardrails, accountability and operating model surrounding those decisions. The question is not simply, “Can the agent act?” It is: “Should it act, within what boundaries and in pursuit of which enterprise outcome?”

The new measure of resilience

No organization can predict every disruption. A more practical measure of resilience is how quickly the organization can understand what has changed, determine what matters, make the necessary trade-offs and mobilize a coordinated response.

That is the real 24-hour test.

The organizations pulling ahead are not simply those with more information. They are those that can connect intelligence, decisions and execution quickly enough to act while choices still exist.

In an increasingly volatile world, resilience will not be determined by who sees the disruption first. It will be determined by who can make - and execute -the best decision before the window to respond closes.

We list the best data visualization tools.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

Agentic AI’s prolific use of tokens can create sizeable, unexpected costs for organizations. But saving on token costs without factoring in risk can be a fatal step.

As upfront prices for flagship artificial intelligence models continue to shrink, organizations have begun to wise up to the hidden costs they encounter with agentic AI models.

Specifically, the costs of tokens, which may look tiny when viewed as individual charges, can add up exponentially as AI agents become more active, leaving organizations with hefty AI expenditures they may not have anticipated.

This is putting CISOs in something of a bind. If they seek to save money on inference costs, primarily driven by token generation incurred by agentic AI, they may increase their security risk and accumulate hidden technical debt that puts their Agentic Development Lifecycle (ADLC) in jeopardy. It’s a problem that many CISOs may not have factored into their security budgets, but it cannot be left unaddressed.

The effectiveness of automated security processes is being impeded by fragmented pricing across the AI landscape, whether we’re talking about hyper-optimized nano models (essentially lightweight, yet powerful models built for a specific use, like Google’s Nano Banana 2 image generator) or premium reasoning engines, like Salesforce Atlas or OpenAI o3.

Organizations do have to keep a close eye on token costs to prevent them from spiraling, but CISOs also need to examine how agentic AI is affecting their security.

The hidden costs of AI agents

Erratic pricing has been a trademark of generative AI pretty much from the beginning.

About two years after OpenAI released ChatGPT, the Chinese company DeepSeek shook up the AI market with the release of a powerful, open-weighted large language model whose training parameters were publicly available, allowing users to customize the model and build on the cheap compared with other generative AI models.

ChatGPT-maker OpenAI and other AI companies started doing the same, and suddenly, the costs of using GenAI systems dropped off a cliff. In fact, prices fell faster for GenAI than for any other technology in history.

The emergence of agentic AI has introduced some stealth costs into the equation, however. The costs of agentic software range from free for open-source models to enterprise agents, with prices that vary from one-time fees (roughly $15,000 for basic models to more than $1 million for global enterprise models) to monthly subscriptions (which can range from a few thousand to $13,000 or more).

But those costs are fixed. Inference costs are another story: they scale with usage and can amount to 90% of AI lifecycle costs.

Tokens come into play when an AI agent requests processing from GenAI models, which charge agents for processing information. At a glance, the costs may appear inconsequential. Input tokens generally range from 15 cents to $5 per million requests. Output tokens, which require slightly more processing, cost from about 60 cents to $25 per million.

They may start small, but can add up in no time, thanks to AI agents that work very quickly, autonomously, and unpredictably. They are designed to interact with systems and other agents throughout the enterprise. A single action might generate scores of LLM calls. Token use, which has grown exponentially with the use of AI agents, has already increased IT budgets by about 20% according to recent estimates.

The accelerating cost of agentic AI is prompting CISOs to look for ways to save money where they can, and one way is to identify LLMs that charge the least per token. But what they may not be considering are the risk factors associated with those LLMs. If CISOs concern themselves only with the costs, they may open themselves up to security risks.

But better security doesn’t necessarily have to cost more. Depending on what they’re using agentic AI for, they may find they don’t always have to trade security for lower token costs.

Getting costs (and risks) under control

There are a few things organizations can do to help stop token costs from getting out of hand, including:

Match Agents and LLMs to the Job at Hand. Commodity AI systems can cost little or nothing, but they lack the deep reasoning for complex security synthesis. But not every application or function within the organization requires a reasoning engine. You can set up agents to work with low-cost LLMs on low-risk projects, while preserving higher-cost LLMs for critical tasks. It’s also worth being aware of which agents are likely to request more LLM calls.

Factor Risk Scores in Choosing Agents and LLMs. The security implications of using AI can’t be ignored. When developing a budget plan, include risk factors.

Monitor Workflows. Keeping a close watch on workflows can help you track costs and performance, allowing you to better understand which tools work best in which situations.

Lean on Human Oversight. Despite agentic AI’s autonomy, in fact, because of agentic AI’s autonomy, forgetting about the importance of the human element is risky business. Teams need thorough upskilling in secure development, with clearly defined ownership roles. And they must be given prominent oversight roles throughout the ADLC.

Agentic AI is fast becoming integral to enterprise operations, and organizations must control its associated costs. But a race to the bottom on token pricing creates hidden technical debt. Instead, CISOs need to weigh security performance when choosing AI tools as part of establishing an up-to-date security maturity model and an AI governance policy that emphasizes performance, costs, and risk management.

Only that approach allows agentic AI to be deployed without either breaking the budget or putting your entire organization at risk.

We've featured the best AI website builder.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

When my co-founders and I started building our platform, the prevailing industry consensus was that the future of AI belonged exclusively to a tiny handful of elite, hyper-capitalized technology labs. The dominant narrative insisted that massive centralized scale and closed proprietary control were the only viable paths to frontier capabilities.

Today, that multi-billion-dollar bet on proprietary infrastructure is facing a massive market disruption. From my perspective as a founder, the era of treating AI as a rented utility is rapidly drawing to a close, replaced by an urgent global demand for open-source independence and data sovereignty.

The first major driving force behind this change is the reality of corporate accounting. Being fully dependent on a cloud provider for the core infrastructure of cognitive capabilities has transformed from a convenient beginning into a huge liability in strategic and security terms. It simply does not make sense to remain in a permanent closed-door monopoly.

As shown by recent research carried out by a scholar at UC Berkeley, moving an enterprise project from a proprietary API to open-source cuts the cost of computation from $3,000 down to $31. Recent reports say that this economic revolution takes place all over the world due to the fact that the quality difference between open and closed systems has become non-existent.

Independent LMSYS leaderboard shows the leaders of open source solutions to be just two per cent behind the best proprietary systems, such as Claude Opus.

Sovereign AI

Beyond the obvious economic advantages, the call for open architecture has become highly geopolitical. Both within the government sector and the business sector alike, institutions are now realizing that there are real risks involved in being tied to a foreign company's products. We are starting to see how such friction points manifest themselves through governmental actions.

In Germany, the increasing conflicts between Bavaria and Microsoft regarding data protection issues and regulations illustrate precisely why businesses cannot afford to be tied into closed systems and rely on overseas hyperscalers.

This is why I am such an avid supporter of Sovereign AI, where one is able to have complete and absolute control of their own data, their own models, and their own jurisdictions instead of continually leasing it from a third party.

This fast deconstruction will result in a great wave of stranded investments due to the immense amounts of capital being channeled into centralized and monolithic data centers.

The trend of technology is no longer towards a reliance on hyper-scale. With open source becoming so much more efficient and smaller in its footprint, there is no longer any need to do everything within a few large-scale server farms.

Engineering advancements mean that highly specialized architectures can do all of the heavy lifting locally or in a distributed network of various hardware configurations. There will be no competition between monolithic centers meant for renting proprietary compute cycles and sovereign, local networks run by companies themselves.

Open source

It is no secret that I believe that attempts to tame AI by limiting access to closed models will always boomerang against the interests of the proprietary software developer. Once there is a looming possibility that they may lose access or encounter political restrictions on export, they simply get pushed into going for open models that they will own and control themselves in their jurisdiction.

That is the whole idea of the open source philosophy. The results of scientific work and technical progress cannot be confined to just a few private labs of major corporations. With an open model, everyone can adjust its parameters for their particular use cases.

Applications of open model architecture clearly demonstrate just how much open-model architectures can achieve in terms of adaptation to a particular localized purpose. Closed-model architectures will surely remain temporarily ahead in individual benchmarks, especially in terms of the most complex and experimental functions. However, most companies do not require a Porsche on each and every trip.

What they require is a highly advanced model that is completely optimized for a specific task, which they can always have access to without being dependent on the decisions of a couple of tech leaders.

Whereas AI has the potential to revolutionize all aspects of our collective existence as well as business operations in the world, such a process cannot be controlled by a handful of providers in any way that would be safe. The general market simply wouldn’t tolerate it. It would be too risky for any contemporary company to tie its strategic development path to a single provider.

That is exactly the reason why technology leaders are pouring resources into infrastructure that is neutral and open. We can’t just invest in software and hardware; we have to invest in sovereignty.

We've featured the best AI tool.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

• Namecheap suffers major outage, with services unavailable for hours
• "Cooling issues" at Phoenix data center blamed for outage
• The issue took around 15 hours to fully repair, with customers raging

Top hosting provider Namecheap suffered a significant outage which led to websites across the globe seeing problems for several hours.

The web hosting company saw issues after what Namecheap called "a failure of cooling systems" at its data center in Phoenix, leaving the company's hosting plans, EasyWP, DNS and email clients all offline.

The issue took around 15 hours to fully repair, with users also unable to access Namecheap services, email, accounts and more during that time.

Namecheap outage

Namecheap detailed the outage in a lengthy update post on its status page, keeping users informed throughout the issue.

Problems first surfaced around 12pm EDT, when two of the four chillers at the data center were brought back online, however it wasn't until 5.30pm EDT that "core databases" came back, and a further 40 minutes before Namecheap.com and Namecheap Live Chat returned.

Even the company's support helpdesk was affected, leaving staff unable to reply to live chat and emails, and many customers frustrated.

Full service was finally restored at 3.50am EDT, with Namecheap CEO Hillan Klein also looking to keep customers updated on progress via his X account, where he tweeted updates throughout the day.

"Once restoration is complete, we will conduct a thorough post-mortem of today’s outage—what led to it, where our existing safeguards and protocols failed, how we responded, and, most importantly, what we need to change to ensure an incident like this never happens again," Klein noted.

"We will be transparent about what we learn and will share the results of that review, along with the actions we are taking, with our customers as soon as it is ready."

Unsurprisingly, Namecheap users were less than pleased with the company's updates, calling for compensation and news on when their services would be reinstated.

"Failed is an understatement," one reply to Klein's message stated. "Sh*t happens but this was unacceptable. Businesses will fail because of today’s shambles. No way I’ll be staying after years of being a happy customer."

In its own update, the PhoenixNAP data center said it had experienced "elevated white space temperatures" as a result of damage from heavy storms overnight in the city.

"I, along with our whole team, are genuinely sorry for the outage today and for the impact it has had on so many of our customers, Klein added, "we failed you today. We know that trust is earned through our actions, not our words, and we will work every day to regain the trust we lost."

"For now, our entire focus remains on the customers who are still affected and will remain so, until all services are fully restored."

Each year, the Mobile Industry Awards shine a spotlight on the organizations pushing the sector forward. From operators and MVNOs to wholesale providers and technology innovators, the winners and shortlisted companies are recognized for delivering better experiences, stronger services and more innovative approaches to an increasingly competitive market.

What's striking, however, is that many of the qualities celebrated have become almost invisible to customers. In a world where connectivity has become increasingly commoditized, mobile providers are no longer defined by who has the fastest network or the cheapest tariff alone.

Instead, they reflect something far more difficult to see: the operational excellence that underpins every customer interaction.

That raises an interesting question: In 2026, what does "good" actually look like in mobile services?

For years, operators have measured success through familiar metrics. Network availability, coverage, latency and service level agreements remain critical measures of operational performance, but they only tell part of the story.

Today's customers don't experience mobile services through dashboards showing network uptime. They experience them through the outcomes those networks enable - whether they could activate an eSIM without friction before a flight, if changing their tariff took minutes rather than days, whether roaming worked automatically, or if resolving a billing query required multiple conversations with customer support.

The network may be performing flawlessly while the customer walks away frustrated. That's because the definition of "good enough" has fundamentally changed.

Beyond network performance

Reliable connectivity remains essential, but it has become the baseline rather than the differentiator. Across many mature markets, customers now expect consistently high levels of coverage and performance. As networks continue to improve, distinctions based purely on technical capability become harder to perceive. Instead, customer expectations have expanded to encompass every digital interaction surrounding the core mobile service.

Consider the lifecycle of a typical customer. They discover a provider online, complete digital identity verification, activate an eSIM, manage their account through an app, receive personalized offers, adjust their plan as their circumstances change, expect seamless roaming when abroad, contact support when necessary, and increasingly interact through automated digital channels.

Each of these moments contributes to their perception of service quality, but none of them is defined solely by signal strength. This is where traditional performance metrics begin to fall short. An operator can meet every contractual SLA while still delivering a fragmented customer experience if these touchpoints fail to work together consistently.

The industry's challenge, therefore, is no longer simply maintaining network performance, but delivering consistency across hundreds of smaller interactions that collectively shape customer trust.

Complexity has become telecom's defining challenge

Delivering these experiences has become considerably more difficult than many people realize. Today's mobile services are built on highly interconnected ecosystems. Behind what appears to be a simple customer interaction may sit cloud platforms, legacy operational systems, multiple software vendors, wholesale partnerships, application programming interfaces (APIs), AI-powered decision engines, digital engagement platforms, regulatory processes and cybersecurity controls.

Every additional capability creates new opportunities to innovate, but also introduces another layer of operational complexity. The challenge is not that these technologies exist - many have significantly improved what operators can deliver - it’s understanding how they interact.

A seemingly straightforward change to a customer proposition, for example, can have implications across billing systems, customer relationship management platforms, digital applications, compliance processes, provisioning workflows and partner integrations. Organizations are increasingly operating within environments where small changes can create unexpected consequences elsewhere.

It’s not that operators lack capability - they often possess sophisticated technology estates and highly skilled teams. What they often lack is clarity. Without clear visibility across increasingly interconnected systems, it becomes difficult to understand where customer friction originates, which operational processes are introducing delays, or how individual technology decisions affect the overall customer experience.

This is where micromoments come in. The small interactions that, when executed well, build trust and loyalty over time. For a customer, a micromoment might be receiving a relevant offer exactly when they need it, increasing their data allowance instantly while travelling, or having a billing anomaly resolved before it becomes a complaint. These micromoments are the true measure of service quality, and they are only possible when complexity is managed effectively behind the scenes.

The result is that complexity itself becomes the barrier to delivering simplicity for customers.

Complexity behind the scenes

The rapid evolution of cloud-native technologies, automation and AI has transformed what is possible within telecommunications. These innovations have reduced deployment times, accelerated development cycles and opened the door to entirely new business models. They have also encouraged a narrative that launching or modernizing mobile services is becoming increasingly straightforward.

There is truth in this: many technical barriers have undoubtedly been lowered. However, there is an important distinction between making technology easier to deploy and making services easier to operate. Complexity has not disappeared; it has just moved behind the scenes.

Customers rightly expect effortless digital experiences, but delivering those experiences requires sophisticated orchestration across systems that continue to grow in scale and diversity. The providers achieving the best customer outcomes are not those avoiding complexity altogether; they are those managing it effectively enough that customers never notice it exists.

This is where modern BSS and MVNE platforms have an increasingly important role to play. The objective is not to eliminate complexity, but to manage and abstract it so that customers experience simplicity, even when the underlying operational environment is becoming more complex. By creating better coordination between systems, processes and partners, operators can deliver more consistent experiences without exposing customers to the complexity behind the scenes.

Operational excellence is rarely visible when everything works as intended. But the lack of it becomes apparent when organizations struggle to introduce new products, resolve customer issues quickly, adapt to changing regulation or respond confidently to evolving market demands.

In many respects, successful mobile services are defined less by the absence of complexity than by an organization's ability to absorb it without passing it on to customers.

The new competitive advantage

As connectivity continues to mature, competition is shifting. Coverage and performance still matter, and price will always influence customer choice. Yet increasingly, these are becoming expected rather than exceptional.

Differentiation now comes from an organization's ability to evolve rapidly while maintaining a consistently high-quality customer experience. That means introducing new services faster, responding more effectively to market changes, integrating emerging technologies without disrupting existing operations, maintaining compliance across increasingly complex regulatory environments and giving customers confidence that every interaction will be straightforward.

In other words, operational agility is becoming just as important as network capability. It also changes how organizations should think about investment. Rather than asking only whether infrastructure can support future services, operators must increasingly ask whether their operational architecture allows them to deliver those services consistently across every customer touchpoint.

Competitive advantage is increasingly linked not to the size of a technology estate or the number of digital capabilities an organization has, but to how effectively those components work together to create a coherent customer experience. That requires visibility, governance and a deep understanding of complexity - not because complexity is valuable in itself, but because managing it well enables simplicity for customers.

We have seen this in practice with operators that have expanded their role beyond connectivity alone by embedding mobile services into broader ecosystems of everyday value, such as retail benefits, partnerships and lifestyle services.

Customers are more likely to engage with brands that are present in more moments of their daily lives. Differentiation comes from creating value that customers cannot easily find elsewhere and building stronger relationships beyond the mobile service itself.

Redefining what "good" looks like

The mobile industry has spent decades asking whether networks are good enough. Today, the more relevant question is whether the services built on top of those networks are.

As customer expectations continue to rise, quality can no longer be measured solely through traditional operational metrics. It must also be judged by how seamlessly organizations bring together the many systems, partners and processes that sit behind every interaction.

Customers are not consciously evaluating network architectures or operational models. They are judging the experience in front of them. They notice whether joining is easy, whether changes happen instantly, whether support feels connected, whether digital journeys are intuitive and whether the service simply works when they need it.

The providers that stand out in the years ahead will therefore be those that recognize that "good enough" is no longer defined by any single metric. It is defined by the countless moments in which operational complexity either reaches the customer or, through careful design and disciplined execution, remains completely invisible.

Working off site? We've listed the best rugged phones.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

As governments around the world seek to regulate artificial intelligence (AI), they are creating a patchwork of laws and policy frameworks that organizations operating across multiple jurisdictions must be aligned to – and understand.

In the European Union, for example, the AI Act – described as the “first-ever legal framework on AI” – takes a risk-based approach as part of a bid to “foster trustworthy AI in Europe”.

While in the US, the federal government has taken an altogether different stance. In March 2026, the White House published its National Policy Artificial Intelligence (AI) Framework, which, among other things, prioritizes deregulation in favor of innovation.

But the picture is becoming increasingly complex at state level. In July 2026, Illinois signed one of the country’s most comprehensive AI safety laws, requiring large AI developers to introduce transparency frameworks, independent third-party audits and formal risk mitigation measures. The legislation follows similar moves in California and New York, adding further momentum to a growing patchwork of state-level governance.

The divergence is not limited to these powerhouses on either side of the Atlantic. According to a recent briefing note published by the House of Commons Library, the UK “does not have any AI-specific regulation or legislation covering AI as a technology”.

Instead, it reports that AI is “regulated in the context in which it is used, through existing legal frameworks, such as financial services legislation”.

A global response to AI

Recognizing the current direction of travel, in July 2026, the United Nations convened its first Global Dialogue on AI Governance, arguing that international cooperation is essential in a world where AI systems, data and economic impacts routinely cross national borders.

“Artificial intelligence is reshaping economies, societies, and daily life,” it said. “Its opportunities are real. So are its risks.

“No country can address either alone. The AI Dialogue exists to ensure that governance reflects the priorities of all nations, not just the most technologically advanced and that the benefits of AI are shared by all,” it said.

So, while policymakers continue to debate the shape of future governance, it begs the question: how can enterprises continue to innovate and evolve when there is so much regulatory uncertainty?

Stop thinking about governance as compliance

For me, the answer is to take a more pragmatic approach. That means first understanding that governance is not the same as compliance. And second, that organizations need to stop treating it simply as a response to regulation.

Instead, business leaders need to understand that governance establishes a framework for who is accountable for AI, how decisions are made, how risks are managed and how systems are monitored over time.

While it’s true that specific regulations may differ from one jurisdiction to another, the fundamental principles that will deliver ‘good AI’ remain remarkably consistent.

Think about it for a moment. How often do AI projects begin with the same set of questions? Is this risky? Who owns it? Who signs it off? What data can we use? How will it be monitored? What happens if something goes wrong?

When you’re faced with a barrage of uncertainty, is it any wonder that projects stall even before they start?

But with the right governance in place, many of these questions have already been answered. Instead of repeatedly reinventing the wheel and debating the same issues, organizations can focus on the task at hand.

This is what good governance looks like. And it’s why it has the ability to remove friction, speed up decision-making and allow organizations to move faster.

The trust advantage

At an operational level, it gives teams an agreed starting point and set of rules. But the benefits extend beyond the organization itself.

Sharing these frameworks with customers, regulators, investors and other stakeholders is a sure sign that governance has been built in from the ground up rather than bolted on as an afterthought.

In doing so, organizations can show that they are taking accountability, oversight and risk seriously. And in an era where confidence in AI remains fragile, that trust can become a competitive advantage.

For anyone looking to implement governance, there are five key areas you need to establish.

Establish clear ownership: Every AI initiative should have a clearly defined owner – someone who is responsible not only for deployment but also for its ongoing governance.

Put oversight mechanisms in place: Organizations need processes to monitor, challenge and review AI systems, particularly as they become more autonomous.

Define your risk appetite: Not every AI application carries the same level of risk. That’s why organizations need to establish clear criteria for assessing use cases before deployment.

Create accountability from the outset: Build accountability from day one. Teams should know who makes decisions, who approves deployments and who owns the outcome.

Review and adapt continuously: AI governance is not a one-off exercise. Review your framework regularly as technologies, regulations and risks evolve.

While each of these has its place on the AI governance leaderboard, if I were pushed to say which is the most important, I would have to opt for clear ownership. Why? Because without clear accountability, governance risks becoming everyone's concern but nobody's responsibility. And when that happens, it can lead to paralysis.

As I said at the beginning, while policymakers may still be figuring out legislation, there is nothing to stop organizations from adopting good governance. In fact, regulatory uncertainty makes good governance more important, not less.

The organizations that establish the right structures, processes and accountability now will be best placed to adapt as the regulatory landscape continues to evolve.

We've featured the best AI chatbot for business.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

• GPT-4 predicted personality scores months before any human ever answered a question
• A Bosch oven manual was enough for GPT-4 to invent personality traits
• DSM-5 predictions hit 0.71 accuracy; astrology predictions somehow hit 0.85 accuracy

Human personality has traditionally been measured by asking people questions and comparing their answers across established psychological frameworks.

A new study published on iScience suggests GPT-4 can go further by estimating how people collectively respond to personality questions before answering begins.

Researchers Rotem Monsa, Aviv Zohar, and Shahar Arzy of the Hebrew University and Hadassah Medical School tested whether the model could infer patterns in human responses from language alone, rather than merely generating plausible questionnaire wording.

GPT-4 predicted responses before humans answered

The team used GPT-4 to create personality questionnaires from two very different textual sources.

One questionnaire drew on the DSM-5, the standard clinical manual used worldwide by psychiatrists to diagnose personality disorders with precision.

The other version pulled statements from an astrology textbook that assigns particular personality traits to each of the twelve zodiac signs.

"We wanted to choose texts that describe human personality in very rich detail but also sit on opposite ends of a spectrum in terms of scientific grounding," said Rotem Monsa, lead author of the study.

“The DSM-5 was refined through decades of clinical research and is the standard diagnostic manual in clinical psychiatry, known worldwide.”

Both questionnaires were then given to 600 participants alongside the Big Five Inventory, the most validated personality measurement tool available.

Before collecting those responses, GPT-4 estimated average scores for individual questions using a five-point Likert response scale.

This gave the researchers a way to compare predictions against subsequently collected evidence.

The model's estimates correlated strongly with actual participant averages, reaching 0.71 for DSM-derived questions and 0.85 for astrology-derived questions.

This decent internal consistency for DSM-5-based questionnaires suggests that traits that typically appear together in real life also correlated in responses.

The astrology questionnaire, by comparison, showed a noticeably weaker internal consistency across most of the personality traits it measured overall.

GPT-4 also estimated relationships between individual questions, producing correlations of 0.74 and 0.69 for DSM and astrology questionnaires, respectively.

The researchers therefore argue that GPT-4 appears capable of capturing population-level response tendencies embedded within language and questionnaire content.

However, they specifically caution that predicting average responses does not establish that the resulting questionnaire accurately measures individual personality differences.

Testing the limits with an oven manual

The researchers also tested whether GPT-4 would generate personality-style questions when given material containing no explicit personality information.

They supplied the model with a Bosch oven manual and a literary description of a landscape from The Lord of the Rings.

GPT-4 still produced questionnaire items resembling personality assessments, even though neither source was designed around psychological traits.

That result suggests the model has learned strong associations between language patterns, personality descriptions, and familiar questionnaire structures.

Yet the researchers found that unrelated source material could produce items that were semantically weak, narrow, or poorly connected.

This distinction matters because a convincing question does not necessarily measure the psychological characteristic that researchers intended to examine.

The study also found important differences between the questionnaires when researchers examined their underlying statistical structures.

The Big Five questionnaire most closely matched its expected organization, while the DSM and astrology questionnaires showed weaker confirmatory factor-model fits.

Astrology-derived items also failed to preserve their intended elemental structure, instead forming broader patterns resembling established personality dimensions.

The results nonetheless showed comparable predictive performance across many life outcomes, although astrology-based measures performed worse for mental-health outcomes.

For now, the evidence supports a narrower claim: GPT-4 can anticipate aggregate human response patterns from textual information with measurable accuracy.

“Given that personality traits are reflected in language, LLMs may have learned the structure of human personality as a natural byproduct of their training,” Monsa added.

“So, while they were not taught specifically psychology or personality theories, these are already embedded in the language that LLMs learn from.”

• QZSS was designed specifically around Japan's difficult positioning environment and geography
• The seven-satellite constellation brings Japan closer to continuous domestic navigation coverage
• QZS-7 successfully reached orbit after QZS-5 was lost during launch

Mountainous terrain and densely packed city blocks across Japan can obstruct ordinary satellite positioning signals from reaching receivers on the ground.

Standard Global Positioning System coverage was never built with those obstacles in mind, leaving gaps in accuracy across parts of the country.

Japan has spent more than a decade building its own answer to that gap, and this week that effort reached a new milestone.

Why Japan built its own system

Japanese authorities developed the Quasi-Zenith Satellite System, a constellation designed to guarantee near-constant coverage and sharper location accuracy across the archipelago.

Japan's Cabinet Office states that the system complements GPS by ensuring a navigation satellite remains almost always somewhere above the country.

Japan's islands span latitudes between 20° and 45° North, a spread too wide for one geostationary satellite to cover alone.

The QZSS satellites therefore fly an inclined geostationary orbit that traces a figure-eight pattern when viewed from the ground.

The newest satellite in that constellation, known as Michibiki No. 7 or QZS-7, reached its intended orbit early Tuesday local time aboard an H3 rocket.

This orbit causes satellites to move more slowly while above the Northern Hemisphere, spending roughly 13 hours near the equator each cycle.

The first configuration used four satellites, and one of them appeared over Japan every eight hours during that early arrangement.

One satellite remained at an elevation of 70° or higher at all times, though not always at the exact zenith overhead.

That detail about elevation, rather than reaching the true zenith overhead, is actually what gave the constellation its distinctive quasi-zenith name.

A troubled rocket finally delivers

The H3 rocket, which carried the QZS-7, is the Japan Aerospace Exploration Agency's (JAXA) newest mid-range launcher, and Tuesday's flight was the rocket's ninth mission since its debut.

Two of those nine missions failed the rocket's short operational history so far.

The most recent failure occurred in December 2025, when a glue-related fault destroyed the navigation satellite payload it was carrying.

That lost satellite was QZS-5, an earlier orbiter meant to expand Japan's navigation constellation before its launch attempt failed.

The first experimental QZSS satellite went into space in 2010, followed by four production satellites launched between 2017 and 2021.

Japan's government later decided to expand the constellation further, growing the planned fleet from four satellites to seven satellites total.

QZS-6 launched successfully in February 2025, arriving in orbit months before the failed December attempt to launch QZS-5 occurred.

Officials say three more satellites are planned sometime during the 2030s, though no direct replacement for QZS-5 has yet been confirmed.

Tuesday's success suggests Japan now holds nearly all of the satellite navigation coverage it originally set out to build.

It also offers JAXA a measure of reassurance that the H3 rocket, despite its rocky record, can still serve as a dependable workhorse for future missions.

Via The Register

The computing industry was once unrecognizable from the one we see today, with the idea of discipline centered around writing code treated as some kind of fantasy. That was before industry pioneer Margaret Hamilton coined the term while leading the flight software team for NASA's Apollo space program in the 1960s.

Pioneering an industry

Hamilton was describing the process by which she coined the term "software engineering" and, by doing so, defined an entire computer science discipline, in an interview.

Rather than an interview for any media outlet, the transcript for this interview appeared in the first edition of the computer science textbook 'Fluency With Information Technology: Skills, Concepts, & Capabilities' by Lawrence Snyder and Ray Laura Henry.

She recalled during this conversation a particularly memorable moment in which a highly respected hardware guru explained to everyone in the meeting that the process of building software should also be considered an engineering discipline – just like hardware. That's because she and her colleagues had earned the respect of others in the room as being in an engineering field in their own right.

Teething issues

Before Hamilton coined the term and defined many of the processes and best practices that drove the software achievements at NASA during the Apollo mission, the approach to computing was in a state of disarray.

People, at the time, valued physical hardware and the components fitted underneath far more than the code that nobody could see. Programming, in a sense, wasn't seen as a real science until Hamilton came along.

During her time at NASA, she created vital on-board flight software, with a particular highlight being her error-detection-and-recovery system. Her contributions were immortalized in the 2022 short film 'To Go to the Moon', in which she was portrayed by Dominque Roberts.

This attitude led to a software crisis in the 1960s through the 1980s, with many projects running over budget and schedule and with the software that was published containing critical bugs, with costly consequences. This was the direct impact of a lack of professionalization in the field.

• China launched a fully autonomous vulnerability hunting attack against Taiwan
• The attack leveraged eight open-source AI models to hunt for new attack vectors
• The attack hit Taiwan government accounts, personnel records, the nuclear safety agency, and more

A first-of-its-kind cyberattack using autonomous AI has been spotted attacking Taiwan, and it compromised 85 government accounts and stole over 2,500 personnel records before moving on to hit the country’s nuclear safety agency and at least seven energy companies.

The attack used eight open-source AI models to build a hacking program that was able to independently conduct reconnaissance and intrusion, and was able to chain vulnerabilities and change tactics whenever it was blocked.

The intrusion took place over the course of four days, and was first exposed by The Financial Times on August 12, 2026. The FT article covered research performed by Dream, an Israeli AI and cyberdefense company that first identified the breach.

Autonomous AI attack

The attack was first uncovered during routine monitoring of cyber criminal activity. Dream found a 160MB online archive of 1,395 files. Further examination of the files revealed that the attack relied on Hermes and OpenClaw - two open-source AI agents.

As is typical of attacks relying on AI models, the hackers had framed the context of the intrusion as a routine cyber readiness test in order to bypass the built-in guardrails of the AI models.

The attack used multiple agents to hunt for new vulnerabilities and access points across the internet, providing the tool with multiple attack paths to take if one failed to gain entry.

AI agents have been quickly integrated into the attacks of cybercriminal organizations and state-sponsored threat actors alike, enhancing their abilities to launch highly complex attacks at scale. “This must be the basic assumption of every government around the globe,” said Amir Becker, Dream's chief strategy officer.

Dream did not tie the attack to any specific cybercriminal group, nor did it confirm the target of the attack, but said it had alerted a government in the “Asia-Pacific.” Documentation within the recovered archive contained Simplified Chinese, which is the official written language used in mainland China.

The archive also contained data collected from the targets, which was written in Traditional Chinese. This form of written Chinese is widely used in Taiwan, Hong Kong, and Macau.

Taiwan's Ministry of Digital Affairs has refused to comment on the breach, and the Chinese authorities have not responded to requests for comment.

China has long considered Taiwan to be a part of mainland China. Taiwan declared its independence following the end of the Chinese Civil War in 1949. A report from Taiwan’s National Security Bureau earlier this year revealed that the country was subject to 2.5 million Chinese cyberattacks per day in 2025.

Expert perspective on autonomous AI attack

Collin Hogue-Spears, senior director of solution management at Black Duck:

The agents ran the intrusion end to end and invented nothing new to run it with. Familiar identity and API failures opened every confirmed path into Taiwan's systems. Dream Research Labs documented up to eight subagents working concurrently across twelve waves, ranking attack paths, redirecting when a technique failed, and researching alternatives online before trying again.

What they found was exposed development endpoints, an API accepting authentication tokens with the signature check disabled, unauthenticated data APIs, and passwords built from employee ID numbers.

The framework also ran its own AI static analysis hunting unknown flaws, but Dream says it worked against two public single sign-on SDK sample projects, and none of those findings produced a confirmed exploit on the live systems. No zero-day appears anywhere in the report, but a nuclear safety regulator does.

In conventional web and identity logs, this reads as a security scan. The distinguishing signal is the sequence across systems, not any single request. The tell is not the request. It is what the same account does next, somewhere else.

Conventional scanners have tested thousands of endpoints at machine speed for twenty years, so raw coverage is not the change here. What Dream Research Labs describes is chaining: password spraying, then fresh SSO sessions, then access to routes an account had never touched, then the same suspected weakness retested until it held, then one identity surfacing across several connected applications.

Simplified Chinese in the operator's notes is one signal. Traditional Chinese in the stolen files is just Taiwan. Dream rested its China assessment on a code-switching observation, and only half of it points at the attacker. Per Chinese-language coverage of the report, Simplified characters appeared in the operators' internal communications and Traditional characters appeared in the exfiltrated data. The first describes the operator's working language. The second describes the victim, because that is what Taiwanese government files look like [Traditional Characters].

The evidence therefore supports a Chinese Mainland-language operator against a Taiwanese target, with a target profile consistent with mainland collection priorities. It does not name a group or establish state direction. The report also publishes no indicators, no hashes, and no victim confirmation; it does not identify the model, and its executive summary claims installed backdoors while its own attack chain says authentication blocked the web shell.

Security leaders must reject unsigned authentication tokens and prohibit the alg:none setting outright, and separately require reauthentication or multi-factor at any single sign-on boundary into a sensitive system. Dream describes two independent identity failures in Taiwan, and closing one leaves the other open. Provider guardrails cannot compensate for a password-only SSO bridge.

They must also monitor route diversity per source, per session, per account, and per device rather than by request rate alone, because a distributed set of agents spreads requests across addresses and sessions that no single volume threshold catches. If your detection assumes one attacker at one address working one path at a time, you have modeled the wrong shape.

Your thresholds were built for one attacker on one path. This was eight, in parallel. And they must ask two questions of any AI attack disclosure before acting on it: which model ran the operation, and what can we hunt on tomorrow morning?"

Via United24

• The Department of War announced an expanded counter-drone marketplace
• It's built by four-year-old startup Kaizen Laboratories under a $15 million prototype contract
• The platform sees defense vendors including Anduril, DroneShield, AeroVironment, and SMARTSHOOTER listing products for sale to approved buyers

The US Department of War has announced an expanded version of the counter-unmanned aerial system marketplace run by Joint Interagency Task Force 401, its coordination hub for counter-drone work.

A report by DroneFront found the platform is built and operated by Kaizen Laboratories, a four-year-old software company in New York, under a $15 million prototype contract.

The announcement is an impressive milestone in a procurement experiment that has been running behind closed doors since February 2026, with access available only to vetted buyers.

AI-assisted digital procurement with safeguards

The inventory available to the Department of War's list of buyers is not visible to third parties, and for those who do have access, it is heavily tiered based on export-control status.

Buyers are manually approved by JIATF-401 admins and granted access based on classification levels; eligible buyers include federal procurement officers, military commands, pre-cleared allies, and defense ministries from partner nations.

Vendors on the platform are also subject to a detailed vetting process that involves business verification, product evaluation, and compliance with export controls before they are allowed to publish anything.

This essentially fits with what a digital weapons catalog is, but with considerably more firepower under the hood than a typical gun shop might carry, and where mistakes or even the display of classified data could have catastrophic security ramifications.

For example, if data about the effectiveness or limitations of certain counter-drone hardware were made public, it could undermine its usefulness in certain theaters of war.

Australia, Poland, South Korea, the United Kingdom and Romania have been publicly named as partner nations, and Army Secretary Dan Driscoll has said state and local law enforcement agencies are being onboarded as well.

JIATF-401 reported $13 million in marketplace purchases as of April. By early August, task force spokesperson Lt. Col. Adam Scher put the figure at $21 million. That shows subsequent sales of about $8 million over the next 4 months, or about $2 million per day.

For a platform that was contracted out to be built at $15 million, one could argue that the sales are moving slower than expected, but that might be the point here: Kaizen CEO Nikhil Reddy has said the platform remains in a slow-rollout phase, with a limited set of orders being used to validate that the ordering and adjudication process works before wider release.

The more consequential piece of the platform, however, is one that has yet to be released: an AI-assisted planning layer that lets a user enter a threat profile, a budget, an operational readiness state and an environment, and then, in Reddy's own description, scans the products in the marketplace agentically and returns a recommended capability configuration across the kill chain. Kaizen said in early August that this was expected to go operational within days.

Given how dynamic the modern battlefield is and how quickly demand shifts among multiple players and allies, having a more centralized yet relatively secure platform to issue orders and conduct threat assessments is a key addition to the Pentagon's arsenal as it continues to support a plethora of allies and domestic partners alike. Whether its website at cuas.mil, which already has restricted access from certain countries, will meet those needs in the future remains to be seen, but it does seem to be off to a promising start.

• Linus Torvalds called Linux 7.2-rc7 part of a "new normal" of oversized release candidates
• Torvalds attributes the volume of fixes to reviews by various AI tools, a week after calling rc6 the biggest in years by commit count
• The growth is in late-cycle bug fixes rather than features, and Torvalds credits AI review rather than AI authorship, with humans still writing and submitting the patches

Modern coding approaches dictate that Release Candidates should be expected to shrink as patches and code fixes get increasingly smaller as the software enters what many call a "stabilization phase".

Linux 7.2-rc7 did not do that, and neither did rc6 before it - and announcing the seventh candidate on August 9, 2026 Linus Torvalds wrote he could not say he was thrilled about the size of it, but that it is what it is: the new normal, with a lot of fixes, many of them the result of review by various AI tools.

A week earlier, opening the rc6 message, Torvalds had been blunter still, calling that release candidate huge and reckoning it the biggest rc6 in years by commit count.

An easier, detailed AI-enabled review process?

Torvalds is not describing AI writing for the Linux kernel; that still requires human intervention. The tools being used are review and analysis agents pointed at existing code, and they produce actionable bug reports. Humans then write the fixes, submit them through the normal subsystem maintainer process, and the creator of Linux, as the maintainer, pulls the code.

Phoronix, which closely tracks the cycle, describes the current environment as one in which AI and LLM coding and review agents keep kernel activity at an all-time high. It has resulted in a much larger-than-normal Release Candidate this late in the cycle; RC7 comes packed with a plethora of bug fixes that would otherwise have been in play much earlier in past cycles.

By one count of the RC7 pull, more than 400 fixes came in from upwards of 230 contributors, a volume more typical of an early merge window than a closing candidate.

Torvalds noted in the accompanying message that nothing looked particularly scary, and that the majority of the fixes were spread across drivers, filesystems, core networking, and architecture code, which was relatively small, making a potential release by the following weekend possible.

The AI approach doesn't have Mr. Torvalds particularly keen, but he has subtly acknowledged its advantages when it comes to code review, and it could bring a different issue into play: the maintainer model, which involves a few trusted individuals to read and test code before it is eventually released to the masses may not be too keen to have the equivalent of an AI-powered firehose of fixes and new information to deal with indefinitely.

For now, it seems larger updates are the new norm, for better or worse, and AI will be powering an increasingly large share of code review in an ecosystem that drives most of the world's servers, smartphones and tablets, as well as a sizeable portion of its personal computers.

• Nightmare Eclipse discloses ShieldBreak, a new Windows privilege‑escalation zero‑day
• Flaw bypasses a recent patch and works on fully updated Windows 11 systems
• Researcher’s ongoing exploit spree leaves multiple Windows vulnerabilities still unpatched

Nightmare Eclipse has struck again! The notorious zero-day researcher with a Microsoft grudge disclosed its latest vulnerability, and just as in previous instances, they picked their timing and released their research hours after Microsoft published its August Patch Tuesday cumulative update in order to maximize the hurt.

The newest flaw is called ShieldBreak, and is described as a local escalation of privilege vulnerability that allows threat actors to gain SYSTEM-level privileges on vulnerable systems. Speaking of vulnerable systems, the list is rather long because it includes all versions of Windows 11, including those with the latest security patches.

“The PoC was tested in the latest version of windows 11 25h2 (+Canary channel) and windows server 2025, the PoC also have a 100% success rate,” Nightmare Eclipse said on their GitHub account. “Please note that Windows 10 (and respective server editions) are not currently supported, they are however vulnerable to ShieldBreak as well.”

A bypass for the RoguePlanet fix

The mysterious attacker also said that the bug was actually a bypass for the patch Microsoft issued to fix their earlier work, called RoguePlanet.

“Microsoft has failed to properly patch the RoguePlanet vulnerability CVE-2026-50656, this PoC demonstrates a full patch bypass,” the GitHub read entry.

Microsoft, on the other hand, responded in pure enterprise fashion, sharing a boilerplate statement that it was “investigating” and that it “supports coordinated vulnerability disclosure”.

In response to an enquiry by The Register, a company spokesperson said Microsoft "is aware of the reported vulnerability and is actively investigating the validity and potential applicability of these claims."

"Microsoft is committed to investigating security issues and updating impacted products to protect customers as soon as possible,” the statement reads. “Importantly, we support coordinated vulnerability disclosure, an industry standard that protects customers and supports the research community by ensuring their findings are thoroughly investigated and addressed before being made public."

A hacker on a mission

Together with ShieldBreak, the number of disclosed Windows vulnerabilities and exploits now counts 10. Nightmare Eclipse’s campaign began in April 2026, when they demonstrated BlueHammer, a Windows Defender local privilege-escalation flaw that gives low-privileged users SYSTEM-level access. The researcher claimed BlueHammer, now tracked as CVE-2026-33825, was previously reported to Microsoft, but the company allegedly mishandled the disclosure.

Just before publishing the work, they said “someone violated our agreement and left me homeless with nothing. They knew this will happen and they still stabbed me in the back anyways, this is their decision not mine.”

At first, Microsoft took a tough stance, calling the public release “never justifiable” and even warning that it might pursue legal cases against people who put customers at risk.

The community interpreted this statement as a threat of legal action against Nightmare Eclipse, which triggered a backlash. Microsoft later backed away, saying “to be clear about our approach to legal matters, we have no intention to pursue action against individuals conducting or publishing their security research.”

In the meantime, Nightmare Eclipse (also known as Chaotic Eclipse) went on a full-blown rampage. They released RedSun and UnDefend (both targeting Defender), YellowKey (a BitLocker bypass), GreenPlasma (a CTFMON-based privilege-escalation flaw), MiniPlasma (a regression of a vulnerability Microsoft had originally fixed in 2020), RoguePlanet (another Defender privilege-escalation bug), GreatXML (a BitLocker/Windows Recovery Environment bypass), LegacyHive (a Windows User Profile Service privilege-escalation flaw), and now ShieldBreak.

BlueHammer was fixed in April, RedSun and UnDefend in May, and YellowKey, GreenPlasma, and MiniPlasma, in June. RoguePlanet was patched in July, while LegacyHive, GreatXML, and ShieldBreak, remain unpatched.

It is also worth mentioning that not all of Nightmare Eclipse’s releases were equally complete or reproducible by third parties. For GreenPlasma, independent researchers said it contained the vulnerability but turning it into a reliable working exploit required significant additional technical work. Some of the early Defender exploits were also apparently difficult to reproduce, mostly because they relied on delicate race conditions and very specific sequences of Windows components.

ShieldBreak, however, seems to be more dangerous in that respect. Speaking to The Register, security researcher Kevin Beaumont confirmed it as working: “I've tried it, it works on latest Windows 11,” he told the publication.

He also said that while ShieldBreak was described as a bypass for the RoguePlanet fix, the two flaws actually operated quite differently.

“RoguePlanet was a filesystem race condition vuln that uses virtual disks and NT native file manipulation to trick quarantine process into overwriting system files,” Beaumont explained. “ShieldBreak user-mode callback hook to change file contents during a Defender cloud-hydration scan via cfapi (Cloud Filter API).”

No one knows how much ammunition Nightmare Eclipse still has, but we will certainly be paying attention to them in the hours after next month’s Patch Tuesday, as well.

• WindRelay campaign used vishing plus custom malware to turn phones into POS skimmers
• Victims installed personalized RATs and NFC malware, enabling real‑time card theft
• Attacks were highly targeted across Eastern Europe, with only a few individuals hit

Hackers are turning people’s smartphones into malicious Point of Sale (POS) devices and stealing their money directly from their payment cards, experts have warned.

Security researchers Group-IB spotted multiple such attacks across Eastern Europe, and named the campaign WindRelay, after the custom-built malware used during the attacks.

The report notes this is a highly sophisticated, custom-tailored attack designed specifically for the victim. It starts with some form of reconnaissance, in which the attackers learn their victim’s identity, phone number, and likely other details. Although the researchers don’t discuss it, it is quite possible that the attackers obtained (or purchased) the data from unrelated data breaches and leaks.

Vishing and malware

After learning a little bit about their target, the attackers get to work. They first prepare a remote access trojan (RAT) named SpyNote. They personalize the label with the victim’s own name (instead of it being a generic or impersonated brand), to build trust with their victim:

“Such tactics are more effective at weakening a victim’s natural defenses and suspicions,” the researchers noted in the report. “It removes the one cue people are trained to check before installing something unfamiliar — a strange or generic app name — right at the moment they’re most likely to hesitate.”

Then, they call the victim on the phone and introduce themselves as employees of their target’s bank. They claim the victim has a problem with their bank card, and instruct them to deploy SpyNote through the device’s package installer (the standard way to sideload apps outside an official app store).

SpyNote is a classic RAT that the attackers then use to deploy stage-two malware themselves. In this next step, they personally (as opposed to having the victim do it) install WindRelay, custom near-field communication (NFC) malware designed to capture contactless payment card data in real-time, when a card is tapped against the phone.

In other words, the malware turns the smartphone into a POS, and when a victim taps their card against it, the information is relayed to an attacker’s terminal.

Vishing + malware

Vishing + malware combo is nothing new. We’ve seen it deployed numerous times before, and ShinyHunters are probably the shiniest example of the practice (pun definitely intended). Over the last couple of years, ShinyHunters have been calling their victims on the phone, impersonating the IT department, and getting their victims to log in via fake login portals which relay the credentials to the attackers.

They then use the credentials to access their victims’ SaaS products, exfiltrate as much sensitive data as possible, and then demand ransom in exchange for deleting the stolen files.

This new campaign, however, is a testament to the technique’s evolution. While ShinyHunters’ operatives only stay on the phone call until the victim logs in, these crooks remain on the line for as long as it takes. Group-IB says the average call lasts around 13 minutes, and by that moment, the victim will have installed both SpyNote and WindRelay, and has tapped their bank card against the phone, making unwanted payments.

In at least one case, the attackers successfully applied for a loan at the victim’s bank, stealing not only the money they had on their account, but also money they would have earned in the future.

The identity of the attackers is unknown at the time. We also don’t know exactly how many victims there were, but given the highly personalized nature of the attack, it’s safe to assume that there were only a handful.

Group-IB says it observed attacks in Czechia, Slovakia, and Slovenia, suggesting a threat actor focused primarily on Eastern European victims. The researchers also said they identified 23 samples uploaded to VirusTotal between November 2025 and July 2026, meaning the campaign was active for approximately seven months, targeting 23 individuals.

“The samples mimic various institutions from the targeted countries and contain text in the language of each targeted country,” the researchers said. “Some samples contain personalized UI elements and labels, such as the name of the victim, similar to the personalized RAT. This suggests the threat actor behind these campaigns most likely has the capability to dynamically build malicious applications tailored to individual victims.”

Group-IB says users should treat personalized app labels as a red flag and should apply extra friction to loan applications.

For years, the PC was a tool that waited for instructions. Even the most advanced software still needed someone to open it, tell it what to do and check the result. AI agents are beginning to change this model.

Rather than simply responding to prompts, AI agents can help users pursue goals across multiple steps. They retain awareness of previous activity and use approved tools and automate portions of workflows within organizational guardrails.

This shift is changing the way organizations think about PCs. For years, the buying criteria were straightforward: performance, reliability, security and cost. The question was whether a device could run the software employees needed.

Now, organizations also need to consider whether a device can support employees working alongside AI agents. That makes a fleet refresh about more than specifications alone.

The PC is no longer just a tool

That collaboration changes the role a device plays in the working day. Traditionally, a PC has been a tool that waits for instructions, responding when an employee opens an application, enters information or starts a task.

AI agents create the opportunity for a more continuous and proactive working relationship, where the device can help employees navigate workflows, surface relevant information and connect work across activities.

Today, that might mean preparing for meetings or helping draft reports. Tomorrow, it could mean helping a project team maintain a shared understanding of a complex program over months, identifying risks, tracking commitments and surfacing relevant information before someone even thinks to search for it.

In that sense, the PC becomes more than a gateway to applications; it’s a platform that helps people navigate their working day by connecting information, insight and actions across tasks.

Delivering that experience requires more than running a browser and a spreadsheet. Agents need to maintain the thread of work across tasks, access information securely and support multiple AI-driven processes without compromising performance or governance.

That creates a new set of requirements and makes fleet decisions more strategic than they have traditionally been. If the infrastructure is not ready, organizations risk limiting both the effectiveness of the agent and employees' trust in it.

Why where AI runs matters

As AI becomes a larger part of everyday work, organizations are paying closer attention to where processing takes place. Advances in PC hardware mean more AI processing can take place locally on the device, which is particularly important given AI agents will likely need to operate continuously and handle sensitive information.

This is particularly important in regulated sectors. A financial services firm reconciling client data or a hospital summarizing patient records needs confidence in how data is processed, stored and governed. Running more AI workloads locally can provide organizations with greater flexibility in how they meet those requirements.

That brings the conversation back to the role of the PC. If organizations want employees to work effectively alongside AI agents, devices need to support AI experiences securely, connect activities across workflows and deliver those capabilities in a way that aligns with organizational governance requirements.

Technology is only part of the answer

Employees are ready for this shift too. Research has identified a “transformation paradox” in which employees are often more ready to work with AI tools than the organizations around them are prepared to support.

Only 30% of AI-using UK employees believe their organization's leadership is clearly and consistently aligned on AI strategy, while 13% say they are actively rewarded for redesigning how they work with AI. The result is a growing gap between what people can do with AI and what organizational structures, incentives and processes are designed to support.

Hardware alone will not close that gap, but it is one of the more straightforward parts of the problem to solve. Fleets that aren’t ready to run agents locally, securely and consistently simply add friction on top of the cultural and leadership challenges organizations already face.

As AI agents become a more familiar part of the working day, the role of the PC is changing. For decades, devices were primarily judged on their ability to run applications efficiently. Increasingly, they will be judged on how effectively they help people work with AI. In that environment, the PC becomes more than a gateway to applications; it becomes a platform for connecting information, intent and action.

Organizations that recognize that shift early will be best placed to capture the opportunities AI creates.

We've featured the best business laptop.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

There are more than 70,000 AI companies operating today.

Most of them will not exist in five years.

Before you can see why — or figure out whether yours is one of them — you need a distinction the market keeps blurring.

Strip away the pitch decks and there are really only two types of AI system being built today.

The first is AI infrastructure: the orchestration and governance technology that makes AI usable at scale. In plain terms, this is the plumbing — agent frameworks, model routing, evaluation and monitoring tools, guardrails, and the controls that let a large organization use AI safely.

It sits between the foundation models and the end user, and it is where an enormous amount of venture money is going right now.

The second is the surface application: the tool an actual person uses to do actual work. The underwriting assistant, the contract reviewer, the sales copilot. The thing with a login screen and a job to do.

What I see in the market is a blending of the two. Some firms are selling architecture.

Some are selling tools. Many are trying to sell both, on the theory that owning the whole stack is the safest position.

And while this market is filled with tremendous exuberance with seemingly everyone starting an AI company, I am very skeptical that many of these firms will ever see profitability as history offers a strong counter.

We've run this experiment twice.

The past and the future

The dot-com era ran the first version of this experiment, and its final tally is worth stating plainly. Researchers estimate that roughly 50,000 startups were founded in the United States between 1998 and 2002 to commercialize the internet.

Of those, something like 8,000 attracted venture funding. About 1,700 internet-related companies made it to an IPO across the whole era — 585 in 1999 and 2000 alone — and at the peak, only about 14 percent of the tech companies going public were profitable.

By late 2002, most internet stocks had lost more than three-quarters of their value and roughly 1.7 trillion dollars had been wiped out. And the number of enduring, large-scale winners from that entire cohort — Amazon, eBay, Priceline, Expedia — you can count on two hands. Run the funnel: 50,000 founded, 8,000 funded, 1,700 public, fewer than ten giants.

A real gold rush works the same way: a few strike it rich, some make a living, and most go home with less than they brought. This is important to remember for everything that follows.

If that funnel looks like a quirk of one bubble, it is not — it is how markets distribute winnings everywhere. Hendrik Bessembinder at Arizona State studied every U.S. stock since 1926, more than 25,000 companies, and found that the best-performing 4 percent account for all of the net wealth the stock market has ever created; the other 96 percent, taken together, did no better than Treasury bills.

Just 90 companies — a third of one percent — produced more than half of it, and the majority of stocks lost money outright over their lifetimes. The market wins; almost no individual company does. Keep that in mind every time someone tells you AI will create trillions in value. It will. That says nothing about whether any particular company captures a dime of it.

The example of cloud

Cloud computing is the sharper rerun. In the early days there were hundreds of cloud providers and a thriving ecosystem of middleware companies selling the connective tissue — provisioning tools, management layers, monitoring platforms.

Today three companies control roughly two-thirds of the cloud market, and their share grows every year.

And here is the part that matters for AI: the middleware layer did not consolidate alongside the platforms. It was absorbed by them. The hyperscalers built the management consoles, the monitoring, the orchestration, and shipped it as a feature. The companies whose entire business was cloud plumbing were acquired cheap or squeezed out.

Meanwhile, the application layer on top of that consolidated infrastructure exploded. Thousands of SaaS companies built durable, profitable businesses without owning a single server. The bottom of the stack ended up in a few hands. The top produced thousands of winners.

It has already happened once inside this stack

If cloud feels like ancient history, look at the data layer — the foundation every AI system sits on. That consolidation already occurred, and it finished recently. The "modern data stack" boom of the last decade funded hundreds of startups selling pipelines, catalogs, transformation tools, and warehouses.

Today the independent tier has settled to exactly two companies at scale: Snowflake and Databricks, each running at roughly five billion dollars in annual revenue, with the hyperscalers’ native offerings holding most of the rest of the market. Nearly everyone else was acquired, absorbed as a platform feature, or left scraping for the remainder.

And notice the shape it settled into. The top five data platforms — Snowflake, BigQuery, Redshift, Databricks, and Microsoft’s offering — hold roughly two-thirds of the market. That is almost exactly where cloud landed: three players, about two-thirds of the market, a long tail fighting over the rest.

Two different layers, a decade apart, ending in the same proportions. That is not a coincidence. It is what happens when competing takes huge capital and the platforms can build whatever sits next to them. Expect the AI orchestration layer to end up the same way.

The consolidation was driven as much by the buyer as by the vendors. Large enterprises learned that scattered data is expensive data: every additional platform meant another copy of the truth, another integration, another security review, another contract.

So CTOs stopped buying data tools one team at a time and started making strategic platform decisions — pick one or two providers, consolidate the estate onto them, and hold that line. A single source of truth became an explicit architectural goal at most large companies, and once thousands of enterprises were making that same decision, the market had no room left for a long tail of vendors.

Look at what it took for Snowflake and Databricks to survive that consolidation: enormous capital, the fact that customers’ data lives on their platforms and is costly to move, and deep ties into how their customers work every day. You can survive as an independent alongside the hyperscalers — but only by becoming one of the few names a CTO puts on the strategic list, and almost nobody makes that list.

The same consolidation is coming for AI

Apply that pattern to the two types of AI company and the forecast writes itself.

The infrastructure layer — orchestration and governance — will consolidate down to a few. Not because the current tools are bad, but because this layer sits directly in the expansion path of the biggest players in technology. The model providers and hyperscalers have every incentive to build orchestration, evaluation, and governance into their platforms, and they are already doing it. Every capability that today justifies a standalone infrastructure startup is a roadmap item at a company with a hundred times the resources and a direct line to the same customers.

If you are building an architecture-only solution, this is the uncomfortable implication: you are likely to be taken out by one of the big players. Maybe you get acquired, if you are early and lucky. More often, the platform simply builds what you sell and includes it for free. Either way, orchestration and governance alone is not a business you can hold. The only real question is how long you have.

Which leaves the application layer as the open field. And this is the counterintuitive part: infrastructure consolidation is good news for application builders. When orchestration and governance become cheap, standardized, and built into the platforms, the cost of building a serious AI application collapses — just as commodity cloud ignited the SaaS boom. We are already seeing a massive increase in the number of AI applications getting built, and most will likely not survive.

Better software, worse odds

Part of what makes this cycle different is how little it costs to enter. Building serious software used to take millions in capital and a room full of engineers — a filter that limited how many companies could even try. Today a handful of people with AI tools can ship in weeks what took a funded startup a year.

So new ventures are multiplying, not because there are more good ideas, but because the cost of trying has collapsed. The scale tells the story: more than 70,000 AI companies operate globally today, roughly 18,000 to 30,000 of them in the United States alone.

The comparison to the dot-com era’s 50,000 is not perfectly apples to apples — that was a five-year founding total for one country, this is a snapshot of companies operating worldwide right now — but the order of magnitude is the same, this wave is global, and the count is still climbing.

Here is the twist that makes the coming shakeout more brutal, not less: the software being built is genuinely good. This is not the dot-com era, where half-finished products hid behind splashy marketing. The tools are now so powerful that quality is the baseline — which means quality has stopped differentiating anything. When every product is polished, capable, and shipped fast, none of that separates you from the next founder who did the same thing last month.

And that is precisely why so few founders see the danger. Every one of them genuinely believes they are building something singular — and by their own measure, they are right. They compare their product to what came before: the clunky incumbent, the manual process, the way the work used to get done. Against that benchmark it looks revolutionary.

What they never compare it to is the tens of thousands of other teams looking at the same models and the same problems, building virtually the same thing at the same time. Measured against the past, every AI product is remarkable. Measured against the field, almost none are. More entrants than either previous cycle, all building excellent software, almost none of it distinguishable.

That is the setup for the largest culling yet, and it will run almost entirely on the moats, because there is nothing else left to separate the winners from the losers.

The delusion of special

I see this up close. I have this conversation with application founders every week, and it always goes the same way. They believe the quality of what they built is their moat: the product works, customers love it, nothing else on the market feels as good.

All of that can be true, and none of it protects them. Quality can be copied. The same tools that let them build an excellent product in months let a competitor build one in weeks. A few founders have built something that truly stands alone, but I just can’t see many finding a way to real profitability.

There will be some winners, but I think they will need to rest on three key differentiators:

1. Data. Not data you scraped or licensed — proprietary data your business generates by operating: claims histories, transaction flows, patient outcomes. If your system gets smarter from data competitors cannot obtain at any price, you compound. If you are building on the same public internet as everyone else, you do not.

2. Distribution. If you already own the customer relationship — an installed base, a trusted brand, an embedded sales channel — you can put an AI product in front of buyers faster and cheaper than any startup. This is why incumbents are more dangerous in this cycle than the last one. The startup has to build the product and buy the audience. The incumbent only has to build the product.

3. Integration into workflows. The one people underestimate. Companies that wire themselves into how work actually gets done — the approvals, the systems of record, the daily habits of thousands of employees — become painful to remove even when a rival ships something better. Switching costs are not glamorous, but they have protected enterprise software for thirty years, and they will protect AI applications too.

Have one of these and you can build a durable business on commodity infrastructure. Have two and you can build a great one. Have none and you are likely running out of time.

Your toughest competitor is your customer

And here is what makes the application layer even harder than the dot-com or SaaS eras: surface applications are not just competing with other vendors. They are competing with the companies they are trying to sell to. The same commodity infrastructure that makes it easy for a startup to spin up an AI application makes it just as easy for the buyer to build one internally.

Every enterprise pitch now runs into a question that barely existed in the SaaS era: why would we buy this when a small internal team could build it in a quarter?

And here is the uncomfortable part. The three advantages that decide the application winners — distribution, proprietary data, embedded workflows — are precisely what the buyer already has. The enterprise owns its data. It is its own distribution. It controls its own workflows. The customer starts the build-versus-buy conversation holding every moat you are trying to claim.

A surface application does not just need to be better than its competitors. It needs to be so much better than what the customer could build themselves that buying beats owning — and that bar rises every time the underlying infrastructure gets easier to use.

Know which company you are

I am not going to pretend to know which specific firms win. But the structure of the outcome is already visible, because we have now watched it three times — dot-com, cloud, and the data layer: infrastructure consolidates to a few, applications proliferate, and the survivors are the ones holding data, distribution, or workflow integration that cannot be copied.

So the first question is not "is my product good?" It is "which of the two companies am I?" If you are infrastructure, your realistic endgame is being bought or being bypassed — plan accordingly. If you are an application, the model is not your moat and the product probably is not either.

So what is?

The good news, and who gets it

One clarification before closing, because everything above can read as pessimism about AI itself. It is the opposite. The technology will create enormous value, and the markets built on it will grow. The open question is who keeps that value, and a century of evidence gives a consistent answer: mostly the consumers of a technology, not its producers.

William Nordhaus at Yale measured this across decades of American innovation and found that producers capture only about 2 percent of the total value their innovations create — the rest flows to the people and businesses that use them. Railroads transformed the economy and ruined most of their investors. Airlines moved the world and destroyed capital for a hundred years. The internet made a handful of platforms rich — and made every company that deployed it more productive.

This cycle is already tracing the same shape: the infrastructure layer consolidates, prices its scarcity, and books historic profits, while the application layer competes and hands its margin to the buyer.

That is the real ending of this story. The coming massacre of AI companies and the coming growth of the AI economy are the same event, seen from opposite sides of the table. If you sell AI, the funnel is your problem and the moats are your only defense.

If you buy AI, the competition among 70,000 firms is working precisely in your favor: every improvement, every price cut, every copied feature moves value from their side of the table to yours. The bad news in this article is only bad depending on which chair you sit in.

We've reviewed, rated, and ranked the best business cloud storage.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

• Ahrefs has launched Letaido, a new agent-powered marketing tool
• Letaido will have native access to Ahrefs data
• It promises to help save users hours every week

Ahrefs, one of the biggest digital marketing, SEO, and AEO platforms, has announced the launch of Letaido, its agent-powered marketing workspace.

The new tool promises to help teams move beyond AI chatbots to plan and execute multi-step tasks, build bespoke dashboards and reports, and automate repetitive workflows. It will also work behind the scenes to continuously monitor websites and competitors.

According to a press release distributed by Ahrefs earlier this week, Letaido will benefit from native access to Ahrefs data, helping marketers and agencies automate content marketing, SEO, and marketing research without any need to manage custom integrations of APIs.

The new tool will also seamlessly connect with widely used marketing platforms with secure integrations, built-in hosting, and always-on infrastructure, supporting workflows that keep on running, even when the instructions have stopped being given.

How will Letaido help marketers?

According to data from BCG’s 2026 global CMO survey, only 32% of CMOs have rebuilt how marketing operates, and 42% admit they use GenAI exclusively as an assistant for individual tasks. Ahrefs promises that Letaido will help close the gap between the first phase of generative AI, which operated largely in silos, and AI that can operate across your full marketing stack, helping marketing teams and agencies connect and plan multi-step tasks.

"Marketing teams have spent years accepting that getting to an insight means clicking through reports, exporting CSVs, filtering data, and manually cross-referencing information across several tools," said Ross Simmonds, CEO and founder of Foundation. "With Letaido, those hours can now go toward the work that actually matters: thinking, strategy, and creativity. Our team has seen keyword research and bottom-of-funnel content audits, work that used to take a full 40-hour week, now take just 60 minutes."

AI-generated content and increasingly fragmented communications are changing how organizations assess digital evidence. For investigators, establishing authenticity is becoming as important as analysing the evidence itself.

For years, digital evidence carried an assumption of authenticity. A document was presumed genuine unless there was reason to believe otherwise.

A screenshot reflected a conversation. A photograph captured a moment in time.

But that assumption is becoming harder to sustain.

Advances in artificial intelligence and editing tools have transformed how digital content is created and shared. At the same time, investigations increasingly rely on evidence gathered from messaging platforms, cloud services and personal devices.

Together, these developments have made questions of authenticity far more prominent.

Before evidence can be relied upon, investigators need to understand where it came from, how it was obtained and whether it accurately reflects the underlying source material.

The end of assumed authenticity

Digital evidence remains central to most investigations. Documents help establish timelines and decision-making. Communications provide insight into intent and behavior. Yet both can present challenges that are not immediately visible.

A document may appear final and authoritative while representing only one version of a longer history. Earlier drafts may be unavailable, edits may leave few traces, and supporting material may never be disclosed.

Communications create similar difficulties. Screenshots and message extracts can appear highly persuasive because they seem direct and contemporaneous. Yet they often lack context. Messages may be missing, conversations may be incomplete, and edits or deletions may not be visible.

In some circumstances, entire exchanges can be fabricated before they are ever disclosed. The issue is not that digital evidence has become inherently unreliable. Rather, reliability cannot be assumed simply because something looks convincing.

Why screenshots deserve scrutiny

Screenshots are among the most commonly encountered forms of digital evidence, particularly in workplace investigations and disputes.

Consider two screenshots of what appears to be the same WhatsApp conversation. Both display realistic timestamps, the same participants and a seemingly authentic exchange. Yet the content differs in a crucial respect: one version appears to confirm a contractual agreement, while the other suggests different terms altogether.

To a casual observer, both screenshots may appear equally credible. For investigators, however, the image itself is only part of the story.

Questions about when the screenshot was created, who captured it and whether it can be traced back to the original data source are often more important than the image itself. Without access to the underlying message database, device or backup, it can be difficult to establish with confidence whether a screenshot accurately reflects the original conversation.

This is why screenshots are rarely treated as definitive evidence in isolation. Their value often depends on whether they can be corroborated by the underlying source data.

Deepfakes and synthetic content

The rise of generative AI has added another layer of complexity. Manipulated images and videos are not new, but the tools needed to create them have become far more accessible. Content that once required specialist skills can now be produced in minutes.

While sophisticated deepfakes remain relatively uncommon in most corporate investigations, their existence has changed expectations around digital evidence. Authenticity can no longer be taken for granted.

As a result, organizations increasingly need processes that establish whether content is genuine before it is used to support important decisions.

Testing evidence, not accepting it

In response, investigators are placing greater emphasis on validating evidence before drawing conclusions. The starting point is often provenance: understanding how material was obtained and whether it came directly from a source system. Evidence collected from original sources will generally carry more weight than material provided selectively by individuals.

Investigators also look for signs that content may have been altered. Metadata, file properties and technical inconsistencies can sometimes reveal issues that are not visible at first glance.

Just as important is understanding what may be missing. A message thread viewed in isolation can tell a very different story from the same conversation viewed in full.

These questions help establish how much confidence can reasonably be placed in a particular piece of evidence.

Confidence matters

Digital evidence rarely falls neatly into categories of authentic or inauthentic. More often, investigators are assessing levels of confidence.

A complete WhatsApp conversation extracted directly from a device and verified through forensic analysis will naturally carry more weight than a handful of screenshots supplied by one individual. Both may support the same conclusion, but the strength of the underlying evidence is different.

This distinction is important because investigations often involve decisions being made before every uncertainty has been resolved. Understanding the reliability of the available evidence helps organizations make those decisions on a firmer footing.

Trust in a digital world

Digital evidence remains one of the most valuable sources of information available to investigators. What has changed is the level of scrutiny required before that evidence can be relied upon.

Documents, messages and media files continue to play a central role in establishing facts and understanding events. However, appearance alone is no longer a reliable indicator of authenticity.

As digital content becomes easier to manipulate and generate, investigators need to pay closer attention to where evidence originated and how it can be verified.

In a world where convincing content can be created with increasing ease, trust is becoming a central part of the investigative process.

We've reviewed, rated, and ranked the best cloud document storage.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

• .com is no longer the highest-priced TLD in terms of average value
• .ai is at an all-time high average value of $535,000, a 260% surge in three months
• .com domains still lead in terms of volume at $86 million, compared to .ai at $37.7 million

.com has long been established as the gold standard top-level domain (TLD). From Amazon and eBay to Google and Facebook, practically every online behemoth has been built on .com.

But with so much new technology flooding the market, could this all be about to change?

New research has claimed .com has lost its standing as the highest-priced web address in terms of average value – beaten by .ai.

The .ai boom

The report by Escrow.com found a record-high average value of .com domains of $529k, beating its previous cycle peak in 2021. Yet, despite its huge value, .ai still beat it, with an all-time high average value of $535k, marking a 260% surge in just three months.

This data reinforces a clear trend, with AI continuing to spread from the stock market into the TLD space and beyond.

The report also highlights a 270% increase in the volume of .ai domains, growing from $10.0 million to $37.7 million during the period. However, .com remains considerably higher by volume at an impressive $86 million.

"This quarter's .ai growth alone beat the combined gain of every other country-code extension. A year ago we didn't know if .ai would actually challenge .com or just fizzle out like the rest. Now we've got an answer. Buyers are paying more, on average, for a good .ai name than a good .com one," shared Matt Barrie, Chief Executive of Escrow.com.

With many AI-focused businesses entering the market, there is little surprise that there is growing demand for the .ai TLD. Startups and small businesses using it gain an immediately recognizable sign of who they are and what they do.

The risks of .ai

Picking a .ai domain isn’t without its risks - using .ai can make it hard to pivot. For example, if a startup launches its brand on a .ai domain, only to later pivot away from AI, the equivalent .com domain may not be available or may be incredibly expensive.

Brands also need to factor in that .ai isn’t a generic domain like .com, but is an official country-code top-level domain (ccTLD) for Anguilla, a British Overseas Territory in the Caribbean.

According to some reports, .ai domain registrations have transformed Anguilla’s economy, which has primarily relied on tourism in the past. However, this reliance on a single local government can create uncertainty for .ai users, subjecting them to potential regulatory changes, geopolitical instability, and political shifts.

• US government to allow private firms to conduct legal cyberattacks on foreign organizations targeting Americans
• Firms will be allowed to disrupt and destroy physical and virtual information systems and networks
• US victims of cyber scams and fraud lose around $20,000

President Trump has signed a memo which allows private US firms to partner with the US government in operations designed to surveille and disrupt Transnational Criminal Organizations (TCOs).

According to the memo, the partnership “will enhance our ability to counter TCO threats and combat transnational cybercrime, fraud, and other predatory schemes against American citizens,” essentially turning private companies into privateers with the ability to launch cyber attacks against foreign entities.

The memo marks a significant shift in how the US tackles foreign cybercrime. “American businesses’ innovative capabilities have historically been underutilized in efforts to identify and disrupt criminal networks operating in cyberspace,” the memo states. But what does this actually mean in practice?

US to leverage private sector for cyber defense

The US is the most targeted country in the world for cyber attacks and cybercrime, with 6.7 million victims losing $138.9 billion in the last year, placing the average loss per-victim at around $20,731.

The program will effectively create a global cyber surveillance network that acts as an early warning system against attacks targeting critical national infrastructure, such as the recent Iranian attacks targeting over 30 US water systems. Private companies that “discover an imminent cyber-attack against United States critical infrastructure” will be required to notify the National Coordination Center (NCC).

The US government isn’t just looking to work alongside the big tech companies. Big companies will be part of the picture to “provide critical capacity,” but smaller companies will also have the opportunity to become involved with the program as they are “more agile,” and “may be better suited for specialized or discrete tasks.”

When a threat is detected, private companies will put together a “cyber operations package” to be reviewed and approved by the Program Executive Directors. These packages will likely include plans for surveillance and offensive cyber operations.

Private companies looking to become part of the program will be vetted according to government guidelines, and will have to operate within a set of operating procedures under the oversight of the federal government. “No operation may be approved unless it complies with these operating procedures,” the memo says.

There is however a caveat that those involved within the program must “maintain a bond or escrow in an amount not less than $1 million,” which would be forfeit should a private company “enter non‑compliance with its contractual agreement.”

The memo also sets our parameters to prevent private companies from accidentally or intentionally targeting US citizens, or US information systems at home and abroad, with the company required to “cease such operation, conduct minimization procedures, and immediately notify the NCC,” in the event that a company “discovers operational activity exceeding the parameters and restrictions of the cyber operation”.

Additionally, “any activity authorized by the Program must be conducted subject to the oversight, operational control, and legal authorities of the United States Government”.

The memo is the latest step in the Trump administration’s efforts to allow private companies to legally launch cyberattacks on behalf of the US government. “The American private sector is the most innovative and technologically advanced in the world, and its scale, speed, and capacity secure a critical offensive cyber advantage for the United States,” the memo states.

The program will also put together a report on its progress every year, as well as reviewing the performance of each private company within the program within the same time frame.

Among the giants of the digital world, hyperscalers like Amazon, Google, Meta and Microsoft will collectively allocate more than $500 billion in capital expenditure to AI infrastructure this year.

When a handful of players invest upwards of half a trillion dollars in data centers, primarily to support AI, that tells us they see a ton of demand.

We can say with total certainty that they have every intention of recouping their investment (and then some).

Indeed, the voracious demand expected from enterprises for AI capability is what has made hyperscalers, frontier labs and GPU makers darlings of the stock market.

However, their anticipated revenues are the future cash flows of companies, including those in industrial sectors like manufacturing, energy and transportation.

The Cost Shift We’re Not Talking About Enough

Up until recently, much of the discussion around AI costs has focused on training large language models. The cost of training just one OpenAI model, GPT-4, for example, is said to have been north of $100 million.

That’s a big number. It’s also largely irrelevant to the industrial organizations that will never train a frontier model themselves.

Inference, the process of applying models in live operating environments to generate recommendations, support decisions and take actions, is a different matter, however. Yet it’s largely absent from most boardroom conversations.

Unlike training, which mostly represents a single upfront investment, inference tasks to support decisions and operational execution run continuously.

As companies shift from experimentation to large-scale deployments of AI across the value chain, inference will soar as a share of workloads and IT budgets.

Tokenomics Take Root

Most businesses do not understand the true cost of AI from a financial point of view. Industry needs a new economic calculus: “tokenomics.”

Tokens—the units through which generative AI systems process and generate information—are the standard mechanism to measure and charge for AI consumption. They’re like the kilowatt-hours for which electric utilities charge their customers.

An AI agent that performs tasks across large operational datasets consumes lots of compute (and storage, networking, electricity, water and more), and the total bill reflects that increased usage.

As companies move more and more to agentic systems, these costs will mount quickly. Agentic workflows require several model calls for each task, involve large context windows and need multiple reasoning loops. By their nature, they are many times more compute-intensive than the generative AI tools most enterprises have used so far.

A Familiar Pattern

Recent research suggests the AI inference market will more than double over the next five years.

Many of us witnessed a similar pattern earlier in our careers during the move to the cloud. In the 2010s when cloud computing was reaching scale, the same economic phenomenon was at play: unit costs dropped sharply while total usage skyrocketed.

The result was, contrary to many expectations, significantly increased total IT spending. There were many business benefits associated with cloud, but it didn’t mean that total costs declined.

The coffers of hyperscale cloud providers swelled. Those industrial enterprises that did not accurately project cloud consumption curves were caught off guard. Today, tokenomics is proceeding on a similar trajectory, only faster.

By now, CFOs understand well the economics of cloud computing and how to optimize the variable cost structures they represent. A similar level of budgetary scrutiny is appropriate for AI.

Token prices are declining as a result of intense rivalry and efficiency improvements. Open-source frameworks, custom silicon and innovative development methods are driving down model costs rapidly, especially in China.

However, as companies look to integrate AI into mainstream industrial operations, plant floors and logistics networks, the absolute number of tokens generated will far outstrip gains in per-unit costs. What looks like a bargain on a per-transaction basis today will look very different as industry marches towards greater autonomy.

A Pragmatic‍‌‍‍‌ AI Approach

Moreover, data-rich environments like industrial plants are sensitive to latency in ways that consumer and enterprise applications are not. Continuous inference over a network of sensors, industrial equipment and control systems requires a very different kind of computation than running a chatbot. The data volumes are different. The infrastructure requirements are different. And the economics are different.

This should not be taken as an argument against the deployment of AI in industrial operations. On the contrary, with a pragmatic mindset about AI’s costs, we can harness the exciting innovations AI introduces and enable all-new use cases.

The good news, though, is that tokenization does not have to mean runaway consumption. Where models are invoked, how often they run, what context they draw on and how intelligence is distributed across edge, cloud and hybrid environments all affect cost.

Importantly, AI in industry is not monolithic. Lighter-weight models, edge-based processing and other techniques will continue to improve the efficiency of AI.

Tried-and-true AI-based predictive maintenance, computer vision and process optimization solutions all provide clear illustrations of mature, winning business cases. Generative, agentic, physical and multimodal AI, however, herald something different and more complex financially.

The upside of AI in industry is significant. So are the potential costs.

They should be accounted for on the same spreadsheet.

Use the best business cloud storage to manage your data.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

The European Commission has just moved to classify Amazon's and Microsoft's cloud businesses, AWS and Microsoft Azure, as "gatekeepers" under the Digital Markets Act. The designation is reserved for platforms with significant market power, even though neither meet the usual size thresholds for that label.

Regulators pointed instead to entrenched market positions, steep switching costs, and the growing weight of AI tools and partnerships in locking customers in.

This is not only a cloud computing issue. It reflects a broader shift in how the world’s largest technology companies exert influence across digital life. Their power increasingly comes not just from attracting large audiences, but from controlling the infrastructure, platforms and distribution systems on which those audiences and other businesses depend.

The same pattern is visible across the media landscape. YouTube, Google, Instagram, Facebook and Amazon rank among the most influential companies across mainstream and social media. Their lead is not simply a product of popularity. It is reinforced by their control of the channels through which information is created, distributed and discovered.

No room for new players

There's a clear divide between this top tier and everyone else, and right now nothing sits in between. The handful of brands that come closest are themselves long-established, globally dominant companies, not new entrants nearing a breakthrough, and even they trail well behind.

This suggests the market is characterized less by active competition and more by a settled hierarchy of platforms. That line separates an established class of technology platforms from the rest of the market, including some of the world's other largest companies.

Perplexity AI is the one worth watching. Its sentiment across social and traditional media already beats Google, TikTok, and Apple, carried by a chief executive who's a near-constant presence in debates about AI and search. However, its influence remains concentrated within specialist and tech communities rather than broader mainstream media.

Keep growing at this pace, and assuming other media issues don't change opinion, it could close much of the gap within a couple years. For now, every one of these platforms sits a long way outside the established top tier.

The self-reinforcing lead

McKinsey found that Generative AI was the primary source of insight for 44% of its users last year. That shifts even more influence to the companies that either build the AI systems shaping discovery, or supply the content those systems draw on.

A similar dynamic is emerging beyond Big Tech. Among leading brands, those appearing most consistently in AI-generated answers aren't necessarily those dominating current media coverage, but those most deeply embedded in AI training data.

Once a brand becomes part of an AI model's underlying knowledge, visibility starts to reinforce itself. Discovery increasingly favors what AI already knows, making it progressively harder for challengers to break through.

Regulatory pressure hasn't opened much of a door either, at least not yet: Apple and Meta have faced significant Digital Markets Act penalties, and several of the biggest platforms that carry below-average sentiment are weighed down by antitrust cases.

Those moves acknowledge how concentrated digital power has become, but they remain reactive rather than having any real redistribution of influence. The influence rankings barely shift because the underlying advantages - audience scale, data, ecosystems and AI integration - remain intact.

What would it take to shift the balance?

The evidence points to a bigger shift than simply market concentration. The platforms at the top aren't just attracting the largest audiences; they're increasingly shaping the infrastructure through which digital services are built and discovered.

AWS and Azure underpin huge parts of the internet. Google, YouTube and ChatGPT increasingly influence how information is surfaced. Apple's and Google's mobile ecosystems determine distribution.

These layers of infrastructure create reinforcing advantages across cloud, search, and device ecosystems. Those advantages reinforce one another in ways that are difficult for standalone competitors - including Reddit, Perplexity AI, Bluesky, Snapchat, and Quora - to replicate.

That's why there are so few credible alternatives today. Perplexity may continue closing the gap in AI search, and regulators may succeed in limiting some anti-competitive behavior. But neither addresses the broader reality that influence has become cumulative across multiple parts of the technology stack.

Until that changes, the question isn't whether Big Tech faces competition, it clearly does. It's whether any competitor can challenge the system of advantages that keeps the current leaders at the top.

We've featured the best AI website builder.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

• Claude Cowork is coming to Google Chrome side panels
• Claude can now connect across multiple systems and tools
• Users will start seeing the upgrade now

Anthropic has revealed it is upgrading the Claude in Chrome browser extension tool into a fully-fledged AI-powered work assistant.

The tool, designed as a side panel, will now be a Claude Cowork session, allowing users to access the company's AI tools in the browser to help crack those tricky work tasks.

"Conversations are saved to your history, your skills and connectors work in the browser, and a task you start in a tab can be finished on the Claude desktop, web, and mobile apps," Anthropic noted in a blog post announcing the news.

Cowork in Chrome

The company notes that while many of the tools that users interact with each day are already connected to Claude, others such as internal dashboards, legacy systems, and vendor portals are not.

This therefore needed a separate side panel session, which could interrupt workflows for some workers.

Connecting these via the new addition will mean that Claude can work in these apps through the Chrome browser, carrying context and conversations across your vital tools, whether that's on desktop, web or mobile.

Anthropic says the side panel will run the same Claude Cowork session being used for longer, multi-step work, such as creating a budget spreadsheet that needs to pull in invoices from multiple vendor portals.

Claude in Chrome would be able to gather up the amounts and dates needed before building the spreadsheet, allowing you to continue the session in the desktop app to add files from your computer, or import last month's budget and ask what's changed, allowing you to maintain context across surfaces as you work.

And because sessions live with your account rather than a single device, you can start work in a browser and pick it up later somewhere else.

Anthropic has also moved to address growing concerns around prompt injection cyberattacks against agents such as Claude in Chrome by adding a step where the tool will ask for a check from the user before carrying out any major action, such as submitting a form, sending a message, or downloading a file.

Users on a Max and Team plans will see the change immediately, or can install the extension from the Chrome Web Store now - it will be rolling out to Pro users over the coming weeks.

With all the buzz around nation-state threats, it’s easy for organizations to focus on threats outside the business – and forget about risks that can spiral outwards from within.

Whilst GenAI tools have introduced undeniable efficiencies for employees, these platforms have also introduced a new class of risk: two in three UK organizations admit they can’t track whether employees are sharing data via approved tools.

Most of the time, employees aren’t sharing sensitive data because they have malicious intentions. They are uploading sensitive information – like contracts, client proposals or supplier agreements to models like ChatGPT and Claude to save time on routine tasks.

Almost all (93%) of CEOs across the globe have adopted generative AI to some extent in the past 12 months (PwC). What’s concerning is that much of this activity is happening without any oversight, in the browser – meaning organizations are failing to track the flow of company information, including when and where it’s uploaded.

This is spiraling into serious risk for businesses.

First, because employees may inadvertently share credentials or other access details with public LLMs, which could result in unauthorized access if the model is compromised.

Second, uploading personal data to LLMs can trigger compliance breaches with laws like GDPR and the Data Use and Access Act – resulting in costly fines as well as reputational damage.

To take control of this issue, leaders will need to implement tools and technologies that provide visibility and control over usage at both the browser and the application levels.

The incentive problem

Employees don’t need more mandatory cybersecurity training – the problem is incentive. Many company-owned gated LLMs are still in the pilot stage, falling short of the speed and precision offered by public alternatives.

While the majority of employees understand the risks, 35% of UK businesses admit data sharing through external tools takes place – indicating many would rather ‘throw caution to the wind’ than waste valuable time using slower tools.

But the risks of this behavior – particularly in highly regulated sectors like financial services, could mean unsanctioned LLMs become 'hidden icebergs’ in an organization. Concealed, but capable of causing catastrophic damage upon impact – like inadvertently exposing customer transaction histories or credit scores.

Part of curbing Shadow AI use in the enterprise therefore starts with designing approved AI tools that integrate easily with existing platforms (for example, Microsoft 365 and Google Workspace). These tools should be continuously improved based on user feedback, ideally avoiding excessive restrictions that make the tool frustrating to use.

But the fact is, nearly two-thirds of organizations are currently stuck in the pilot stage when it comes to their AI initiatives and haven’t started to scale across the enterprise (McKinsey). So, what can organizations do today to gain control of the Shadow AI problem?

The solution – tools to bring unsanctioned AI usage under control

You can’t control what you can’t see, which is why organizations need a real-time view of who, or what, is accessing what data, from which devices, and where it’s being shared.

Next-gen identity security platforms can help organizations to gain an immediate understanding of how employees interact with consumer AI tools like ChatGPT, Claude, and Gemini, tracking interaction frequency and monitoring document uploads.

Once high-risk behavior is identified, organizations can then automate corrective actions, redirect users to secure AI alternatives, or prompt users to justify their business use case before proceeding.

Visibility will become even more important with the emergence of ‘nested’ agents. In this scenario, employees might believe they’re only interacting with a single AI agent, but that chatbot may delegate tasks to multiple underlying agents. The organization has no visibility into how many downstream agents or services its information is being shared with.

An identity security tool makes these identities ‘discoverable’ via a real-time ‘agent ledger’. This ledger acts as a complete, unchangeable trace of all agent activities and interactions. It also applies controls to each agent in the database.

Only the absolute minimum privilege required for a task is granted, at the exact moment it is needed, and for the shortest possible duration. In this way, agent permissions don’t automatically ‘cascade’. If an agent wants to connect with another agent, it must be verified by the system first.

Closing the visibility gap

Shadow AI is more than a tooling problem: it’s an identity problem. Organizations can close the ‘visibility gap’ by using tools that track interaction frequency, block sensitive document uploads, and prompt employees as well as AI agents to justify their business case before they use unsanctioned tools.

Once organizations know which tools are being used, what data they're accessing, and where that information goes, they can apply effective guardrails to secure behaviors – both human and non-human. A simple inventory of AI agents is not enough; now, organizations need to move beyond flat inventories and develop an understanding of the context and relationships that surround every agent.

In essence, identity security platforms become adaptive – moving from static to dynamic, real-time approaches to access. This is helping organizations to operationalize zero trust by ensuring that no identity, human or non-human, is trusted by default.

In the era of AI agents, securing identity has become a prerequisite for innovation.

We've featured the best AI tool.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

As governments and organizations rethink their reliance on foreign-owned IT infrastructure, data sovereignty has become a boardroom priority.

However, the conversation has become overly focused on where data is stored, overlooking the legal, operational and resilience factors that determine whether organizations are truly in control.

Whether through misunderstanding or a deliberate attempt to mislead, the term data sovereignty is often misused.

Data residency and data sovereignty are being conflated, despite being very different. Data residency is about the physical location of data, while data sovereignty is much broader and also includes legal jurisdiction, operational control, resilience, governance and the ability to manage risk.

Concerns about dependence on foreign-owned digital infrastructure have brought added urgency to issues of digital independence and control over critical technology.

In response, various vendors - particularly the big US-based hyperscalers - are now repositioning themselves with “sovereign” alternatives based primarily on where they store customer data.

While this might address some of their customers’ needs, reducing sovereignty to a question of geography creates a misleadingly simple narrative: if an organization moves data to the “right” country, it will somehow become compliant. In reality, the core issue is not just where data should be hosted, but understanding who may seek access to it and who ultimately controls the infrastructure supporting it.

This misunderstanding is giving rise to what could be described as “data sovereignty washing”, with simplified claims that don't reflect legal or operational reality.

Data sans frontières

Governments the world over have well-established legal mechanisms for requesting information held in other jurisdictions. While data residency influences which laws apply and how requests are handled, it does not provide immunity from lawful access or eliminate international cooperation.

An example is the US CLOUD Act. Under certain conditions, it enables US authorities to request data from US service providers even when it is stored outside the United States. So, even if a UK or European-owned organization hosts data with a US-owned provider in a UK or European data center, it may still be reachable under US legal process. Being physically ‘local’ doesn’t change that.

The US is far from unique in this regard. Many other countries have legislation in place allowing authorities to access data for law enforcement or national security purposes, often supported by cross-border agreements and established legal processes.

The risk is that enterprises treat location as a complete sovereignty strategy, rather than one element of it. Threat actors care about the value of the data, not geography. As a result, organizations can spend significant time and money migrating data to new locations while leaving their biggest security risks fundamentally unchanged.

A more useful starting point is to ask what risks the organization is actually trying to reduce. That shifts the focus and any subsequent changes in approach away from maps, and towards threat modelling, which provides the right context for meaningful conversations about sovereignty.

Start with the threat model

Different organizations have fundamentally different threat models. A local retailer, a multinational bank, a defense contractor and a government department are unlikely to share the same priorities, even if they all process sensitive information.

For some organizations, regulatory compliance or data residency requirements may be the primary concern. For others, resilience against cyberattack, protection of intellectual property, or reducing dependence on a particular technology provider may be far more important. It’s generally a matter of sector-specific and business priorities.

So, rather than simply asking where data is stored, leaders should consider who ultimately controls the infrastructure, how dependent they are on individual cloud providers, what happens if services become unavailable, and whether they retain sufficient visibility and control over critical systems. If any of this raises operational or regulatory concerns, it may be sensible to adjust strategy.

The resilience paradox

An unintended consequence of pursuing absolute data localization is that it can reduce resilience. Organizations often improve availability and data recovery by maintaining geographically separate copies of critical data. Restricting everything to a single jurisdiction can reduce those options.

Decisions about sovereignty should therefore factor in availability, confidentiality, and integrity – all of which are important. The most effective approaches recognize that resilience sometimes requires carefully managed distribution rather than rigid localization.

Consider this scenario: an organization has ensured all their email is stored within a single jurisdiction to meet sovereignty objectives. Months later, their provider has a major outage in that one region, and they lose access to their email for days. Even worse, a serious data loss event affects their backups, which are also stored in the same region. If they had optimized for resilience instead, their data would have been safe and available throughout.

Even though the organization successfully addressed one aspect of sovereignty, they weakened another by reducing their ability to recover critical business information. That's ultimately the difference between treating sovereignty as a marketing claim and treating it as a genuine risk management exercise.

We've reviewed, rated, and ranked the best cloud backup.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

The past two years have seen unprecedented adoption of generative AI. This was driven largely by LLM platforms such as Claude, which reported 28 million paying US customers in March.

In the corporate world and startups, this has put the pressure on companies to invest in AI, to keep up with the latest models and bolster productivity and efficiency. But now that the world has adopted AI, a stark readiness gap is emerging as governance and AI skills lag behind. AI fluency is now a baseline expectation for employees as leaders feel the pressure to prove the returns on their hefty investments.

Tokenmaxxing is the latest AI trend to come under fire as businesses want employees to use more AI in their work. In essence, this means boosting AI input to maximize AI output. On the surface, it sounds efficient and harmless. But underneath, it poses major security risks.

Tokenmaxxing explained

So what actually is tokenmaxxing?

A ‘token’ is a unit of data processed by an AI model. For example, a word or character inputted into an LLM search. So ‘tokenmaxxing’ quite simply means over-engineering generative AI prompts to get the most out of one search input. This can be anything from overly detailed prompts, to overloading an LLM chat with information, to asking an AI model for step-by-step breakdowns, as opposed to short summaries.

The trend is driven by businesses as leaders face pressure to prove the ROI of AI. It became a tongue-in-cheek benchmark of AI performance. Some companies, such as Meta, even gamified tokenmaxxing, measuring and ranking AI usage and citing the highest scorers ‘Token Legends’.

Their assumption is that AI usage means being AI-forward. But instead, companies need to consider the value they get from AI tools. They also need to keep security at the center of the conversation.

The hidden costs of tokenmaxxing

Although it might look like harmless corporate showboating, this trend poses a wide range of security risks given that LLM vendors are 52% more likely to be designated as “high risk” than traditional SaaS. This is due to access to sensitive data, IP, and internal workflows, so it’s imperative that businesses have oversight of how employees use these resources.

Rapid adoption of AI has led to an experimentation mindset. This is a positive shift from an innovation standpoint, but from a compliance perspective, a ‘trial and error’ approach is more error than trial.

The legacy tech systems most major enterprises are still reliant on are controlled by procurement and security teams and were not designed for the agility of AI technology. Meaning both innovation and compliance are lagging behind. It’s the latter that’s causing major concerns in the cybersecurity world.

When employees face mounting pressure to get the job done, they won’t wait for security teams to approve new tools, which results in shadow AI. This is where unmanaged, unapproved AI tools operate inside company environments without oversight.

Industry data shows that 70% of 16k cybersecurity customers currently have some form of shadow AI lurking within their organization, largely due to AI tools introduced through improper procurement channels that now have access to company data without oversight or guardrails.

There’s also been a 36% increase in shadow IT year-on-year, with organizations discovering, on average, around 140 Shadow IT tools accessing their environment within 90 days of connecting to the platform.

The bottom line is that AI adoption is drastically outpacing governance, and employees are prioritizing speed over control.

How businesses can defend against Shadow AI

The core issue isn’t tokenmaxxing itself; it’s businesses' inability to keep up with employee demand for speedy access to the latest AI tools. This ultimately results in friction between the desire to safely onboard new tools and the ongoing pressure to use AI.

When security teams intervene and revoke access to unmanaged tools, employees just reinstall them. Industry data finds that within a 30-day period, the average enterprise sees employees reinstall revoked tools 100+ times. Within one year, it happens 1,000 times.

To bolster defenses, organizations must design their procurement systems to match the speed of AI innovation, so they can keep up with the rate of AI usage, as demonstrated by so-called ‘token legends’.

Three actions businesses can take now

The more generative AI gets adopted in the corporate world, the more employees will face pressure to adopt and prove its ROI. Tokenmaxxing is just one hype within this wider picture. Businesses need to act fast to stop the gap between experimentation and control widening.

Three things leaders and compliance teams can kickstart today to bolster defenses against shadow AI are:

• Shrinking vendor review timelines so they match the speed of AI adoption
• Set up continuous monitoring systems to detect threats caused by tokenmaxxing before they jeopardize safety
• Implement employee training and policies for AI usage to ensure employees don’t expose sensitive data or IP

The new mandate is matching speed with governance, and it’s up to security teams to lead the charge.

We've featured the best AI chatbot for business.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit

• A 150M model reached 29.5% while costing just $0.0007 per task
• ChatGPT scored higher, yet its comparable reasoning runs cost substantially more
• BDH-CQ performs reasoning internally instead of generating lengthy intermediate text

Pathway, an AI lab focused on building Post-Transformer architectures, has released new benchmark results for its BDH-CQ reasoning model.

According to the researchers, their 150M-parameter model scored 29.5% pass@2 on the public ARC-AGI-1 evaluation set.

It achieved this at a computed inference cost of $0.0007 per task, roughly eleven times cheaper than ChatGPT's underlying GPT 5.6 Luna (Low) model.

A cheaper way to reason

Today, many AI tools waste computing power because of how they are designed, not because deep reasoning demands it.

"Today's AI pays a steep token cost for reasoning, but that cost is imposed by architecture, not by any law of intelligence," said Zuzanna Stamirowska, CEO and co-founder of Pathway.

“We show that a different architecture changes the game and opens up a whole new space in terms of how much intelligence per dollar.”

Amazon Web Services believes that BDH-CQ’s result is a promising step toward using advanced AI reasoning in real products more affordably.

"Customers are increasingly exploring how to move advanced reasoning from experimentation into production, where performance, efficiency, and scalability all matter," said Nicolas Tarducci of AWS.

ARC-AGI-1, a widely used reasoning benchmark for AI systems, checks whether a system can infer an underlying rule from limited examples and apply it correctly to new inputs.

In this test, OpenAI's Luna model scored only slightly higher at 34.2%, yet running it still costs significantly more ($0.008 per task).

That price gap already includes OpenAI's recent 80% price cut on Luna, which began on July 30th of this year.

Further up the chart, Claude Opus 5 and Gemini 3.1 Pro reach 97–98% but cost around $0.5 – $0.6 per task, meaning the frontier's very top costs close to a thousand times more than BDH-CQ for the highest scores.

On the cheap end, Qwen3 235B costs over three times more than BDH-CQ while scoring worse than even its Low variant, so it isn't a real competitor on either price or performance.

"Pathway shows that model architecture, not just scale, can drive the next leap in AI reasoning," said Łukasz Kaiser, co-author of the original 2017 Transformer paper.

Why it costs so much less

The efficiency gap stems mainly from a structural difference in how each system actually performs reasoning during inference computations.

Many reasoning AI systems generate intermediate text, adding one token after another before producing their final answers.

The longer that written reasoning becomes, the more it costs to run and the slower the AI responds to each request.

BDH-CQ works quite differently, quietly solving problems inside its own memory instead of writing everything down first as visible text.

Pathway also confirmed that early experiments already follow standard Transformer-like scaling laws across model sizes from 1B to 600B parameters.

The company also plans to extend this approach toward harder benchmarks, including mathematical reasoning, ARC-AGI-2, and eventually full ARC-AGI-3 evaluations.

If these efficiency gains hold across larger and more difficult tasks, cost rather than raw capability could increasingly separate rival reasoning systems.