• Nvidia's RTX 5080 Super was rumored for January 2027, but fresh gossip suggests it's delayed
• Meanwhile, a new report on GPU prices observes a hefty price jump month-on-month
• Even lower-end GPUs are getting more costly at a faster pace, and the high-end is a nightmare — with no RTX 5080 Super riding to the rescue, it seems

Nvidia's RTX 5080 Super has been strongly rumored of late, but there's now been a cloud of doubt cast over this potential addition to the Blackwell range — just as we've got more evidence of how dire the situation with GPU price hikes has recently become.

VideoCardz spotted the latest from the rumor mill on the RTX 5080 Super, a graphics card that supposedly packs 24GB of video RAM, with Taiwanese tech site Benchlife reporting that Nvidia has delayed the GPU (assuming it's a real product in the first place, of course).

We were told via other rumor peddlers that the 24GB spin on the RTX 5080 is theoretically planned for a January 2027 launch, but Benchlife asserts that the release date "remains unclear" and that Nvidia's board manufacturing partners have yet to receive any design details for the new model.

Given that, according to Benchlife, the "probability is extremely high" that the RTX 5080 Super won't debut in January, considering that board makers were expecting to receive the mentioned design info back in October.

This pushes the rumored GPU back at least a month, very probably more, and the idea of that is hardly surprising given the ongoing RAM crisis, which has been getting worse as 2026 rolls on.

That affects video RAM, and it's the main driver of price spikes with graphics cards, which are starting to get out of hand, as evidenced by TechSpot's latest monthly report highlighting the current cost of GPUs across various regions.

What TechSpot's monitoring data for October 2026 shows is that, as we already knew, high-end pricing has gone even further off the rails. The RTX 5080 is up 39% in the US and 40% in Germany month-on-month, for example, which is ludicrous, frankly — in the US it's now 120% over the MSRP at $2,200. The UK has seen an 8% hike over the past month, with this GPU now 40% over its MSRP in the country. Globally, it's 77% over Nvidia's recommended price on average.

The RTX 5090 is even sillier, as we've seen for some time now, and stands at $9,400 in the US in October, some 370% over the already very expensive MSRP. If you can find the Blackwell flagship at all, that is.

What's really worrying is that TechSpot's data shows lower-tier GPUs are now seeing some major price bumps. Want a humble RTX 5050 in the US? It's gone up 35% over the past month (an increase of over $100), so you'll now pay 68% over the MSRP.

The RTX 5060 is creeping up in price in some European countries, too — it's up 7% in the UK and 13% in Germany. The RTX 5060 Ti 8GB has gone up more, with this GPU getting hiked by 10% in the UK, so it's now £440 compared to £350 back at the start of 2026. The 5060 Ti 8GB went up 6% in the US, too, so it's now 40% over the MSRP in total.

There are hikes across the board for Nvidia, and AMD's Radeon GPUs are getting more expensive at an even faster rate currently. Again, lower-end graphics cards are suffering, such as RX 9060 XT models being up 10% on average month-on-month. The RX 9070 XT is also up 10% in the US since last month, and the increase in Canada has been a sickening 30% — in just one month.

Analysis: a deteriorating situation for GPUs

What does all this tell us? That spiralling graphics card prices are very real, with some huge ongoing hikes for higher-end models — but this isn't just a top-end GPU problem any longer. The real concern is that lower-tier boards are also seeing more price inflation, and I don't think this is a problem that's going away.

Indeed, based on the weighty amount of chatter about the memory crisis continuing to get worse in this final quarter of 2026 and into next year, there's every reason to believe that GPU prices could keep going up. That's why I've been advising PC owners thinking of a GPU upgrade to treat Black Friday as a prime (pun not intended) opportunity to get in and buy a graphics card before any more hikes.

With the RTX 5090 reportedly effectively discontinued — something Nvidia strenuously denies, but the stock situation speaks for itself in countries like the US — and the RTX 5080 pricing itself well out of the range of the vast majority of consumers, those who were anticipating the RTX 5080 Super to ride to the rescue have suffered a fresh blow here, too.

I never thought that was a realistic proposition anyway, mind, but still, it's clear enough that the GPU market is in trouble, and Nvidia seemingly doesn't have any answers in the short term for the upper echelons of its Blackwell range. As noted, though, the real worry is that there's more pain coming to the more affordable end of the graphics card spectrum, where mainstream buyers make their purchases.

Meta might be building a 4.5ft-tall, 106lb humanoid robot that could look you in the eye, mimic human expression, and be a help around the home — a description that would fit almost any consumer-grade, humanoid robot from almost half a dozen companies. The report from The Information adds, though, that this bot will be " a physical embodiment of Muse."

Rarely has anyone put a finer point on why companies like Tesla, Meta, Unitree, Figure, X1, and Boston Dynamics are now racing to deliver home companion, assistant, and janitorial robots into homes.

These robots will essentially be Trojan horses for the increasingly unpopular AI systems these robotics companies, and their frontier and LLM partners, desperately need consumers to adopt to justify the hundreds of billions of dollars they are pouring into every part of the AI sector.

In the US, we sit on the precipice of a midterm election in which AI and data centers are, mostly indirectly, on the ballot. Voters are all but asking for an AI litmus test for candidates: where do they stand on the rapid expansion of data center construction? Obviously, it's not just the data centers. Voters are concerned about AI's role in misinformation, infrastructure security, as well as its role in the rapid deconstruction of the job market and workplace.

In short, AI is, for many (though certainly not all), if not a villain, then certainly something to be examined critically.

If more than half of Americans are at least "concerned" about AI, companies like OpenAI, Meta, Claude, and others may soon face a backlash that results in people pulling away from the use of chatbots and AI agents like Meta Muse.

AI and robotics — perfect together

Perhaps robotics faces some of the same challenges. Science fiction has played its role in making many fear the robot uprising. Still, social media is lately filled with humanoid robotics feats of dexterity, dancing, and even terrifying fighting skills. People might be worried about Terminator, but can't help but be intrigued by a tireless android ready to take on all your least favorite chores.

It can be easy to forget that AI is the thing that's supercharging robotics development or that Nvidia, the most valuable company in the world, is the one most responsible for the chips powering the AI inside many data centers.

Partnerships between AI and robotics companies are everywhere. Meta's rumored robot is reportedly using Nvidia's Jetson Thor chip. That platform can also be found in Figure AI robots and Agility Robotics, among others. Google's DeepMind is partnered with Boston Dynamics. Over the years, Amazon has bought its share of robotics companies, often redirecting them to work in its vast supply chain and inventory operations.

For any of these companies, though, I think it's safe to assume that the potential home robotics revolution is an attractive target for their AI aspirations.

Even Apple has reportedly looked to combine robotics with Siri AI. Its new HomeView 7-inch desktop display will arrive with Siri AI baked in, and a rumored follow-up might include an animated head that can look at and track you as you, say, move about the kitchen and consider your next Siri AI prompt.

What's really inside your robot

If any of these companies succeed in integrating home-targeted AI systems and delivering them in useful humanoid robotics, it will not only be a game-changer for consumers, but will also open the door for what I assume is the ultimate goal: locking consumers into long-term AI and robotics subscription programs.

After all, most of these robots will be far too expensive for any consumer to buy outright, but monthly leasing fees that might wrap in model access and even AI tokens would likely appear more approachable for most consumers.

Instead of paying for access to Meta Muse, Claude, ChatGPT, Alexa+, or SiriAI, they'd be paying for cleaning and assistant services that come with their friendly neighborhood humanoid maid.

Getting real about humanoid robots

Assuming that's the plan, it suffers from one major flaw: the readiness of any of these humanoid robots for the home.

Don't be misled by those aforementioned viral videos. Boxing robots, for instance, are almost entirely teleoperated. Most of the best stuff many of these robots do is still via remote control (or is canned programming and not true autonomy).

I'm not arguing whether these robots can walk, run (quite fast), and dash up a flight of stairs, but they remain unpredictable and, I think, potentially unsafe in the average home.

That will, of course, change. Five years from now, the first safe and reliable humanoid robots may be ready to clean your home, walk your dog, and even look after you and your children. And you'll be paying pricey monthly fees that will funnel back to AI companies building more data centers and developing AIs that are, themselves, building better models.

The robots are coming, and they're bringing the future of AI with them. Just be ready to pay for it.

• Microsoft fixes long reboots, screen flicker, and unexpected restarts on the Surface Laptop Ultra
• Launch reviews revealed slow restart issues even as most of the flicker and crash reports came directly from owners
• No update can fix the Mini LED panel's slow response, grayish blacks, and weak dimming, or RTX Spark's efficiency limitations

Microsoft has shipped the first firmware update for the Surface Laptop Ultra just six days after its RTX Spark-powered flagship went on sale, and the changelog reads like a tidy summary of launch-week grumbles being ironed out.

The fix, which was listed in Windows Update as Microsoft Corporation System Hardware Update (200.0.7.0), resolves issues that could cause extended reboot times, screen flickering, and unexpected restarts, and includes a line promising additional bug fixes and performance gains.

The rollout is staggered, so not every machine will see it at once, and Microsoft warns that firmware updates can't be uninstalled or reverted once applied, and owners can check for it in the Surface app under Help & support.

Six days, three fixes with more unfixable issues to address?

Microsoft's speed in addressing the issue seems excellent, but it has more than a few teething issues under the hood.

To its credit, this is Microsoft's first Surface with Nvidia's custom silicon and its first with a MiniLED screen, so some teething trouble was always likely, and a patch inside a week suggests the Surface team is watching closely.

Zac Bowden of Windows Central, whose review complained about restart times, reports that the reboot problem is now fixed and that the update also targets blue-screen crashes, though he notes those were "something I am yet to encounter on Surface Laptop Ultra anyway." Beyond the unexpected restarts, some reviews noted the flickering issue, while others reported no such issues on their units.

These are, however, sadly not the only issues in play: reviewers have also flagged ghosting, battery life, and limits on app and game compatibility on Microsoft's Arm-based platform.

The display is where the issues may be most pronounced: Notebookcheck measured a 57 ms black-to-white response time, visible ghosting, and a 0.61-nit black level that leaves dark scenes looking grayish, though its 4,650 Hz PWM backlight frequency should be high enough for most flicker-sensitive users. Its review also reported no local dimming at all, which prompted Microsoft to tell Windows Central that dimming is active by default, with 5,400 Mini LEDs spread across 1,344 zones.

Windows Central found that it depends on the "Change brightness based on content" setting and the Enhanced display profile, and that it works best with HDR off. The catch is that HDR off makes ghosting and blooming more visible, while HDR on reduces them at the expense of dimming. Its review summary seems to capture it quite well, stating: "It's just not very good local dimming."

Battery life is another key issue: at Computex in June 2026, an Nvidia executive told TechRadar to "expect all-day battery life" from RTX Spark laptops in regular workloads, and Microsoft rates the Surface Laptop Ultra at up to 15 hours of local video playback. Linus Tech Tips measured about nine hours in a 1080p playback test at 200 nits and 60Hz, according to Notebookcheck, the same as a 14-inch MacBook Pro whose battery is 27% smaller, and just 53 minutes under a GravityMark stress load.

That is roughly six hours shy of Microsoft's own rating, though the test conditions aren't completely identical, even as Windows Central's own runs landed at around 14 hours using the Procyon benchmark and closer to 10 in real-world use.

Either way, efficiency is a property of the silicon, not something a System Hardware Update rewrites. Some owners have also reported the laptop running warm under light loads, a topic Microsoft has not yet directly addressed.

A beta tester for Nvidia silicon or a first mover with an AI advantage?

Any decision when it comes to buying the Surface Laptop Ultra comes down primarily to price; it starts at $2,599.99 for an 18-core chip with 24GB of memory and 512GB of storage, and tops out at $5,899.99 for the 20-core, 128GB, 1TB build, which our review roundup notes already carries lead times into late December.

The roundup's verdict is also the clearest buying test available, and it still holds true: "Grab it if local models above 24GB are your daily work." Other outlets that tested local AI models put the Surface ahead of Apple's M5 Max and AMD's Ryzen AI Max+, while gaming and single-threaded work seem to be middling at best.

Waiting for more patches only makes sense for software problems, and there are a few. Microsoft told Windows Central that Copilot+ PC features aren't ready for the platform yet and should arrive in the coming months. The laptop ships in the "Best Power Efficiency" mode, which Windows Central found could cut performance by a quarter or even half until you switch to another preset.

To Microsoft's credit on the software front, the device's support runway is long, with driver and firmware updates promised until October 16, 2032. Microsoft is also running a somewhat stingy trade-in promotion that adds up to $1,000 in cash back on top of a qualifying MacBook Pro's value for orders in the US and Canada through November 23 2026, but Windows Latest notes that trade-ins are final, even though the Surface itself can be returned until January 31, 2027.

If the screen is a significant factor, other RTX Spark machines ship with OLED panels. Asus's ProArt P16, for instance, earned praise in our review for its vibrant OLED display and Nvidia silicon, though the reviewer also judged its price a tough sell. This isn't a far cry from earlier coverage suggesting that, before launching, the RTX Spark would sink or swim on price-to-performance, and this first patch doesn't change that equation much.

The answer therefore is one that might be more nuanced than just a simple yes or a no: If you run models larger than 24GB locally every day, this update removes a couple of irritants from what reviewers rate as one of the better Windows options for that work, and buying now, inside a return window that stretches to the end of January, is reasonable.

If you were hoping for a MacBook Pro-beating screen or all-day battery under load, save your money: the problems that held back the reviews are baked into the hardware, and nothing in Microsoft's release notes suggests that it can fix all of them in the near future, if at all.

• Nvidia has quietly cut Rubin's listed HBM4 bandwidth from 22 TB/s to 19.2 TB/s per GPU and NVLink 6 from 3.6 TB/s to 3 TB/s
• The eight-GPU HGX Rubin NVL8 fares worst, losing 26% of its memory bandwidth and dropping from 2.3 TB to 2 TB of HBM4 capacity
• On paper, AMD's MI455X now leads on per-GPU memory bandwidth and Helios on scale-up bandwidth, though neither side's figures are benchmarked

Nvidia has quietly lowered the memory and interconnect bandwidth it publishes for Vera Rubin, cutting per-GPU HBM4 bandwidth by about 13% and NVLink bandwidth by about 17%, without a press release or explanation.

The Vera Rubin NVL72 product page now rates each Rubin GPU at 19.2 TB/s of HBM4 bandwidth, down from 22 TB/s, and 3 TB/s of NVLink 6 bandwidth, down from 3.6 TB/s.

Memory capacity holds at 288 GB, and the headline 50 PFLOPS of sparse NVFP4 inference remains unchanged; however, at rack scale, aggregate memory bandwidth slips from 1,580 TB/s to 1,400 TB/s and the NVLink total from 260 TB/s to 216 TB/s, while the two-GPU Vera Rubin Superchip drops from 44 TB/s to 38.5 TB/s.

An edit without an announcement?

The revision surfaced in a report from IT Home, which compared Nvidia's October documentation with a copy saved in June 2026. Wccftech then narrowed the window with Wayback Machine snapshots, placing the edit between September 1 and September 11. Nvidia, as Hardware Busters observed, went "the quiet way: by editing the product pages."

The timing matters here: the change landed shortly after Nvidia's August second-quarter results said Vera Rubin racks were already running at partners including CoreWeave, Google Cloud and Microsoft Azure, months after early systems began reaching select customers, and just as the fall production shipments Nvidia promised in May 2026 were due to begin.

This suggests the new figures describe hardware going out the door rather than a distant goal, and the nerfed figures sat on Nvidia's site for at least six weeks before anyone reported them.

The downgrade wasn't completely proportional, though, with some configurations taking a bigger hit.

Most noticeably, the HGX Rubin NVL8, Nvidia's standard eight-GPU server board, fares worst. The HGX page now lists 2 TB of HBM4 at 130 TB/s for the whole system, down from 2.3 TB and 176 TB/s, a 26% bandwidth cut.

NVLink Switch bandwidth falls from 28.8 TB/s to 24 TB/s, and dense training figures slip from 280 to 270 PFLOPS of NVFP4 and from 140 to 130 PFLOPS of FP8. Scale-out networking is the only line that rises, by 12.5% in most configurations, and Nvidia hasn't said why.

Scaling back to ensure delivery?

Interestingly, the math does not hold completely anymore: its per-GPU table still advertises 288 GB at 22 TB/s with 3.6 TB/s of NVLink, figures that no longer add up to the system totals printed beside them, and Nvidia's July Rubin architecture deep dive still promises 22 TB/s. Split across eight GPUs, the new totals come to about 16.25 TB/s and roughly 250 GB each.

This fits two potential readings: eight smaller HBM4 stacks running at just under 8 Gbps per pin, the industry baseline, or one of the usual eight stacks disabled per GPU, a trick Nvidia used on the 80 GB H100, but Nvidia hasn't said which approach it took.

The retreat makes more sense given the history here: JEDEC's HBM4 standard, published in April 2025, specifies up to 8 Gb/s per pin across a 2,048-bit interface, or 2 TB/s per stack. Nvidia wanted far more, but it might have had to settle for less.

According to a TechPowerUp report from March, Rubin's per-GPU target climbed from 13 TB/s at GTC 2025 to 20.5 TB/s by that September and 22 TB/s at CES 2026, and a SemiAnalysis note it relayed said SK hynix and Samsung were struggling to reach the final figure, with early shipments expected to be nearer 20 TB/s.

TrendForce reported in September 2025 that Nvidia was pressing suppliers for 10 Gbps per pin as AMD readied Helios, and warned that if volumes fell short or costs ballooned, Nvidia "may abandon the upgrade or segment its platform products."

By January, TrendForce said the request had passed 11 Gbps and sent all three HBM makers back to redesign, though an Nvidia spokesperson told Tom's Hardware that "our HBM4 memory partners remain on track for production shipments." SK hynix finished HBM4 development first, and Samsung began shipping in February at 11.7 Gbps, above Nvidia's stated requirement.

More than a year after TrendForce's warning, the revised sheet reads like its fallback scenario: a lower overall ceiling, plus a further-reduced HGX tier, and this seems to be because Nvidia needs to deliver solutions to its customers as a higher priority than chasing an elusive memory bandwidth goal.

Team Red for the win?

Because of the retreat, perceived or otherwise, the competitive picture shifts considerably, at least on paper. AMD's Instinct MI455X spreads 432 GB across 12 HBM4 stacks for 23.3 TB/s, which needs only about 7.6 Gbps per pin, below the JEDEC baseline. AMD bought its bandwidth with width rather than speed, so it never needed above-standard memory.

When it launched Helios in July 2026, AMD claimed 6% more HBM bandwidth than Vera Rubin NVL72's published figure; against the revised 1,400 TB/s, Helios' 1.67 PB/s is roughly 19% ahead. Scale-up bandwidth, Nvidia's traditional stronghold, was level at 260 TB/s per rack and now favors AMD by about a fifth, while the scale-out increase trims AMD's claimed 50% lead there to about a third.

For buyers running inference, the bandwidth line matters more than the FLOPS line. Nvidia's engineers write in that July deep dive that decode is bound by the memory subsystem, though they argue that what counts is how efficiently kernels use it, and that it is "not about peak bandwidth specs."

The marketing hasn't moved, however: the NVL72 page still promises up to 10x more tokens per megawatt and one-tenth the cost per million tokens versus GB200 NVL72, footnoted as subject to change, and Nvidia hasn't said whether it reran those projections on the lower figures.

None of this makes Rubin slow. At 19.2 TB/s, it still offers about 2.4 times a Blackwell GPU's 8 TB/s, rather than the 2.8x Nvidia's blog still claims. Jensen Huang said in August that "Vera Rubin, now in full production, was built to power exactly this moment." The moment, it turns out, arrived with a little less bandwidth than advertised.

• Nvidia reveals plans to invest $3 billion in Meta’s share offering
• Meta could raise $95 billion to finance its expanding AI infrastructure
• Nvidia hardware will support Meta’s next generation of artificial intelligence agents

Nvidia has revealed plans to invest $3 billion in Meta’s share offering while both companies increase cooperation on artificial intelligence systems designed to operate as personal agents.

The agreement will give Meta access to additional Nvidia computing technology as Facebook’s parent company develops its Muse models and associated artificial intelligence services.

It also includes discussions about agent supervision technologies, although the companies have not disclosed a timetable for implementing every proposed collaboration.

Nvidia hardware will power Meta’s next generation of AI agents

Meta expects to install more Nvidia equipment, including its forthcoming Vera Rubin platform and Vera central processors, to support training and execution of advanced agent systems.

The companies also intend to improve performance throughout the computing process, with particular attention to processing efficiency, system throughput and infrastructure operating expenses.

Networking equipment will form another part of Meta’s planned deployment, including Nvidia Spectrum-X technology and BlueField-4 components for handling communication between computing resources.

The companies are considering Nvidia’s OpenShell architecture for safer agent deployment, alongside Sentry technology intended to track agent activity and support data processing.

“AI agents represent a new computing paradigm — software that understands our goals, reasons, uses tools and gets work done on our behalf,” said Jensen Huang, founder and CEO of Nvidia.

“This changes how we use computers, from operating applications to working with intelligent agents.”

Meta claims its Muse systems have achieved strong results in benchmark testing, while consumer downloads have climbed since their initial release.

“Muse is how we bring personal superintelligence to everyone. We’re excited to keep working with Nvidia to deliver the best personal agents and frontier models for billions of people,” said Mark Zuckerberg, founder and CEO of Meta.

However, the companies have not supplied specific download totals or independent comparisons that would establish how widely these models outperform competing products.

Meta’s multibillion-dollar fundraising supports its AI expansion

Nvidia’s investment comes as Meta prepares to raise substantial capital through share sales intended to finance its expanding artificial intelligence infrastructure.

Meta initially outlined a fundraising programme worth approximately $35 billion, comprising $22.5 billion in ordinary Class A shares and $12.5 billion in convertible depositary securities.

The company subsequently increased its ordinary share sale to $25 billion, with each Class A share offered at $625 and depositary securities priced at $50 per share.

Its broader fundraising plans could generate up to $95 billion through successive offerings, though the full amount is not guaranteed.

Nvidia’s proposed $3 billion contribution would make it a participant in Meta’s financing, linking its investment directly to a major customer’s infrastructure spending.

The companies have not disclosed how much additional Nvidia equipment Meta will purchase under the expanded partnership or when all planned deployments will occur.

• FADU's File-Semantic SSD runs path lookup, permission checks, and file-to-flash mapping in firmware instead of on the host
• NFSv3 requests ride inside standard NVMe commands, so GPUs, DPUs, and remote clients can reach files without requiring a host file system
• Its headline 460,000 ops/s result came from a PCIe 5.0 host, with PCIe 6.0 results promised by Q1 2027

South Korean controller designer FADU has shown off an SSD that understands files rather than just blocks, betting that the next bottleneck in AI storage is not raw bandwidth but the CPU time spent working out where data lives.

The File-Semantic SSD (FS-SSD), unveiled at the OCP Global Summit, moves file lookup, access checks and file-to-address translation into firmware on FADU's sixth-generation controller.

FADU's whitepaper promises GPUs, DPUs and network cards a route to their files with "no host filesystem, no block layer, and no storage-server CPU in the path." Partner samples are not expected until Q3 2027.

Not the first approach to the 'CPU problem'

FADU is no newcomer to data center flash. Its Gen5 controller sits in the SN861, which Western Digital launched and SanDisk now sells; last year, the company said it had signed controller supply deals with two of the four hyperscalers, weeks after revealing its Sierra PCIe 6.0 controller.

Its pitch here, however, isn't entirely new, with the recognition coming from a not-so-surprising front: Nvidia. Nvidia has been doing the same from the other end of the PCIe bus. GPUDirect Storage removed the CPU from the data path years ago, letting drives move data straight into GPU memory, but host software still decides what to fetch and issues every request.

Its decision to open source cuFile and shift that control path onto the GPU, with SCADA attacks adding accelerator-side overhead even as FADU attacks it from the drive end of the equation.

FADU says its own measurements show the host file system and block layer consuming about 30% of CPU cycles in a GPUDirect Storage random-read workload, a vendor figure that has not been independently reproduced. It also argues that existing bypass schemes only get the CPU out of the way by abandoning files altogether.

How does this work?

Rather than invent a new language for its drive, FADU borrowed an old one. NFS, short for Network File System, has let computers open files stored on other machines for decades. FS-SSD tucks NFS requests, such as looking up a file by name or reading part of it, inside the standard NVMe commands that SSDs already use.

Those commands can come from inside the same server or, using a networked version of NVMe called NVMe over Fabrics (NVMe-oF), from other machines in the data center. FADU chose NFS version 3 for three reasons: it is small, with just 21 operations; it works with any operating system; and it is stateless, meaning each request carries everything the drive needs to answer it, so the drive never has to remember who has which file open.

FADU has also merged two lookups that normally happen separately. On a regular system, the computer’s file system first works out which numbered block holds each piece of a file, and the SSD’s own internal software, called the flash translation layer, then works out where that block physically sits on the flash chips.

FS-SSD does both jobs with a single file layer based on F2FS, a Linux file system designed for flash memory, laid out to match how the chips are physically organized. One map then takes the drive straight from a file to the spot where it is stored, and that map is much smaller. A typical SSD carries about 1GB of DRAM for every terabyte of flash and fills almost all of it with its block map. FS-SSD uses the same amount, but FADU says about 95% of it is left free to keep details of frequently used files close at hand.

That layout also makes cleanup simpler. Flash memory cannot overwrite old data in place, so drives regularly move the data they still need elsewhere and wipe whole blocks clean, a chore known as garbage collection.

When F2FS runs on an ordinary computer, the file system and the SSD each do their own version of this cleanup, and the two end up multiplying each other’s work. FS-SSD runs just one, and FADU says that cuts write amplification by 63% compared with F2FS on a regular host. Lower write amplification means less wear on the flash, but the 63% figure comes from simulated tests on NVMeVirt, software from Seoul National University that imitates an SSD, rather than from a real drive.

In FADU’s own preliminary test, remote file throughput held at about 460,000 ops/s as the server was cut from 16 cores to one, while a standard Linux NFS server fell from roughly 150,000 ops/s to 40,000.

The testing has its limits, though. FADU’s Gen6 drive sat in a PCIe 5.0 host, so every sequential read result topped out near 14.1GB/s, roughly the Gen5 link limit, even as Gen6 drives such as Micron’s 9650, rated for 28GB/s reads, are already in mass production.

Both the client and server ran Intel’s Core i5-13600K desktop chip with 128GB of DDR5 and a 100Gb ConnectX-6 NIC. That processor has 14 physical cores and 20 threads, so FADU’s test likely counted threads rather than physical cores.

The world’s first file-semantic SSD is an idea that isn't entirely new

FADU calls FS-SSD the world’s first file-semantic SSD, and it may well be the first hardware demonstration of its kind, but the idea isn’t new. The DevFS research paper, presented in 2018, described a direct-access file system “embedded completely within a storage device,” though the authors tested it only as an emulated prototype.

NVM Express has standardized a key-value command set that addresses data by key rather than block address, and SK hynix has developed a key-value computational storage drive with Los Alamos National Laboratory. Unlike the key-value command set, FS-SSD has no standard behind it yet. FADU has not submitted a file-semantic command set to NVM Express or OCP Storage, so for now every client needs FADU’s own software.

FADU’s longer-term goal is a flash enclosure with nothing in front of it but a PCIe switch and a DPU (a network card with its own processors) or NIC handling NVMe-oF in hardware, with a reference design built on Xsight Labs’ 800G E1 DPU due by the end of 2027.

FADU’s model puts that at about 42% lower power and 57% lower cost per 16-drive system than a conventional storage server, though those are projections rather than measurements. DPU-fronted flash boxes already exist, such as Supermicro’s BlueField-powered JBOF, but those still run the storage software on the DPU’s Arm cores. FADU wants the drives themselves to serve the files.

FADU chief executive Nam Yi-hyun said the company will use FS-SSD as “a springboard to become a leading company in AI storage solutions,” according to The Herald Business. The bottleneck FADU is chasing is real, and building on NFSv3 rather than inventing a new protocol gives it a credible starting point.

For now, though, the evidence amounts to a Gen5 test bench, an emulated write-amplification figure, and modeled savings, with OEM and ODM talks still early. The Gen6-host results FADU expects by Q1 2027 will be the first real sign of whether the drive can do the most important part: handling the CPU’s job at full speed.

• Nvidia wants to turn Germany's retired nuclear sites into AI data centers
• Germany's nuclear shutdown leaves dozens of former power stations awaiting dismantling
• Nvidia and PreussenElektra are investigating new uses for Germany's retired nuclear facilities

Nvidia has agreed to examine whether shuttered German reactor sites could become computing hubs, though the company rules out restarting any idle units.

The US chipmaker signed a preliminary agreement with PreussenElektra GmbH (PEL), a utility from Germany, to study the idea at decommissioned locations.

The arrangement does not involve generating electricity from fission for any future facility, since the work concentrates on existing grid links and water access.

Reusing sites without reviving reactors

Following the Fukushima incident, Germany permanently closed its final operating reactors in April 2023 and is now dismantling about 30 atomic stations nationwide.

PreussenElektra oversees the closure and teardown of eight of those stations, and these are probably the locations that will receive scrutiny.

Neither company has named the sites under review, so the eight locations managed by PreussenElektra remain an assumption rather than a confirmed list.

“Our former power plant sites have valuable industrial infrastructure and offer opportunities for new, future-oriented uses,” said Guido Knott, chairman of PEL’s management board.

“Together with NVIDIA, we intend to conduct a sound assessment of whether and how these conditions can be leveraged for AI factories.”

The assessment will review grid connections, open land and industrial infrastructure to judge whether each site satisfies the strict operating requirements of AI factories.

The two companies say they will test whether the sites can meet engineering and running needs while supporting Germany's computing capacity for AI.

Cost savings and an unproven case

New data center facilities normally need well-located land, industrial zone approvals, access to water for cooling, and powerful grid connections, which former atomic stations already possess.

Nvidia hopes that reusing these locations could lower expenses and shorten development schedules compared with constructing new campuses during an unprecedented expansion of AI infrastructure.

“AI factories require access to power, land and high-capacity infrastructure, and former industrial sites could provide a strong foundation for this next generation of computing,” said Markus Hacker, senior regional director at Nvidia.

“Together with PreussenElektra, we will assess how its sites could meet the technical and operational requirements of AI factories and contribute to Germany’s high-performance AI infrastructure.”

Both companies describe the effort as exploratory, and neither has disclosed funding figures, sanctioned designs or binding building contracts, so the idea remains unproven.

Reusing atomic stations looks efficient on paper, yet neither company has shown that grid capacity at the sites matches the power appetite of AI clusters.

Approval processes for dismantled facilities, local acceptance and the cost of conversion remain unaddressed, so any savings in time or money are still speculative.

The refusal to revive reactors may ease political concerns in Germany, although the announcement does not say whether regional authorities or residents have been consulted.

Via Tom's Hardware

Asus ProArt P16 RTX Spark: two-minute review

The Asus ProArt P16 RTX Spark is one of the new generation of laptops powered by Nvidia’s RTX Spark chip. This all-in-one system-on-a-chip is packing 20 ARM CPU cores, 6144 CUDA GPU backed up by 64GB of RAM and promises to offer fantastic AI, creative and gaming performance in a lightweight, portable form factor. So how does this new SoC version of the Asus ProArt P16 laptop fare?

When powering up the Asus ProArt P16 RTX Spark for the first time, perhaps the first thing I noticed is its huge, vibrant 16-inch OLED screen. With a peak brightness of 1,100 nits and a 100% DCI-P3 color gamut, it looked gorgeous whether I was loading up my digital artworks or watching movies. And its 120Hz VRR makes it perfect for playing games on.

That’s not the only part of the ProArt P16’s design that feels premium. Its overall build quality feels suitably luxe, with the smooth metal of its upper casing feeling wonderfully tactile. Inside, its keyboard is well designed and feels pleasant to type on, while its haptic trackpad is precise and gives snappy feedback, even if I did find it too easy to trigger the right click at first. I will say its lid has a little bit too much wobble for my liking though: if I’m honest, I’d expect a laptop at this price to have a slightly more stable hinge.

But let’s not just focus on the superficial here: the ProArt P16 RTX Spark has a lot going on under the surface. That’s predominantly thanks to that Nvidia SoC, which enables it to tackle a wider range of tasks than even many of the best laptops.

Perhaps the most attention grabbing feature is its AI capabilities. Designed for hybrid workflows, the RTX Spark makes it easy to use a combination of local and cloud models to run your AI agents. Testing it out with the Qwen 3 8B open-source local model, I certainly found it had pretty snappy performance for basic workflows, with a basic data analysis only taking about 12 minutes.

More involved tasks took much more refining and prompting over multiple sessions though — attempting to automate 3DMark benchmarks took many 75 minute runs before the completion time reduced significantly. Still, being able to run these workflows on such a lightweight, portable machine was pretty impressive

Another area the Asus ProArt P16 RTX Spark delivers potent performance is during creative tasks. Running multiple PugetBench benchmarks for automating Photoshop and Premier Pro workflows, the ProArt P16 was only a little behind Apple’s M5 chip when running image-based tasks and beat it when it came to video tasks. And this was borne out by my own testing: even my largest 4K 300dpi Photoshop file crammed full of clipping masks didn’t make it wince, with filters and generative fills executing smoothly. Even downscaling the whole image to 2K only took around five seconds, which isn’t bad for a nigh-on 2GB file.

Few SoCs to date have been designed with gaming performance in mind. But given Nvidia’s heritage in this space, it’s hardly surprising that the ProArt P16 RTX Spark is capable of some serious gaming power. While it’s not quite able to deliver at the same level as RTX 5080 and 5090 GPUs, I enjoyed a very smooth 60fps on Cyberpunk 2077 on Raytracing Medium settings, while Gears of War: E-day ran at more solid than bedrock at over 110fps. So this laptop can achieve some stunning results, even if the performance you’ll get isn’t 100% the top of the range.

Given it can handle all of these tasks on the go, you’d hope it has a pretty capacious battery to back up that powerful chip. And it’s relatively good news on this front. While it only managed 11 hours of continuous video playback during my testing, during productivity based work it lasted well over 14 hours. That’s nowhere near as long-lasting as the Methuselah-like MacBook Pro M5 Pro, which clocked over 21 hours in our testing, but it’s certainly enough to get you through a standard day’s work.

Honestly, the Asus ProArt P16 RTX Spark is pretty adept at everything it tries to do: it’s a talented all-rounder. The only caveat really is its price: like many of the RTX Spark laptops, it isn’t cheap. Starting from $4,399.99 / £3,899.99 / AU$7,999, it costs just as much as a high-end gaming laptop or a pro-level creative workhorse — so, honestly, if you care more about getting the best performance you can in a specific niche, you may be better off investing in one of those instead.

However, if you need a single bow that can handle every single one of your strings, the Asus ProArt P16 RTX Spark is a good shout. It’ll handle running your own AI agents, whipping up creative work and even gaming for your downtime, all in a single, portable package.

Asus ProArt P16 RTX Spark review: price & availability

Having launched on October 9, the Asus ProArt P16 RTX Spark is available now. Unlike its smaller sibling, it ships with the higher-end Nvidia RTX Spark N1X chip.

As a result, it’s a pricey meatball: its list price starts from $4,399.99 / £3,899.99 / AU$7,999 for the spec with 64GB RAM, 3K screen and 1TB SSD. That increases to $6,999.99 / £5,999.99 / AU$10,999 if you want the version with 128GB RAM, a 4K screen and a 2TB SSD.

Even with its impressive specs, there’s no getting away from the fact that this is a premium-priced laptop. Both in terms of price and spec, the ProArt P16 RTX Spark is pretty similar to the Samsung Galaxy Book6 Ultra with Nvidia GeForce RTX 5070, which will set you back $3,799.99 / £3,649 (around AU$5,410).

Alternatively, this kind of spend could net you the MacBook Pro M5 Pro 16-inch with 64GB for $2,999 / £3,999 / AU$6,299, which should give you a lot of performance and creative power but I wouldn’t expect to get quite as much gaming power for your buck.

Asus ProArt P16 RTX Spark review: Specs

Asus ProArt P16 RTX Spark review: design

• Gorgeous 3K OLED display
• Slim, sleek design
• Slightly flimsy hinge

Like all of the initial lineup RTX Spark laptops I’ve seen, the new Asus ProArt P16 is a sleek-looking unit. While its 16-inch span is pretty expansive, it’s still light in terms of weight, clocking in at 1.77 kg compared to my 16-inch MacBook Pro M4’s 2.14kg. It’s also got a pretty sleek silhouette — unlike some of the RTX Spark laptop models I’ve seen, it’s not concealing a much chonkier base with the help of gradually sloping bevels. And its nano black and neo white colorways feel suitably premium, although I marginally prefer the black version I tested here.

It’s not all about style though: the ProArt P16 RTX Spark has plenty of substance too. Manufactured from smooth metal, its external and internal surfaces feel wonderfully tactile, although I’d say the inner ones feel a tiny bit less sturdy than the solid bodywork of the outside. On top of this, there’s a tiny bit of wobble to its screen when you touch it — I’d expect a slightly firmer hinge on a premium laptop like this.

Speaking of that screen, I found Asus’s 3K touchscreen Lumina Pro OLED to be pretty impressive on the whole. Not only can my unit hit a peak brightness of 1,100 nits and a max 120Hz VRR but comparing it to my MacBook Pro M4’s Liquid Retina XDR display while watching Pixar’s Elemental was — sorry for this — pretty illuminating. While the latter can go slightly brighter at 1,600 nits peak, I didn’t find the difference all that pronounced. And the colors of the P16 RTX Spark’s OLED panel were noticeably more vibrant, with the reds, blues and greens of the fire, water and earth elements really popping.

To say laptop keyboards can be hit or miss might be the understatement of the century but, fortunately, the ProArt P16 RTX Spark’s keyboard falls more in the former category than the latter. Although the keys don’t have a huge amount of travel, I still found the typing experience to be pretty fluid. I rarely ever struggled to find my chosen key when touch typing — although I did occasionally knock both parts of the dual enter key, accidentally sticking random #s at the end of sentences. Still, it felt nice to use, whether I was prompting an AI agent or rapidly splatting demons in Doom: The Dark Ages.

For the most part, the P16 RTX Spark’s trackpad is also well designed. It’s very accurate, gives you a huge area for controlling your cursor and the haptic feedback is spot on, feeling exactly like a mechanical click. Really the only issue I found with it is the right click triggers far too far to the left, meaning I’d sometimes accidentally right click on certain elements. Still though, it’s a decent trackpad for the most part.

As you’d expect from a laptop this premium, the P16 RTX Spark comes with access to a wide range of connectivity options as well. On top of offering Wi-Fi 7 and Bluetooth 5.4 connectivity, it rocks a pretty decent selection of ports for such a slim laptop. These include three USB-C and one standard USB port, an HDMI 2.1 port, a 3.5mm headphone jack and an SD Express card reader. So even for most creative workflows, you should find you’re able to plug in plenty of devices to get things done.

• Design score: 4 / 5

Asus ProArt P16 RTX Spark review: performance

• Strong all-round performance
• Local AI workflows impressive but require some patience
• Specialist laptops can give better gaming or creative performance

I’ll be honest: by far the most exciting part about getting my hands on the Asus ProArt P16 RTX Spark is that eponymous chip. Similar to Apple’s M-series silicon, Nvidia’s RTX Spark is the company’s first attempt at a full system-on-a-chip architecture. And given Nvidia’s involvement in both the AI and gaming spaces, it’s safe to say I was looking forward to putting the Nvidia Spark RTX through its paces.

Combining CPU, GPU and RAM all in one compact package, the version of the RTX Spark included with the Asus ProArt P16 I tested here rocks 20 ARM CPU cores backed up by 64GB of RAM. Meanwhile, on the graphics side, it wields 6144 CUDA GPU cores with 30GB integrated memory. In terms of raw hardware, it's less cooked and more meaty than steak tartare.

While this is the equivalent of using a steam hammer to obliterate a nut, general productivity tasks ran perfectly smoothly on the ProArt P16 RTX Spark. Whether I was browsing the web, tinkering with spreadsheets or leafing through huge PDFs, I never experienced any awkward hanging or slowdowns. Even on the odd occasion when I had to look something up online while the laptop worked on an AI task, there wasn’t any unpleasant gear-crunching, showing the ProArt P16 is plenty fast enough for general use.

But you’re unlikely to be buying an RTX Spark laptop just for doing a bit of Googling or filing your taxes. Instead, the most obvious use case for the RTX Spark chip in the ProArt P16 is its hybrid agentic AI functionality. Put simply, rather than being forced to choose between using cloud models that potentially jeopardize the security of your data and less powerful local models, this blended approach means you can pick and choose which parts of your workflow you’re going to outsource to the cloud and which you’ll keep under strict local control.

With its integrated fifth-generation Tensor cores, the RTX Spark should be perfectly placed to support this. To see how fast it could run some of these models locally, I used the MLPerf Client benchmarking tool to mimic data analyst and SWE agentic workflows using the Qwen 3 8B local model. And the results felt pretty rapid, with the pipelines completing after an average duration of 63.7 and 50.8 seconds and processing up to 14,005 and 10,238 tokens respectively.

But I wanted to see for myself what it could do. So I fired up Hermes, the open-source AI agent tool, and set it up to run the local open-source model Qwen3.8 27B to see what having my own personal AI version of Alfred could help me achieve.

Some of the basic tasks I threw its way, the RTX Spark practically scoffed at. When I asked Hermes to organize a spreadsheet full of web analytics for me and pull out actionable highlights, it whipped the file up into a pivot table and gave me six bullet points within about 12 minutes. Asking it to convert this info into a presentation with slides and prompts took a little longer at 23 minutes, but it helped highlight the best information to pitch — even if the slides themselves needed a considerable glow up before I’d ever call them up in a Zoom call.

In theory though, these hybrid agentic AI tools can go much further than this. Not only can you convert your prompts into skills or schedule them to run as daily tasks but you can even set them to target your apps and control your screen directly.

Some of my experiments paid off here: after one abortive run, I eventually managed to get Hermes to access my emails and create a daily digest in 70 minutes. But trying to automate some of my 3DMark benchmarking process ended up being a real rabbit hole — even with judicious prompts throughout the process and prompting Hermes to find ways to boost the skill’s efficiency, it took around eight runs to significantly improve its workflow. So it’s important to recognize that while some agentic pipelines may save you time in the long run, without utilizing some cloud models, training them may take a decent investment.

Agentic AI isn’t the only use you can put all of that processing power to though. While I’m not a massive fan of generative AI, I can definitely see how fast the RTX Spark is at whipping up new images and videos. Automating some imagery and video pipelines using ComfyUI, the laptop completed tests of Flux2.Dev - NVFP4 and Flux2.Klein 4B Base - NVFP4 models in just 46.5 and 16.2 seconds, while the LTX 2.5 - Distilled - NVFP4 video model took 88 seconds. Those are pretty swift run times for local generative models.

And if, like me, you prefer creating your own stuff from scratch, my testing showed the ProArt P16 RTX Spark can help here too. Using PugetBench to automate a range of creative workflows, I found Asus’s laptop was capable of some lightning fast processing.

Our standard Photoshop workflow completed in 10.83 minutes and achieved a score of 11,801 — that’s not far short of Apple’s M5 chip at 13,755. And the RTX Spark offered even speedier performance for video work, securing a score in our Premier Pro benchmark of 76,234 in 8.12 minutes, giving the M5 chip a bit of a drubbing at 69,887. Unfortunately, it can’t keep up with the similarly priced Apple MacBook Pro M5 Pro though, which managed to best it for both still imagery and video with scores of 15,204 and 105,580 respectively.

Naturally, I’m not just going to take these numbers on trust though, so I dug out one of my most demanding Photoshop files — a 4K, 300dpi PSD with dozens of clipping masks and adjustment layers — and checked how the ProArt P16 RTX Spark handled it.

Manipulating layers and applying filters caused scarcely any slowdown, in spite of the fact I’d tripled the number of layers in the file. Even downscaling the image to 2K — no small task given the total file size was now in excess of 2GB — only took around five seconds, which is pleasingly swift. And thanks to the optimization work that Adobe has been doing behind the scenes to allow its apps to take advantage of the RTX Spark’s AI capabilities, even generative fills and neural filters didn’t take long to generate.

When it comes to gaming performance, I’d say the ProArt P16 RTX Spark hovers somewhere between mid-range laptops with RTX 5060 and RTX 5070 GPUs. For example, when playing Cyberpunk 2077 on Raytracing Ultra settings at 1080p, I found the Asus hovered around 30.16fps — but, by contrast, the Acer Predator Helios Neo 16S AI with an RTX 5060 only reached a marginally smoother 34.47fps. Other games told a similar story. For example, Doom: The Dark Ages on Ultra settings sat at a pretty decent 48fps on the ProArt P16. In comparison, with its slightly beefier RTX 5070, the Samsung Galaxy Book 6 Ultra whipped the game like meringue into a much smoother 63.98fps.

But that’s running things on max settings. In practice, when actually playing games on more moderate settings, they ran plenty smoothly enough, especially those that are already well optimized for the RTX Spark. As I explored Dogtown in Cyberpunk 2077 at Raytracing Medium settings, it chugged along at a perfect stable 60fps. And playing Dogs of War: E-day on Ultra settings, I never saw it drop below 110fps, which shows what wonderfully smooth experiences you should still be able to expect from native gaming on the RTX Spark.

Honestly, I’d say that if you put the Asus ProArt P16 RTX Spark up against any similarly priced, specialist laptop focused on AI, creative workflows or gaming, you’ll likely find its performance can’t quite keep up. But that’s the point — RTX Spark laptops are a triple threat: they’re designed to be a swiss army knife that can handle every task you might throw at them. So if you want a lightweight laptop that combines the agentic AI hosting of a compact desktop, the creative power of the Apple M5 and the gaming oomph of the Nvidia Geforce RTX 5060, all in one handy package, the Asus ProArt P16 RTX Spark should have your back.

• Performance score: 4.5 / 5

Asus ProArt P16 RTX Spark review: battery test

• Battery should last 14 hours of standard use
• Some similar rivals can outlast it

Being able to achieve all of this on the go is impressive. But for that to be sustainable, it needs a battery that can keep up with these kinds of heavy duty tasks. Fortunately, the Asus ProArt P16 RTX Spark is capable of decent battery life, even if it’s not quite as exceptional as some of its rivals.

Running it through our standard battery tests, I played a 1080p video at 50% brightness and volume to see how long it took to drain the battery from 100% to 0%. The ProArt P16 RTX Spark ran out of juice after 11h 23m — that’s not at all bad but it’s also not quite at the level I’d expect for a premium laptop like this.

But a video test doesn’t necessarily give you the full picture. One of the big benefits of an SoC is through boosted efficiency during CPU usage, so I wanted to also run some tests that showed how the laptop fared in action. Running PCMark 10’s applications battery test — which runs various Microsoft Office based tasks until the battery is drained — the ProArt P16 lasted 14h 24m.

That’s definitely a better score, even if it falls a little ways short of some of the best laptops for Windows. For example, the Samsung Galaxy Book6 Ultra managed well over 15 hours in our benchmark tests and that wasn’t even the strongest contender. The MacBook Pro M5 Pro 16-inch clocked over 21 hours, which makes the ProArt P16’s look pretty short-lived in comparison.

Fortunately, the Asus ProArt P16 RTX Spark does have one extra trick up its sleeve: its charging rate. Thanks to its 140W fast charging, topping your battery up again should be a relatively speedy affair. Sadly, my US unit didn’t come with a UK plug — slight oversight there — so I wasn’t able to test this as I would have liked. Still, my 65W charger topped it up relatively quickly, so you’ll definitely get pretty speedy charging out that 140W cable.

So I’d say that, all in all, the Asus ProArt P16 RTX Spark is capable of competitive battery life. When you’re doing standard office work, it’ll get you through a full day without you having to hunt out a charger. But it’s still not quite able to keep pace with the most enduring batteries on the market and particularly intensive tasks like AI or gaming are going to drain it much faster.

• Battery score: 4 / 5

Should I buy the Asus ProArt P16 RTX Spark?

Buy it if…

Don’t buy it if…

Asus ProArt P16 RTX Spark review: also consider

How I tested the Asus ProArt P16 RTX Spark

• Tested over the course of several weeks
• Ran a huge number of benchmarks for different workflows
• Conducted experiential AI, gaming and creative testing

Testing the Asus ProArt P16 RTX Spark over the course of several weeks, I ran a wide range of benchmarks. Not only did I run a variety of baseline hardware tests using Geekbench 7.1 and Cinebench R23, but I tested performance across a wide range of tasks as well. I used MLPerf Client to test agentic workflows, ComfyUI to simulate generative workflows, PugetBench to automate creative tasks and 3DMark and gaming benchmarks to test in-game GPU performance.

On top of this, I did a wide range of experiential testing. Not only did I use the laptop for plenty of daily and productivity-based work but I also extensively tested agentic AI workflows using Hermes and the local open-source model Qwen3.8 27B. I also tried carrying out creative work using large, hi-resolution Photoshop files and played a range of games including Cyberpunk 2077, Gears of War: E-day and Planet Zoo 2.

As TechRadar’s reviews editor, I have plenty of experience covering laptops. As well as editing scores of laptop reviews over the last few years, I’ve reviewed multiple gaming handhelds and have used a huge variety of laptops over the years, both Windows and MacOS.

• First reviewed: October 2026
• Read more about how we test

• T-Mobile, Nokia, and Nvidia claim the first AI-RAN call through a commercial cell site in Bellevue, Washington
• The call ran Nokia's radio software on Nvidia GPUs, but no AI performance data was provided
• Commercial AI-RAN gear is due in 2027, with 6G networks expected around 2030

T-Mobile, Nokia, and Nvidia say they have completed the world's first AI-RAN (Artificial Intelligence - Radio Access Network) call, routed through a commercial cell site in Bellevue, Washington.

The call ran over T-Mobile's live production network using a Nokia Massive MIMO radio, Nokia's 5G anyRAN software and servers built on Nvidia's Aerial RAN Computer (ARC) platform.

This wasn't a lab rig with test handsets, either, but a real-world setup that held; according to Nvidia's announcement, the site connected to T-Mobile's existing 5G Standalone core and a commercial smartphone.

What is an AI-RAN and why is it relevant now?

The radio access network (RAN) is the part of a mobile network your phone actually talks to: the antennas, radios, and baseband computers that process signals at each cell site. That processing has traditionally run on custom chips.

AI-RAN moves it onto accelerated hardware, in this case Nvidia GPUs running Nokia's CUDA-based radio software, so the same box can run AI models that fine-tune the radio link and, in theory, rent out spare capacity to other AI jobs.

The prize on offer is spectral efficiency, or more data squeezed through the same slice of airwaves. In a blog post about the call, Nokia's chief technology and AI officer Pallavi Mahajan writes, "With AI-RAN, we are on a path to 2x spectral efficiency."

Nokia's July 2026 platform launch claims gains of more than 20% efficiency gains so far, without saying where they were measured, and puts 50% on the calendar for 2027 and more than 100% for 2028.

Interestingly, while the tests worked in principle, details remained limited on actual throughput, latency, or efficiency figures from any of the three major players. Put simply, the call shows Nokia's radio stack can run on Nvidia silicon in a live network without dropping anyone.

It doesn't yet show that the AI makes your calls any better. The most concrete numbers for the Nokia and Nvidia platform, ironically, came from a lab: Deutsche Telekom reported typical uplink throughput gains of 10% to 15% from AI-based channel estimation in initial tests in Bonn and Oulu, with a field trial still to come.

The timing is also awkward for those pushing GPUs as an eventuality here; A day before Nokia's post, T-Mobile chief network officer Ankur Kapoor published results from AI-managed beamforming trials with Ericsson in New York City, with the AI model running on Ericsson's own silicon rather than Nvidia GPUs. At peak hours, T-Mobile reported a 16% improvement in downlink user throughput and 12% better spectral efficiency.

Those figures are company-reported, and when Fierce Network asked how many sites were involved, T-Mobile did not specify a number.

Despite this, Omdia senior principal analyst Gabriel Brown told Light Reading, "This is a wow moment," adding that the gains came from software updates to equipment already in the field. T-Mobile and Ericsson had already reported close to 10% better spectral efficiency from a separate AI scheduler in May, again with no GPUs involved.

T-Mobile is clearly backing both horses, as it opened its AI-RAN Innovation Center with Nvidia, Ericsson, and Nokia in 2024. Kapoor's post argues that "AI-RAN is a continuum, not a single product," and he told Fierce the carrier is developing AI-RAN software with both vendors.

For anyone actually making calls, nothing changes yet. Your voice sounds the same, and any benefit arrives as extra capacity in crowded places rather than as a new feature. Nokia's AI-RAN products enter pilot deployments around the end of this year, with commercial availability slated for 2027.

T-Mobile CTO John Saw, who favored field trials, said they would help the carrier get "a sense of what the total cost of ownership is," and is aiming to keep expectations in check. When asked about Nokia's spectral efficiency claims, he told Fierce that "the proof is going to be in the pudding." The Bellevue test shows the plumbing works, but scale has yet to be played out.

Whether the AI earns its keep is arguably a more crucial test, and it has yet to play out. But that hasn't stopped Jensen Huang from framing it a few days ago: telecom networks were built for conversations, and "Now, they will also carry the world's intelligence."

Every cell site that gets a GPU is a sale for Nvidia as it chases its next multi-billion-dollar growth idea, betting on reprogrammable AI infrastructure and continuing to target custom ASICs in multiple industries.

Via DataCenterDynamics

• XMG warns that 'substantial price increases for upcoming GPU shipments' mean it's hiking RTX 5090 laptop prices
• One model is already pricier and bigger hikes for XMG's other notebooks could be coming soon
• At the same time, we're hearing more about the struggles of Nvidia's partners for supply of components for the desktop RTX 5090

The position of Nvidia's RTX 5090 is looking shakier by the day, as even more chatter is emerging suggesting that this GPU is effectively sidelined at this point, even if it hasn't actually been discontinued.

You may have read up on those recent rumors claiming that Nvidia has discontinued the RTX 5090 — which is speculation that Team Green has firmly denied, I should make clear.

The fresh negative news from the grapevine comes firstly from laptop maker XMG on Reddit, which is sharing more bad news on the RAM crisis affecting its prices, specifically with the RTX 5090.

XMG says: "We expect laptops with the RTX 5090 to become significantly more expensive again over the coming weeks." (We already saw price hikes from XMG last month, if you recall).

The company explains why: "Procurement conditions for the RTX 5090 Laptop GPU have deteriorated considerably within a short period. We have received new information indicating substantial price increases for upcoming GPU shipments.

"At the same time, availability is becoming increasingly constrained. Replenishment orders that have already been planned or, in some cases, even confirmed are being significantly reduced in quantity."

XMG notes that beyond these existing orders, it is "currently difficult to make reliable predictions about future supply volumes and delivery dates."

The laptop maker also tells us that "other laptop GPUs are also facing supply constraints", but that the "combination of massive price increases and extremely limited supply" is "particularly pronounced" with the RTX 5090.

That doesn't sound very comforting, does it? The immediate result is that the Schenker Key 18 Pro laptop (Schenker is XMG's parent company) with the RTX 5090 has increased in price by €250 (around $280 / £210 / AU$400) as of today, and the XMG NEO laptops with the RTX 5090 will see at least a similar price increase — although we're warned that the "increase could be even higher" for those notebooks depending on how the situation develops. Nasty.

The second shot of negativity here is from overclocker Der8auer in Germany, who as Wccftech reports, has heard from one of Nvidia's board making partners.

Add seasoning as appropriate here, but the unnamed graphics-card maker claims that while it hasn't been informed that the RTX 5090 is discontinued — Nvidia Germany also told Der8auer this (backing up other statements to this effect from Team Green) — it cannot currently purchase the GB203 chip plus VRAM package (32GB of GDDR7 video RAM) to make the RTX 5090 laptop card.

The more worrying part is that the board maker says that this kind of component shortage has happened before, but it has always been able to place a backorder to be fulfilled later. However, right now it's not even possible to obtain a backorder from Nvidia.

A maker of coolers for the RTX 5090 further observes that it has witnessed reduced orders from board makers, and it doesn't have any orders at all for heatsinks for Nvidia's flagship GPU for Q2 2027 at this point. That obviously suggests that the production quantities for the RTX 5090 are looking even shakier for the future.

Analysis: bleak times

At some point, the accumulation of rumors begins to reach a kind of critical mass where it seems very unlikely that they're all off base, and that in all likelihood, something really is going on.

We appear to be arriving at this point with the RTX 5090 rumors, in my opinion. There are too many of them, and the indications at the retail level — vanishing stock, and ever-increasing RTX 5090 prices, which have become truly ridiculous — also back up the grapevine here.

What's worrying is the further gloom from the likes of a cooler manufacturer, and the fact that the RTX 5090 laptop GPU is also involved. Remember that this doesn't use the GB202 chip seen in the desktop RTX 5090, but Nvidia's GB203 silicon; however, the supply problem lies in getting the video RAM (3GB GDDR7 modules, which are now thinner on the ground and far pricier than the 2GB modules used in the RTX 5080 laptop graphics card).

So, while the RTX 5090 may not be discontinued, it would appear that supplies are getting so low for this GPU — on both desktop and laptop fronts — that while technically it's still in production, the quantities are dipping to such an extent that the flagship is effectively being pushed aside.

All in all, it feels like an especially bleak time for higher-end Nvidia graphics cards, and the fear is that the situation may worsen from here.

Certainly when we look at the predictions for the RAM crisis next year, they're universally pessimistic, and there's a greater level of belief these days that memory supply woes could last for the rest of this decade. Of course, VRAM is bound up in that, as well as AI demand for GPU chips, and all this could affect the launch timeframe of Nvidia's next-gen GeForce graphics cards, too.

• Sky Elements and Damoda launched 71,207 drones during the record attempt
• A formation of 10,183 drones created the record-breaking blue dragon
• Each performance involved approximately 10,200 drones flying several times daily

Sky Elements teamed up with Damoda America in rural Texas to secure two Guinness World Records, launching a total of 71,207 drones within one week.

That weekly total surpassed the previous record of 55,939 launches, set earlier in 2026 by Aerial Illuminations and UVify.

On September 28 2026, the teams used 10,183 drones to form a giant blue dragon, breaking the record for the largest mythical creature depicted by multirotors.

Shenzhen engineers bring charging launch trays

Damoda supplied the hardware and sent specialists from its Shenzhen base to work alongside American operators from Sky Elements throughout the three nights.

Sky Elements sought out Damoda for both titles because the manufacturer offers launch equipment that differs from regular setups in how each aircraft gets positioned.

Crews normally place each drone by hand, whereas Damoda units arrive in trays that store and charge them and receive each aircraft after landing.

The system simplifies transport and warehousing, since every aircraft departs from one dedicated tray and returns there after completing its flight.

Their assignment was to help the Texas crews handle faults as they arose during repeated launches of a fleet that exceeded 10,000 aircraft.

The dragon glowed in electric blue and served as the centrepiece of a programme that also featured oceanic scenes, two astronauts and a lion.

The entire event unfolded at a remote farm with few residents nearby, where the crews operated the whole fleet from one site.

Every performance involved about 10,200 drones, a volume that the crews found demanding because the fleet flew several times each day.

A thunderstorm struck after the opening show on 27 September, yet the remaining performances continued and ran through the closing night on 30 September.

“This was easily one of the most challenging drone attempts I’ve adjudicated, with the weather and the sheer scale of 10,000 drones flying multiple times a day,” said Michael Empric, the adjudicator who oversaw the attempts for Guinness.

“Yet after several very long nights (and early mornings), it was great to see everything finally come together with five successful shows on the final night.”

Fleet and Dragon breaks previous records

The dragon display broke a mark of 2,247 drones set on the final day of 2025 by an Emirati tourism authority and the emirate itself.

The new weekly total exceeded the earlier figure by 15,268 drones, while the dragon formation surpassed the Emirati mark by 7,936 aircraft.

Sky Elements already holds other titles, among them a gingerbread village built from 4,981 drones and a fictional figure made from 4,979.

Its other entries include a cowboy hat image made from 519 drones and an animation that lasted 30 minutes, the longest by unmanned aircraft.

Though an adjudicator attended the attempts, no independent data in the available account shows how reliably the tray system handles storms or daily launches.

• Nvidia's DSX POD aims to cut AI infrastructure deployment to 14 days
• Factory-built containers move major AI data center assembly away from deployment sites
• Vera Rubin systems form the computing foundation of Nvidia's DSX POD

Nvidia has introduced DSX POD, a modular infrastructure design which brings its upcoming Vera Rubin computing systems together with essential data center equipment.

The system places computing hardware, networking, electrical distribution, and liquid cooling inside factory-assembled containers, reducing the amount of installation work required at deployment sites.

Nvidia claims the approach could shorten deployment schedules from six months to just 14 days, provided locations already have the necessary supporting infrastructure.

Factory-built containers aim to accelerate AI deployment

The DSX POD design moves major integration tasks from construction sites into manufacturing facilities, where equipment can undergo assembly and testing before shipment.

Each container arrives with internal cabling, power connections, cooling pipes and networking installed, allowing operators to connect the unit to prepared infrastructure at their chosen location.

Nvidia CEO Jensen Huang linked faster deployment with the ability to begin generating revenue from AI infrastructure sooner.

“An AI factory creates value when it starts producing tokens — not when construction is complete,” he declared.

“With NVIDIA DSX POD, we’re bringing factory-built precision and speed to AI infrastructure construction…so customers can scale capacity and bring revenue-generating production online faster.”

According to Nvidia, reducing installation time allows operators to start training models and running production workloads earlier, potentially bringing forward opportunities to generate revenue.

The reference design starts with Vera Rubin NVL72 systems, which Huang described as comprising seven liquid-cooled computing racks connected through approximately 17 kilometres of Spectrum-X cabling.

However, the 14-day deployment estimate depends on site readiness, meaning power availability, cooling capacity, networking and other supporting systems must already be prepared.

Nvidia claims lower energy and water consumption

Beyond faster construction, Nvidia says its DSX infrastructure platform can improve resource efficiency, although the claimed savings depend on the facility design and operating conditions.

The company claims its cooling approach can eliminate water evaporation during normal heat rejection, using closed-loop liquid cooling alongside outdoor dry coolers to release heat.

Nvidia estimates that a deployment operating at 2GW could save 54 million litres of water daily compared with conventional systems relying on cooling towers.

Extreme temperatures may require additional mechanical cooling, potentially reducing water savings depending on local climate and operating conditions.

Nvidia also claims its DSX MaxLPS technology can deliver equivalent token output using 30% less power through dynamic allocation within an AI facility's available electricity budget.

Another component, DSX Flex, is designed to adjust data center power consumption in response to electricity prices, grid conditions and utility requirements.

Nvidia is collaborating with companies including Dell Technologies, HPE, Lenovo, Schneider Electric and Vertiv to manufacture and supply DSX POD systems worldwide.

Deliveries are expected to begin in early 2027, while Nvidia’s projected efficiency gains still require independent verification under real-world operating conditions.

• Tiny Corp's Tinybox configurator is now AMD-only, after a new commit titled "rip nvidia" removed every Nvidia option and the Tinybox Pro
• The six-GPU, 192GB Radeon AI Pro R9700 build rose from $19,000 to $22,000, and a lower-end four-card, 128GB version now starts at $17,000
• For those wanting Nvidia only, a $91,000 tinybox green v2 with four RTX Pro 6000 cards is still listed in the company's Shopify storefront, about 40% above the $65,000 it charged in May

When TechRadar Pro covered Tiny Corp's relaunched Tinybox on October 13 2026, the headline figure was $11,000 for a server-grade workstation pairing a 32-core AMD EPYC processor and DDR4 memory with four Radeon RX 9070 XT graphics cards.

Less than two weeks on, that configuration is still listed at the same price, but what surrounds it has changed considerably.

Tiny Corp's online configurator is now AMD-only, according to the configurator itself and the public commit history of its website; the RTX 5090 options disappeared on October 11, and the $19,000 RTX 6000 was marked out of stock on October 21, while those thinking this was just a temporary setback were greeted by a commit titled "rip nvidia" that deleted every remaining Nvidia option, along with the $32,000 tinybox pro platform.

AMD GPUs galore?

The AMD side has moved in the opposite direction: Tiny Corp added a six-GPU build based on AMD's Radeon AI Pro R9700, with 192GB of GPU memory, on October 13. It raised each card from $2,000 to $2,500 on October 23 and added a four-card version the next day.

The cheapest six-card build, the "tinybox red flash" preset that the configurator now opens on, has gone from $19,000 to $22,000. The other two presets ship with no GPUs, described as "Bring your own GPUs."

The new Tinybox launched on October 8 with a short GPU list: four RX 9070 XT cards at $1,000 each, or one or two RTX 6000 cards at $19,000 apiece. RTX 5090 options at $8,000 per card were added a day later and were gone by October 11, removed in a commit with the message "rtx5090s aren't made anymore."

Nvidia disputes this, however: at an event in Taipei, Nvidia's Jeff Fisher said that despite a seemingly massive shortage of the chip on multiple US storefronts, "it has not been discontinued."

Fisher's comments followed a leak claiming Nvidia would divert its GB202 silicon to RTX Pro cards, and a warning from custom PC maker StinceBuilt that "Retail GPU pricing is expected to climb significantly," as we reported on October 12.

Tiny Corp's list included Nvidia's 96GB RTX Pro 6000 Blackwell workstation card, and its price has followed a curve similar to the RTX 5090 on consumer storefronts. It launched at $8,565 in March 2025, reached $13,250 on Nvidia's own marketplace by June, according to Tom's Hardware, and hit $16,000 there in August, per TweakTown.

Nvidia hasn't completely disappeared from Tiny Corp's storefront; The company's Shopify store, still linked from its FAQ, lists a tinybox green v2 with four RTX Pro 6000 Blackwell cards and 384GB of GPU memory at $91,000, made to order in two to eight weeks.

The website's own history shows Tiny Corp listing the same configuration at $65,000 as recently as May, which puts the current price about 40% higher. The store also still lists the $32,000 Tinybox Pro, but its description tells buyers to use a configurator that no longer offers it, suggesting potential orders may not be processed.

There is some irony here: in 2024, Tiny Corp put its AMD-based box on hold over unstable firmware and launched an Nvidia version, telling customers, "If you want 'it just works,' buy green," as The Register reported. The October 24 commit scrubbed green from the configurator right down to its stylesheet, deleting the green preset label and Nvidia's accent color.

Each R9700 carries 32GB of GDDR6 with 640GB/s of bandwidth and a 300W board power rating, according to AMD, so six of them provide 192GB and 3,840GB/s in aggregate.

Tiny Corp caps each card at 225W in the six-card build, and the arithmetic suggests why: six cards at their full rating could draw up to 1,800W on their own, more than the Tinybox's 1,700W power supply, while the cap brings them down to 1,350W. The four-card build has no cap, so the same math applies: four cards at 300W total 1,200W.

The price rise is steep, but not completely unwarranted. Partner versions of the card were listed between $1,244 and $1,277 when it first went on sale in July 2025; AMD later set a $1,299 retail price, and one buyer's guide recorded street prices of roughly $1,400 to $1,900 when it rechecked them in September. At $2,500 per card, Tiny Corp now charges nearly double the official price.

Despite using an older 'G4' configuration locked to PCI-E 4.0 speeds, DDR4 memory, and an older 32-core EPYC CPU, the configuration can, when you plug in the maximum number of configurable GPUs and memory, cost as much as $25,000 without any RAID configuration memory whatsoever.

That is a long way from the first Tinybox, a $15,000 box of six Radeon RX 7900 XTX cards that drew 583 preorders in 2024. The FAQ still says the goal is to "commoditize the petaflop and enable AI for everyone."

The configurator now offers that through three AMD GPU options and no Nvidia ones, while a $91,000 Nvidia box sits in the shop, but it is unclear if such orders will be honored. Whether that reflects Nvidia's supply squeeze, the years Tiny Corp has spent building its own AMD software stack, or both, a two-word commit message doesn't convey well enough, one way or another.

• Philips helped Taiwan begin its decades-long technology transformation
• Li Kwoh-ting pursued foreign expertise to build Taiwan's domestic electronics industry
• TSMC manufactures advanced chips, while ASML builds equipment used in production

Nvidia CEO Jensen Huang has recently discussed how Taiwan’s semiconductor leadership emerged through decades of industrial partnerships, government planning and manufacturing expertise.

During a recent conversation with Lex Fridman, Huang discussed how European technology relationships helped establish capabilities that later supported Taiwan’s electronics industry.

The island now manufactures more than 90% of the world’s most advanced semiconductors, giving its relatively small territory enormous importance to global technology supply chains.

Philips helped establish Taiwan’s electronics foundations

The story dates back to the 1950s, when Taiwanese economic planner Li Kwoh-ting pursued foreign partnerships to develop domestic manufacturing technology.

His efforts included attracting Dutch electronics manufacturer Philips, whose involvement helped introduce industrial knowledge and production capabilities into Taiwan’s developing electronics sector.

Philips subsequently became part of the historical background behind companies that helped shape semiconductor manufacturing and equipment development.

Taiwan Semiconductor Manufacturing Company (TSMC) eventually became central to advanced chip production, while Dutch equipment supplier ASML emerged as a leading manufacturer of lithography systems.

These businesses occupy different positions within semiconductor supply chains, with TSMC fabricating integrated circuits and ASML supplying machinery used during chip production.

Their historical connections to Philips help explain how international industrial relationships contributed to Taiwan’s technological development.

However, their subsequent success involved different corporate histories, investments and technological advances rather than one partnership determining the industry’s eventual direction.

Taiwan’s manufacturing capabilities developed through sustained industrial expansion, foreign relationships and growing demand for increasingly sophisticated electronic components.

Computer manufacturing created another opportunity for Taiwan

Taiwan’s electronics industry continued expanding as international demand for personal computers created new opportunities for component suppliers.

IBM helped accelerate this development when it published specifications for its personal computer architecture, enabling other manufacturers to build compatible machines and supporting components.

Taiwanese businesses capitalised on the growing market by designing chipsets and supplying computer components to manufacturers.

TSMC and United Microelectronics Corporation (UMC) subsequently became important semiconductor manufacturers, although their expansion involved distinct strategies and developments across the industry.

The growth of compatible personal computers also created opportunities for motherboard suppliers serving companies assembling and selling complete systems.

Michael Dell, who founded his computer business during this period, initially built machines using commercially available components, including parts sourced through international supply chains.

Taiwanese suppliers formed part of that developing network, although Dell’s eventual success also depended on his direct sales approach and business operations.

IBM’s influence therefore represents a separate stage in Taiwan’s industrial development, rather than a direct consequence of Philips’ earlier involvement.

Huang’s account connects foreign investment, manufacturing expertise and changing market opportunities to explain Taiwan’s rise within the semiconductor industry.

Its current leadership, however, rests on decades of further investment, engineering advances and specialised production capabilities.

The island’s concentration of advanced chip manufacturing has become particularly significant as AI increases demand for sophisticated processors.

Governments and technology companies are consequently seeking reliable supplies, while Taiwan’s manufacturing expertise remains difficult to reproduce quickly elsewhere.

Via Charles Xiong on X

• Marvell reveals Photonic Fabric Memory Module, the building block of a 16-module appliance sharing 32TB of DDR5 and 1,152GB of HBM3e across 16 hosts over fiber
• The design dates back to Celestial AI's 2025 presentations, currently remains pre-silicon, and is targeting customer sampling by 2027
• Nvidia is already hedging with a $2 billion Marvell stake in addition to its own BlueField-4 STX context memory, Enfabrica's pooling tech, and a co-packaged optics roadmap

Marvell has put physical hardware behind one of the more ambitious ideas in AI infrastructure.

In a video posted to X, and earlier in a theCUBE interview recorded around its October 6 Investor Day, the company showed off its Photonic Fabric Memory Module, which it bills as the "first optically interconnected high-capacity, high-bandwidth memory."

The pitch is simple: use light instead of copper so that dozens of terabytes of memory can sit in a separate box, racks away, and be shared by many processors at once, something that could effectively birth the widespread use of communal or shared memory across multiple AI servers.

Two terabytes at a time

Some clarity, however, is warranted: the 32TB figure belongs not to the module but to the Photonic Fabric Memory Appliance, which gangs 16 modules together and shares the pool across up to 16 hosts.

Those hosts can be whole servers or individual accelerators, so 16 GPUs might understate the potential reach in one sense and at the same time, overstate the bandwidth any single GPU will see in another. Nothing is shipping just yet, either: Marvell is targeting customer sampling in 2027.

Each module pairs 72GB of HBM3e with slots for up to eight DDR5 DIMMs, a Photonic Fabric ASIC, and 7.2Tbps of optical I/O, according to Marvell's August 2026 Hot Interconnects presentation.

Hitting 2TB per module means filling all eight slots with 256GB RDIMMs and sixteen modules, joined by a passive all-to-all fiber shuffle rather than an electrical switch, yields 32TB of DDR5 plus 1,152GB of HBM3e, which Marvell rounds to about 1TB and runs as a faster write-through cache in front of the DRAM.

Marvell is going after the economics of today's accelerators. HBM sits roughly 1.5 millimeters from the processor inside the same package, so the only way to buy more of it is to buy more processors. This is increasingly expensive because HBM is extremely costly, and it also means a GPU essentially acts as "the world's most expensive memory controller," as per Dave Lazovsky.

The target workload is the KV cache, the working memory a large language model keeps for each conversation, which balloons with long contexts and agentic sessions. Once it outgrows local HBM and host DRAM, servers either evict it and recompute, or fetch it from much slower storage. This has significant performance implications, and Marvell plans to mitigate it differently.

Ravi Mahatme, who leads product management for the Photonic Fabric unit, said the biggest payoff of a shared tier is "a cross-node KV cache," letting one accelerator reuse context another has already computed. Will Chu, Marvell's EVP of Custom Cloud Solutions, made the broader case in the company's FMS announcement: "As AI scales, memory must scale more independently of compute."

The CAPEX argument is rooted in this: If operators can buy memory separately from accelerators, they no longer need to buy GPUs just to get capacity. This makes sense, but Marvell has yet to put a number to it, even as a May 2026 company blog talks about lower memory cost per gigabyte and lower inference costs without figures, and Celestial AI's 2025 conference slides promised benefits beyond TCO reduction with no supporting numbers.

The timing also isn't ideal for harnessing cheap memory: the approach's 128 DIMM slots draw on the same DDR5 caught up in the current memory crunch, where server RDIMM prices have climbed faster than consumer memory, now around 5.5 times their pre-crisis price in the US.

The appliance design is not entirely new either; Celestial AI presented the same architecture, 16 modules holding 32TB of DDR and about 1TB of HBM, at Hot Interconnects in August 2025, and TechRadar has covered its DDR5-plus-HBM approach back in 2024. What has changed is the Marvell badge, a physical module, and a sampling date.

Nvidia is already hedging, both as an investor and a pioneer

The idea that Nvidia won't sit idle is more than an assumption, because it has been moving for more than a year, including taking a stake in Marvell itself. In March 2026, Nvidia invested $2 billion in Marvell in a deal covering NVLink Fusion and silicon photonics, with Jensen Huang declaring that "The inference inflection has arrived."

It is also important to understand that every NVLink Fusion platform must include at least one Nvidia product, so Nvidia gets paid even when custom silicon wins the socket, but that doesn't mean it doesn't want a piece of that pie anyway.

It is bringing its own answers to the same memory problem in multiple ways; A single Vera Rubin NVL72 rack carries 54TB of LPDDR5X, more than Marvell's entire appliance.

At Nvidia GTC 2026, the company launched BlueField-4 STX, a flash-based context memory tier built for KV cache, claiming up to five times the token throughput of CPU-based storage, with Huang saying reasoning AI needs "a new class of storage." Partner systems are due in the second half of 2026. Back in 2025, Nvidia reportedly spent more than $900 million on Enfabrica, whose EMFASYS system pooled up to 18TB of memory for GPU clusters.

On optics, Nvidia has put $4 billion into Lumentum and Coherent, unveiled photonics switches at GTC 2025, and will offer co-packaged optics as an NVLink option in Feynman-generation systems due in 2028. On the choice between copper and optical scale-up, Huang has said: "We're going to do both."

Marvell still has a long way to go: working silicon, a named customer, a memory partner, and measured throughput on real hardware rather than simulated curves all need to play out exactly as planned.

Until then, the appliance is a well-argued blueprint, and the CAPEX savings remain an argument rather than a number. The irony is hard to miss: the company whose GPUs Lazovsky casts as the world's priciest memory controller owns a $2 billion slice of Marvell, and through NVLink Fusion, stands to earn on the racks either way.

• A new report claims that Amazon is buying up more AI GPU racks than it can deploy in data centers
• An analyst links this report to Nvidia's GeForce GPU supply woes, theorizing that there's a connection
• The contention is that Nvidia is deliberately 'channel stuffing' or in other words, pushing an excess of AI GPUs to the detriment of gaming GPU chips

Nvidia's supply issues with its gaming GPUs are seemingly the result of a strategy to get AI giants to stock up on more graphics chips than they can handle, if fresh reports are to be believed.

You can't have failed to notice the stock issues around Nvidia's GeForce GPUs, and the spike in prices that has happened over the past few months, particularly for higher-end graphics cards.

Well, according to the musings of tech pundit and analyst Ed Zitron (via RespawnFirst) — who picked up on a report by Business Insider — the consumer GPU shortage is a direct result of Nvidia pushing AI graphics cards on the likes of Amazon.

Business Insider informs us that Amazon (or rather AWS) is buying more Nvidia GPUs than it can deploy. Apparently, racks of the AI graphics cards are being kept unused in holding warehouses until there's power available at the data centers where they'll be used (electricity capacity, not space, is the issue here).

Amazon is paying out huge sums for these GB300 racks from Nvidia, and according to an internal email that Business Insider got hold of, a new 'Early Rack Receipt' process (buying these GPUs upfront, before they're installed) saw some $5 billion worth of server racks paid for in August 2026. Seemingly $3 billion more are on order every month for the "foreseeable future" going by another email.

We're told one AWS team didn't have a "clear forecasted path" for getting the purchased racks installed — in other words, they're not even sure when these GPUs might be used.

Zitron's argument is based on a key phrase in the Business Insider report, where the author states that the reason for the switch in policy — Amazon only used to pay for racks when they were installed in a data center — is that "Amazon has to compete more for scarce resources as it jockeys with rivals for access to Nvidia GPUs and other AI components."

The conclusion Zitron draws is that this is basically indicating that Nvidia is making Amazon, and others buying up AI GPUs, compete with each other by agreeing to order a whole pile of them there and then — more than they possibly need — otherwise they won't get their allocation (presumably).

Analysis: GPU misery is clear enough to see

In short, these AI GPUs aren't actually needed as such — not by Amazon in this quantity, as per the Business Insider report — but Nvidia is effectively strong-arming the company into buying the chips, or as Zitron puts it, this is "deliberate channel stuffing." Meaning Nvidia is forcing this AI demand artificially to the seeming detriment of GeForce GPUs, with lesser chip supply for those gaming cards causing price inflation and general misery for gamers.

Obviously we need to take this with a good deal of caution — that goes for the initial report from Business Insider, and doubly so with the theorizing that Zitron applies on top of that article.

There's no denying, however, that we have witnessed a frustrating shortage of consumer GPUs, and in particular Nvidia's high-end gaming graphics cards have become painfully thin on the ground — and pricey. The RTX 5090 is either unobtainable, save through third-party marketplace sellers trying to gouge you out of the best part of 10 grand (in the US), or it's hideously expensive to the tune of over £6,000 in the UK, for example.

What's worse is the RTX 5080 is now getting ridiculously pricey too, reaching £1,400 in the UK. And in the US on Newegg, there's now only one model for sale (direct from the retailer) — and it'll set you back $2,000. It's all getting silly, frankly, and with rumors that the RTX 5090 is being discontinued doing the rounds — speculation Nvidia has denied, I should make clear — it's currently a disappointingly gloomy picture for GPUs.

The expectation is that matters are only going to get worse for RAM, SSDs, and GPUs, as next year rolls into town — and that makes all the above theorizing seem a bit more plausible. It's certainly clear that Nvidia's priority is AI GPUs rather than GeForce graphics cards because of the profits involved with AI, but of course, Team Green is a business — so what do we expect? Prioritization is one thing, though, and the levelled accusation of the practice of 'channel stuffing' is entirely another.

• Nvidia simulates 104 qubits using JUPITER, Europe's first exascale supercomputer
• The technique reportedly reduces physical qubit requirements by 90%
• Magic state cultivation could reduce hardware demands for fault-tolerant quantum computing

Nvidia researchers have simulated a 104-qubit quantum computing system on JUPITER, Europe's first exascale supercomputer, validating a technique that requires 90% fewer physical qubits.

The experiment took place at Germany's Jülich Supercomputing Centre, where researchers examined how quantum computing systems could operate reliably.

The team investigated magic state cultivation, an approach to fault-tolerant quantum computing that could reduce the physical resources required for complex calculations.

New technique reduces quantum computing hardware requirements

Quantum processors are vulnerable to environmental interference, which introduces errors and makes accurate calculations increasingly difficult as systems grow larger.

Researchers are developing fault tolerance methods to address these problems, enabling quantum computers to maintain computational accuracy despite errors affecting individual physical components.

One approach involves producing specially prepared quantum resources called magic states, which are consumed during operations required for fault-tolerant quantum calculations.

Producing these resources requires substantial computing capacity, making improvements in their preparation important for reducing the hardware demands of useful quantum systems.

Nvidia's simulation examined magic state cultivation, incorporating realistic hardware noise, which allowed researchers to investigate errors that earlier simulations conducted under ideal conditions could not adequately represent.

"Useful quantum computing will require fantastically complex, fault-tolerant systems—and understanding how they behave at scale is essential to building them," said Jensen Huang, Nvidia's CEO and founder.

"Our results from Jülich show how quantum processors and GPU supercomputers can work together to tackle research that would otherwise be out of reach—accelerating the discovery of transformative quantum applications."

Earlier research at Jülich simulated 50 qubits under ideal conditions, while the latest experiment incorporated realistic hardware noise.

The experiment nevertheless used classical computing infrastructure to simulate quantum operations rather than testing the technique on a functioning quantum processor.

JUPITER enables large-scale quantum simulations

Located in North Rhine-Westphalia, Germany, JUPITER provided the computing capacity needed to model complex quantum operations that conventional systems struggle to simulate.

The supercomputer used Nvidia GH200 processors for the experiment, with researchers scaling their methods to 8,192 GPUs.

Before moving to JUPITER, the team developed and tested its simulation methods using Nvidia's GB200 NVL72 system at the company's Accelerated Quantum Research Center.

Researchers expanded their initial tests from four graphics processors to hundreds, evaluating performance and accuracy before deploying the methods on the larger machine.

These preparations reduced simulations that previously required several weeks to just days, helping researchers use JUPITER's computing resources more efficiently.

The team also used Nvidia's cuQuantum software and CUDA Q platform to simulate quantum circuits and represent the operations involved in magic state cultivation.

The results could help researchers assess quantum processor designs before constructing hardware, potentially reducing the need for costly physical experiments during development.

However, Nvidia's result remains a simulation, and further experimental work must establish whether the technique delivers comparable benefits on actual quantum hardware.

• Nvidia's RTX 6060 may not arrive until 2029 a leaker claims
• They note that when there are supply issues (the current RAM shortage), Nvidia tends to drag out the release of lower-end GPUs
• They also believe it's unlikely the RTX 6060 will match the PS6 for frame rates

Nvidia's budget GPU for its next-gen range may not be released for three years yet, if a new rumor about the RTX 6060 turns out to be correct.

Wccftech spotted a post from leaker Kepler on the NeoGAF forums, who chipped in with speculation on Nvidia's graphics cards in a thread about the cost of the PlayStation 6.

Kepler was responding to a comment observing that the PS6 is going to go up against the RTX 6060 when it launches in terms of comparisons to PC gaming, and the leaker said — add seasoning with this as ever — that the "6060 will launch much later (and based on recent trends I doubt it will match PS6)".

In a further post, the leaker added: "When there's a supply shortage Nvidia tends to really drag the release of lower end cards. There's a good chance RTX 6060 comes out in 2029."

This would mean we're looking at an early-to-mid 2028 launch for the RTX 6090, 6080 and 6070, which lines up with more recent rumors. Of course, that's not a foregone conclusion, and Kepler is clearly speculating here.

Recent reports from analysts indicate that the PS6 will likely launch in 2028, though some have argued that Sony's next-gen console could appear late in 2027 (alongside other devices, perhaps).

Analysis: late, as well as short on performance?

Of course, what Kepler contends here is what Nvidia is supposedly planning right now, and just as with the PS6, a lot of the direction that Team Green goes in with its next-gen GPUs will depend on the memory crisis. Graphics cards do, of course, use video RAM (VRAM), and the next generation of GeForce products will be expected to up the ante on that score. (Gamers are already unhappy with Nvidia's current-gen VRAM configurations).

The problem is that there's no shortage of predictions that RAM cost inflation is going to get even worse in 2027, after it has already spiked in a big way through the middle of this year. The theorizing then runs that 2028 is also going to be a difficult year, which lends weight to the argument that the RTX 6060 will indeed be late to the GPU party.

In short, I wouldn't rule out the possibility of a 2029 arrival for the RTX 6060 as suggested, which obviously isn't great news for those hoping to score a more wallet-friendly next-gen Nvidia GPU.

It's also interesting to note that Kepler believes the RTX 6060 won't be as powerful as the PS6 for gaming frame rates, but I'd take that claim with even more salt than the rumors around timing.

Gauging hardware capabilities at this stage is obviously a trickier task, but Kepler's assertion isn't an impossibility of course. Note that the current RTX 5060 matches the PS5 Pro's gaming performance (roughly, for rasterization as opposed to ray tracing), so if the RTX 6060 couldn't equal the PS6, that would be a clear disappointment for PC gamers on a budget (particularly if the VRAM loadout again turns out lean).

Having been at Nvidia's RTX Spark event in San Francisco earlier this month, I was lucky enough to get a firsthand walk through of the many things the 'superchip' can do. Since then I've spent almost every waking moment putting the ARM-based RTX Spark through its paces and seeing what makes it so different from other chipsets that have come before.

The RTX Spark is Nvidia’s first System on a Chip (SoC). Combining an 18-core CPU and 5120 Cuda core Blackwell RTX GPU all in one, or a 20-core CPU and 6144-core GPU for its meatier version, it offers an integrated chipset in the vein of Apple’s M series chips. While Apple’s silicon is known for its oomph, that Blackwell RTX GPU means you’ll also get considerable gaming performance here to boot.

How does this translate in terms of raw numbers? It’s certainly got a lot going on under the hood, as the RTX Spark N1X chip I’ve been playing around with squeezes 20 cores into its CPU. While the 2,623 single-core performance I clocked on Geekbench 7.1 falls somewhat short of the 4,288 we got for Apple’s M5 chip, its 21,899 multi-core score massively outstripped the latter’s 17,926. That means that for tasks like content creation, software development and AI, you should be able to expect more oomph in your corner.

So given the potential it promises across agentic AI, creative workflows and gaming, let’s take a look at how it performs across each of those areas in turn.

Hybrid agentic AI

During the launch of Nvidia’s RTX Spark SoC, there was a huge focus on its potential for hybrid agentic AI. Now, cloud-based AI models are undoubtedly pretty powerful but I absolutely have security concerns about uploading my most private data to the cloud and I’m not entirely sure I want to blindly trust the companies behind some of these models.

Using local, open-source models that I can vet myself and that never transmit sensitive data outside of my network feels like a much safer alternative — but not everybody necessarily wants to cut themselves off from the power of frontier models. That’s why Nvidia has introduced a hybrid-system for the RTX Spark instead.

Instead, the hybrid-workflow that the RTX Spark supports allows you to select your choice of local models and ring-fence them for certain parts of an agentic workflow, while outsourcing specific, demanding tasks to cloud frontier models. As a result, users should be able to take a far more granular approach, deciding exactly how much they want to rely on public vs private models.

For people who like to vibe-code complex products or are generating demanding 3D scenes, not being forced to choose between binary security options of a lax, YOLO approach and a full-blown lockdown will likely be a valuable tool.

So now that I’ve spent some hands-on time with the RTX Spark, how has its agentic performance delivered on these promises? To get a broad picture of I ran the MLPerf Client benchmark to simulate data analyst and SWE agentic workflows using the Qwen 3 8B local model, with each run hitting an average duration of 63.7 and 50.8 seconds and processing up to 14,005 and 10,238 tokens respectively.

I also benchmarked some generative image and video workflows using ComfyUI. When testing workflows using Flux2.Dev (NVFP4 and Flux2.Klein 4B Base) NVFP4 models for image generation, it took 46.5 and 16.2 seconds to generate respectively, while the LTX 2.5 (Distilled) NVFP4 model took 88 seconds. That’s broadly in line with what Nvidia has been promising and shows how rapidly the RTX Spark can generate media using local models.

Raw data will only tell you so much though. That’s why I spent days training my own AI agent using Hermes and the Qwen3.8 27B open-source local model to see what the RTX Spark could do.

Certainly, it seems to have absolutely ripped through some tasks on its Extra High setting. Feeding it some analytics data in CSV form, I asked it to tidy up the spreadsheet, analyze the data and present its key findings and after 12 minutes of pondering and tweaking, it happily spat out a pivot table and some observations on the data. While I wasn’t hugely surprised by any of its conclusions, I’ve spent months poring through this data and the fact it could identify the most essential points without any of my secure information leaving the laptop was pretty impressive.

Some experiments I conducted involved a little more patience and tweaking of the agent’s workflow to get the best results though. The first time I asked my Hermes agent to connect to my online email client and provide a digest of the day’s emails, it spent 75 minutes trying various methods to exchange tokens with Outlook’s online client, before crashing unexpectedly. However, the second time I played a slightly more active role, providing it feedback when I felt it was drifting or had come up against hurdles and it successfully set up the brief in 70 minutes.

So with some patience and judicious prompting, you can definitely get some good results utilizing a local model. Additionally, you don’t have to go through this full process every time — ask Hermes to save something as a skill and it will automate this process in future or even schedule it as a recurring Cron Job. You can even ask it to work out ways to improve the efficiency of existing skills, meaning it should be easier to minimize the run time and resources they require in future.

Creative workflows

You’d be forgiven for thinking this means the RTX Spark laptops are predominantly focused on AI tools. But that couldn’t be further from the case: while a lot of the press surrounding the chip has been on its agentic AI tools, this really isn’t the sole string to its bow.

As you can tell from the fact that the RTX Spark is used in the latest Asus ProArt P16, creative work is a big focus for the new SoC. Even if you’re not interested in the power it offers for generative AI, the sheer number of RTX-compatible creative apps out there, whether that’s Blender, Adobe Creative Suite, Canva or Unity shows how many creative giants are throwing their weight behind the new SoC.

Eager to test how well the RTX Spark handles creative work, I used Puget Bench to automate a range of tasks in Photoshop and Premiere Pro. PugetBench’s standard Photoshop workflow took 10.83 minutes and netted a score of 11,801, while its standard Premiere Pro pipeline took 8.12 minutes and achieved a score of 76,234, putting it firmly in the range of a true creative workstation. That means it sits in the rough region of Apple’s M5 chip, which scored 13,755 and 69,887 respectively, besting it on complicated video work but not quite equaling it for still imagery.

For those who use generative AI in their work, I also benchmarked some generative image and video workflows using ComfyUI. When testing workflows using Flux2.Dev (NVFP4 and Flux2.Klein 4B Base) NVFP4 models for image generation, it took 46.5 and 16.2 seconds to generate respectively, while the LTX 2.5 (Distilled) NVFP4 model took 88 seconds. That’s broadly in line with what Nvidia has been promising and shows how rapidly the RTX Spark can generate media using local models.

While I’m neither an aspiring cinematographer or a budding Blender artist, I’ve spent decades tinkering with static images so I dug out an old 4K, 300dpi Photoshop file. It was already pretty huge as it featured lots of hi-res layers, clipping masks and adjustment layers but I tripled the number of layers for good measure and added a bunch of filters and effects. Processing time remained less than a second and even halving the image size only took about five, which is pretty impressive given the sheer number of layers and groups I was working with.

And the RTX Spark should even be able to speed up these applications’ built-in AI tools, thanks to the recently announced collaboration between Nvidia and Adobe. Optimizing the apps to take full advantage of the chip’s GPU, Adobe promises that RTX Spark laptops offer 2x AI processing, editing, color and effects. That’s a real advantage for anyone that uses these AI tools as part of their workflow.

Gaming performance

Given Nvidia’s history, gaming is naturally also a big focus for the RTX Spark chipset. As most games run on x86 architecture, SoCs haven’t been the most natural home for gaming in the past. However, Nvidia’s stature in this space means plenty of developers launched RTX Spark native versions of their games when the chip was launched on October 16, including Cyberpunk 2077, Doom: The Dark Ages, and 007 First Light — while the Call of Duty series is due to land in 2027. And for those that aren’t yet compatible, Prism emulation will mean you’ll be able to run many titles with only a slight impact on your CPU performance.

When it comes to gaming performance, the RTX Spark is certainly no slouch. It netted a 3DMark Port Royal score of 8,133, Steel Nomad score of 2,846 and a Speed Way score of 2,872. That’s a touch below Asus Zephyrus G14 with Nvidia GeForce RTX 5070 Ti’s respective 10,119, 3,488 and 4,031 scores. But it’s considerably better performance than you’d get out of most laptops with an integrated GPU such as the pre-RTX-Spark Dell XPS 16, which secured values of 3,888, 1,488, and 916.

But while these simulated benchmarks are helpful for getting a broad view of gaming performance, in-game benchmarks are the best way of gaining real insight into how actual titles perform. While testing the RTX Spark, Cyberpunk 2077 delivered a stable 30.16fps at 1080p on Raytracing: Ultra settings and Doom: The Dark Ages averaged out at 48.58fps at 1080p — very similar to the 32.33fps and 55.63fps we got out of the Asus TUF Gaming A14 with Nvidia GeForce RTX 5060.

Nothing compares to actually trying out games first hand though, so I fired up a few of my favorites and tried them out. Cyberpunk 2077 honestly felt pretty playable even using the above Raytracing Ultra setting and once I’d switched down to Ray-tracing Medium, it ran at a wonderfully fluid 60fps.

Meanwhile, running Gears of War: E-Day at Ultra settings felt as smooth as tahini drizzled over a marble countertop. As I crunched through waves of the Horde, there wasn’t any noticeable slowdown even at Ultra settings, with the action never dipping below a seriously fluid 115fps. So it might not match the RTX 5090 for sheer power but anyone barring elite gamers will find a lot to like here still.

So even though it’s naturally not at the level of dedicated gaming laptops like the MSI Vector A18 HX A9W, I’m still pretty impressed with gaming on the RTX Spark chip. And that only becomes more pronounced when you compare it to similar lightweight SoC systems. While the MacBook Pro M5 is similarly impressive in terms of creative performance, it was barely able to scrap 11.5fps in Cyberpunk 2077 on Raytracing: Ultra settings.

As a result, if you’re looking for a chipset that will give you not just great performance in one specific task but great performance across everything you do, the RTX Spark has proven it will be a solid choice.

Nvidia CEO Jensen Huang has sung the praises of Microsoft's new Surface Laptop Ultra, describing it as "so beautiful" and "incredible".

Speaking at a developer event in London, Huang outlined the development process that went into building the hardware powering the new product, which is the first Microsoft device powered by an Nvidia chip.

“This computer is going to be incredible - it took us four years to build it," Huang told attendees including TechRadar Pro, "this project is really a work of love - we had several thousand engineering years dedicated to it, we build a brand new revolutionary chip, a whole new line of PCs, it’s so beautiful - you’re going to love it.”

The end of the traditional PC?

Revealed in September 2026, the Surface Laptop Ultra is targeted towards developer and AI usage, powered by Nvidia’s RTX Spark platform with an N1X processor and RTX Blackwell graphics, supporting up to 128GB unified memory

This is alongside 24GB RAM and 512GB storage, while the next version has 32GB RAM and 1TB storage, making it ideal for creators, AI specialists, engineers, scientists, developers, and other users running demanding applications.

But as Huang tells it, the creation of the device, set to launch on October 16 2026, was a long and arduous process.

“It took four years to do this," Huang said of the hardware development process. "We were thinking about agentic systems, AI and autonomous systems, and how they’ll be able to access your files, and be grounded on information, and be able to reason about it and use tools - but the ultimate tool is your personal computer, right?"

"So wouldn’t it be cool if we could create a personal computer, a PC that actually runs an agent locally - so it can go off and just work on things for you - and not only could it just run locally, it would be smart enough to use the tools."

Huang explained his vision of "texting" your PC to ask for help, to get it to do tasks like rendering images and updating slides, then delivering them straight to you.

"Wouldn’t it be awesome if you could just text your computer - you don’t have to bring it with you anymore - just leave it at your desk!”

“So we created this thing - it took us four years to do it, so I called Satya (Nadella, Microsoft CEO), and the two of us talked about it…and so we had to create a brand new chip, we had to wait until Blackwell was done - and this thing has 1 petaflops of computing, and because we knew the models were going to be big, we had to redesign how the GPU uses memory - so this is the first Nvidia GPU with shared memory, one petaflops, Blackwell architecture - and also it has to run everything a PC doesn’t.”

“There’s three types of PCs, right," Huang outlined, "there’s the PC that basically you use to type in - that’s called a PC, there’s the Nvidia gaming PC, you’ve got GeForce in that, and you’ve got the Nvidia workstations that runs all kinds of applications…so we figured that if you’re going to have an AI go use the tools, it should be able to use every single tool…it runs everything, it runs every AI framework, it runs any AI model as well."

• Microsoft will make Maia its main inference engine at multi-gigawatt scale, following Google and Amazon into in-house silicon
• Microsoft's upcoming Maia 300 trades a modest 36% compute gain for 2.5x Maia 200's memory bandwidth, betting on memory-bound decode, with production in 2027
• Hyperscalers supply just over half of Nvidia's data center revenue at a 75% gross margin, so each move to in-house chips affects not only its pricing power but the chip industry's sustainability as a whole

In 2024, Microsoft told the European Commission that Google was the only company vertically integrated at every AI layer, from chips to a mobile app store, and that "everybody else must rely on partnerships to innovate and compete."

Now, it has revealed plans to follow the same route without breaking a sweat.

In an Official Microsoft Blog post, Scott Guthrie, executive vice president of Cloud + AI, said Microsoft will scale its Azure Maia accelerators to multi-gigawatt deployments starting in 2027 and make Maia the primary inference system for Microsoft's products and services.

The vehicle of its choice is Maia 300, which Rani Borkar, president of Azure Hardware Systems and Infrastructure, and Saurabh Dighe, corporate vice president of systems planning and architecture, detail in an architecture deep dive.

Feeding the beast, not growing it, as it takes the Google route

The move is a pointed statement from one of Nvidia's biggest customers, and it lands as bubble worries grow louder. Gartner forecast this week that AI providers will spend nearly $1 trillion on infrastructure in 2026 alone, and that AI revenue won't cover the build-out profitably until at least 2032.

If the largest buyers of AI hardware stop paying Nvidia's margins for the work they do most, the most richly valued part of that build-out starts to look exposed. Whether Maia 300 is the final catalyst for such a move, though, is a harder case to make.

Maia 300 moves from Maia 200's single die to a multi-chiplet package, with compute on TSMC's N2P process and memory-controller and I/O functions split onto separate dies. Eight HBM4 stacks deliver 288GB at 18TB/s per accelerator, against 216GB at 7.1TB/s for Maia 200, according to StorageReview's breakdown of Microsoft's spec sheet.

Microsoft hasn't abandoned compute so much as rebalanced around memory. In decode, the token-by-token phase that dominates long reasoning chains and agent loops, accelerators spend much of their time waiting on data rather than doing arithmetic.

The Maia 300 tries to close that gap by keeping data near the compute that uses it, through a non-uniform memory layout across chiplets, software-managed on-die SRAM, and narrower number formats, including an enhanced MXFP4 and sub-3-bit storage formats, with hardware conversion between them. The Maia 300 announces something the rest of the industry already knew, albeit more subtly: Custom silicon is no longer a side project for hyperscalers.

Google, the company Microsoft singled out two years ago, began recognizing revenue from TPU system sales delivered to customer data centers in the second quarter, CFO Anat Ashkenazi said on Alphabet's July earnings call, while CEO Sundar Pichai said the company is "seeing strong demand for TPUs as standalone systems from other customers."

It isn't the only one either; Anthropic has lined up 3.5GW of TPU capacity through Broadcom for 2027, on top of an earlier deal for up to a million TPUs, and Meta was reported in 2025 to be in talks to rent-cum-purchase them. Amazon has gone the same way. CEO Andy Jassy said in April 2026 that Amazon's chip business, spanning Trainium, Graviton and Nitro, had passed a $20 billion annual revenue run rate and was now "one of the top three data center chip businesses in the world." OpenAI, meanwhile, has lifted the lid on its own Jalapeño accelerator.

Nvidia's margins: the AI holy grail?

The incentive is plain from Nvidia's own results. In the quarter to July 2026, the company reported $96.2 billion in revenue, $89 billion of it from data centers, at a 75% gross margin. CFO Colette Kress put hyperscale revenue at $49 billion on the earnings call, meaning a little over half of Nvidia's data center business comes from a small group of buyers, the biggest of which now design their own accelerators.

Every inference workload that moves in-house is one where that buyer pays closer to manufacturing cost instead of Nvidia's margin. Microsoft told investors in July that Maia 200 was up to 40% cheaper to operate than Nvidia's latest chips on OpenAI and Microsoft models.

It did, however, fail to name a particular GPU it called a "leading GPU-based system," an important distinction given Nvidia's focus on efficiency via Vera Rubin, which OpenAI even refused to compare against outright.

The reveal didn't move stocks much; on Monday, Nvidia shares closed down 1.29% after the reveal, while Microsoft's rose 1.14%, and Microsoft's opaqueness about efficiency gains may be one reason.

The other reason is equally important: Evercore ISI analysts estimate Maia will account for only about 10% of Microsoft's AI chip spending by the first quarter of 2027, CNBC reported, and Nvidia has guided to $108 billion in revenue for the current quarter.

Scale is arguably an ongoing constraint, however; at 167 kW per rack, a gigawatt of rack power is roughly 6,000 Maia 300 racks, or about 575,000 accelerators, before cooling and facility overheads.

Going the Google route hasn't stopped anyone buying from Nvidia either. Amazon, with its "top three" chip business, is still set to take more than 3 million Nvidia GPUs by mid-2028.

Alphabet, which now sells its own TPUs, raised its 2026 capex guidance to between $195 billion and $205 billion in July, lists Vera Rubin alongside its own TPUs, and Ashkenazi said it is "still in a supply constraint environment."

Microsoft keeps Nvidia and AMD for training, and Satya Nadella said earlier in 2026 that "because we can vertically integrate doesn't mean we just only vertically integrate."

Jensen Huang's answer to the custom silicon question on Nvidia's August call could have been written about Maia: "Whereas many of these XPUs are inference-specific chips for one cloud or one service, NVIDIA is a platform." Microsoft's counter-bet is somewhat different: one cloud is big enough.

The deeper problem for the pin theory is that custom silicon changes who collects the margin, not how much gets spent. Gartner warns that the industry needs about $8.5 trillion in revenue between 2026 and 2030 to earn a reasonable return on its build-out, versus no more than $3.6 trillion it expects AI to generate. Microsoft isn't responding by building less; it is planning gigawatts more, and cheaper tokens arguably help close that revenue gap rather than widen it.

Some of what Microsoft saves will also flow elsewhere in the supply chain. Maia 300's pitch rests on HBM4, the same constrained memory Nvidia's Rubin uses at the same 288GB per package, and its compute dies come from TSMC, which supplies everyone.

Google's TPUs are developed and supplied under a long-term agreement with Broadcom, whose CEO Hock Tan said that "many investors in AI don't understand what they're putting their money into," and who expects chip shortages to ease only around 2028 or 2029. Nvidia isn't sitting idle either; it already paid $20 billion to license Groq's SRAM decode technology, a sign it sees the same inference threat coming.

Until production silicon ships and someone outside Microsoft can measure it, Maia 300 is a well-argued design whose most important numbers are ones that only Microsoft has seen, and a reminder that the biggest threat to Nvidia's margins may be its best customers rather than its rivals.

It's been a theory for quite some time now that Nvidia is — what's the polite way to put this? — let's say deprioritizing PC gamers, perhaps, or more specifically, its gaming graphics cards. But this week has added more evidence to the collective weight that GeForce GPUs are increasingly at risk of being sidelined further by Team Green.

I'm talking first about the story that emerged a few days ago, where a leaker claimed Nvidia is diverting all GB202 (Blackwell) chips to make AI (RTX Pro) graphics cards. That's the graphics processor used in the RTX 5090, but apparently not anymore; it's for AI boards only now, we're told.

That makes some sense as the RTX 5090 is already disappearing from shelves, certainly in the US, where you can't buy one for less than the best part of 10 grand (from gouging third-party sellers, with no major retailers actually having stock anymore). So arguably this effective discontinuation of the RTX 5090 is already happening.

On top of that, the latest developments are that Micro Center is reportedly listing the Blackwell flagship as 'not carried', and Nvidia itself has removed all mention of the RTX 5090 from its own GPU marketplace (though it is still mentioned elsewhere on the site). Search for the RTX 5090 on that marketplace, and you won't find it any longer — the highest-end graphics card listed is now the RTX 5080 (hat tip to Wccftech for spotting this being flagged up on Reddit).

Aside from an actual announcement from Nvidia (which probably wouldn't ever happen anyway), that seems like about as clear an indication you can get that the RTX 5090 has been shoved out of the exit door, never to return, because Nvidia needs those GB202 chips where they are most profitable. RTX Pro GPUs offer a much heftier premium that Team Green can heap on top.

A new Blackwell flagship?

So, what does this mean? For starters, if it is true that the RTX 5090 has been scrapped, the RTX 5080 becomes the flagship Nvidia GeForce GPU. Once the existing stock of RTX 5090s has been depleted, this could mean the price of the RTX 5080 is inflated. When I say inflated, I mean jacked up even more, because it has already seen some nasty price hikes. Currently, the RTX 5080 is at around $1,800 in the US on Newegg, or £1,450 on Overclockers in the UK – and those are the lowest prices; it goes up from there. It's a very similar story in many other regions.

This is clearly a concern for gamers as you can see from the Reddit thread I pointed out above, with comments such as: "If you want a 5080 better buy one while you can still afford one." Really, it's debatable if the card is anything like 'affordable' relative to its graphical frame rate muscle even now.

Here's where another rumor comes into play. Elsewhere on the grapevine, there's talk that Nvidia may actually release the RTX 5080 Super variant which has long been speculated about. This would be a 24GB graphics card, and it'd essentially replace the discontinued RTX 5090 – with the theory being that Nvidia would also ditch the vanilla RTX 5080 (16GB). This would leave the rumored RTX 5080 Super as the sole high-end GeForce graphics card, quite some distance above the RTX 5070 Ti.

If that happens, imagine the pricing concerns here – what MSRP would Nvidia attach to the RTX 5080 Super? And then consider the cost inflation that will be added when you take into account that this card will also be sought after for AI usage (as seen with the RTX 5090).

If an RTX 5080 Super is indeed inbound for early 2027, that would also suggest that the next-gen RTX 6000 family is further away than some rumors believe – as this Super version would be a stopgap GPU to keep the Blackwell range ticking over. The RAM crisis may have a fair bit to do with this strategy also, given that most predictions foresee it getting worse next year (and the misery continuing into 2028, and maybe beyond). Of course, launching a new GeForce range into such memory headwinds wouldn't be easy.

Software focus and streaming dreams

All in all, it feels to me like there's a stream of subtle – and not so subtle – hints and rumors that Nvidia is very much slowing down the development of gaming GPUs (while not wanting to give that impression, naturally). The other part of the equation here is the increased focus on the software side, of course, rather than the hardware. DLSS 5 is out and Nvidia is already pushing hard with honing and refining that neural rendering tech, and something else that happened earlier this week was that Team Green highlighted some big leaps in performance here. Pretty huge strides forward, in fact.

Of course, you could argue that all the development resources being ploughed into DLSS – including frame generation, which got a hefty boost alongside DLSS 5 – shows Nvidia isn't abandoning gaming GPUs for PC gamers. And I concede that's true to an extent, but there is very much a vibe of a major slowdown leading towards the feared deprioritization of GeForce GPUs – and there's another angle to consider here too: cloud gaming.

Don't forget that GeForce Now also benefits from DLSS. Just this week, Nvidia added DLSS 4.5 override controls to the streaming service to help push up frame rates – those on the RTX 5080 (Ultimate) plan can use Dynamic Multi Frame Generation of up to 4x now. Ray Reconstruction for bolstering ray-traced visuals is available, too, among other options.

Would Nvidia rather have you subscribed to its cloud gaming service, making monthly payments, than buying a GeForce GPU for a one-off sum? It seems an obvious direction for the company to be heading in, and with increasing warning signs around Team Green's gaming GPUs, I'm becoming more convinced that this is where Nvidia's priorities are shifting for the future of PC gaming.

This isn't to say GeForce graphics cards are definitely on their way out, or that an RTX 6000 line-up won't happen – it's already well underway after all – but beyond the next generation, I'm increasingly dubious about where GeForce GPUs might fit into the big PC gaming picture. Between AI demand and cloud gaming (plus increasingly faster internet connections), there's a squeeze underway that could prove painful for the GeForce line-up – and the ongoing RAM crisis is hardly helping, either.

Nvidia CEO Jensen Huang has laid out his vision of how OpenClaw will help shape the future of technology - with a particular focus on robotics.

Speaking at a “Build a claw” event in London, Huang explained how OpenClaw and AI agents as a whole can spur on incredible advancements in a range of fields.

However it was the field of humanoid robotics that seemed to catch his fancy, with Huang outlining the need for proper controls to ensure the potential of the technology can be reached.

AI is "becoming useful"

Huang opened by noting how OpenClaw was “one of the most important models” ever seen, helping to further progress across multiple industries.

“We’ve now seen an inflection in artificial intelligence - it’s become useful,” he declared. “The type of things people are doing with it, as simple as software coding, being a marketing agent assistant, research assistant, lab assistant, helping scientists do drug discovery and engineers run experiments - we’ve got agents everywhere.”

Spotting a robot built by Nvidia partner Humanoid, Huang noted how OpenClaw is the “first credible” agentic harness, which he likened to LLMs inside an exoskeleton, before extolling the possible benefits to using the technology for robotics.

“When it first came out, we thought wow, what a really terrific digital robotic system,” he said, “kind of like a human robot, that is basically the physical harness of a large language model, essentially - the motors and the articulation and the wheels and the arms, those are just tools that are used by the agent that’s sitting inside the computer.”

“So when we first saw that we said, this is incredibly useful, and we’re going to help make it even better - but the one observation that we made was that these agentic systems, if they were able to be deployed in a safe and secure way, we’re going to have to create some kind of container for you.”

Huang noted that the current situation was similar to the birth of the internet, when a browser itself was not isolated from the computer, so when a user interacted with a website, it would download an app which had full control of bare metal - and often delivering what he described as “a special package to your PC”.

“The obvious thing to do is create essentially a container, an isolation layer - essentially another operating system inside the operating system - so we created this thing called OpenShell,” Huang said, highlighting how this creates an isolation system for agentic systems.

“When you’re building an autonomous system, the first thing you should do is assume zero trust. So you give it a policy, you get minimal privileges - so if you’re asking an autonomous system to make you an omelet, you don’t have to give it access to a chainsaw - all you need is a frying pan…so whatever the task you give your agent, you give it minimum rights, minimum privileges, and then you get to monitor it to make sure that it’s compliant with policy”

“Zero trust, minimum policy, continuous monitoring, contained inside an isolation system that is separated from bare metal - that’s how you develop and deploy agents safely.”

But what if unexpected incidents or complications arise? Huang was optimistic that things would not get out of hand.

“We will monitor the autonomous systems continuously, and if that robot were to start to violate or do something that’s inconsistent with the privileges we gave it, it would escalate to us, and we can decide whether we want to shut it up!” he laughed.

“These kind of systems, these escalation systems, these security systems, all have to be built so that we can give all these autonomous machines and autonomous agents and deploy them in a way that is able to do so safely and securely, so we can realise the incredible future potential of the technology that we’re building.”

• Nvidia claims that for each gigawatt an AI data center uses, only 70% is used for AI work
• Inefficiencies include cooling and power distribution, failures, and restarts
• While a chip shortage was considered to be the main bottleneck to AI data center capacity, it is now a facility’s connection to the grid

AI data centers are inefficient, says Nvidia, with various physical and systems issues contributing to power use beyond AI processing. This opinion, shared by data center builder Vertiv at its recent annual EMEA event, underpins future development steps, in which it plans to assemble sustainable facilities.

Vertiv and Nvidia have been collaborating since early 2026 on co-designing the next generation of data centers, with Vertiv providing power and infrastructure management for Nvidia’s GPUs.

With a maximum of 70% of electricity drawn from the grid used for AI processing, the remaining 30% represents inefficiencies that optimized data center designs for the future should reduce.

That's not efficient

Although the various inefficiencies affecting AI data centers may not come as a surprise, the scale of them might. According to Nvidia – via its Vice President of Supercomputing, AI Cloud Infrastructure, and AI/HPC for EMEA Rod Evans – that 30% consists of infrastructure and system issues.

For a one-gigawatt data center, 20% is lost through cooling, power distribution, and conversion. A further 10% is lost to incorrectly configured racks, as well as failures and restarts.

Evans says that “You sign a contract for 100 megawatts, but you never actually get that. With a bit of luck, you’ll get 93 megawatts, and at certain times of the day, even less.”

Nvidia believes some of the 30% (which can be as high as 40%) can be recovered. Evans notes that “A data center is actually designed for peak load and outages, not for what it typically uses.” This means that unused capacity can be tapped, and power and cooling can be revised to find the optimum configurations.

Bottlenecks of power

Rod Evans states that Nvidia and Vertiv are planning data centers that are more efficient and sustainable. “Together, we want to build sustainable AI data centers where the focus isn’t on using as much power as possible, but on getting the most out of every watt.”

One of the most significant details from the press conference is the challenge posed to AI data centers not by a shortage of chips, but by a lack of power. Bottlenecks are resulting in underpowered data centers, with some waiting for several years to receive the power to match their capacity.

This can be impacted by location. Vertiv’s EMEA president Paul Ryan explained that while you can get power in some locations but no permit. “In another, the permit is granted quickly, but there’s no power. And sometimes you have both, but the location is so remote that you can barely get the equipment there.”

• Nvidia has considerably improved the AI model for DLSS 5
• This means the tech now offers up to 40% faster frame rates
• Team Green has also improved frame generation, and is bringing DLSS 5 to two new games: Hogwarts Legacy and Where Winds Meet

If you were wondering how soon Nvidia might start ramping up the performance of DLSS 5, well the answer is that it just happened.

Team Green's neural-rendering tech, which officially debuted in NBA 2K27, has only been out for just over a month, but Nvidia is claiming that it has doubled the performance of the AI that drives DLSS 5.

When it was first previewed in March 2026, DLSS 5 needed a pair of RTX 5090 GPUs to run, with the company improving the performance of the AI model by 5x for the release of the tech in NBA 2K27. With this latest update, it's now 10x more performant than that baseline.

Nvidia also made it clear that it has reduced the video memory (VRAM) usage of the tech considerably, noting that: "The DLSS 5 October update shrinks the [AI model] footprint by 24% to roughly 118 million parameters while reducing 4K VRAM usage by 20%. The result is half the GPU overhead of the September [launch] version while maintaining comparable visual fidelity."

Based on Nvidia's testing in NBA 2K27, that translates to boosts in frame rates in the order of 30% to 40%. Check out the video clip below to see the gains in the basketball game.

This new DLSS 5 model launches on October 16, with availability for RTX 5000 GPUs, RTX Spark laptops and the GeForce Now streaming service.

Nvidia also announced that DLSS 5 is being introduced for Hogwarts Legacy on November 5, and also Where Winds Meet in the "near future", although we aren't given a specific launch date in the latter case.

Finally, away from DLSS 5, Nvidia announced that it's bringing in a new performance mode for frame generation, speeding up the tech considerably. The company explains that this benefits from a "faster and lighter algorithm" that boosts frame rates even further, and that it'll offer up to 35% better frame rates with an RTX 5060.

This performance mode is called 'Preset C' in the Nvidia app's frame-generation override.

DLSS 5 is also coming to RTX Remix on October 23, meaning old games can get a graphical makeover with added neural-rendering goodness. That only makes sense. given that gamers have been fudging support (in various ways) with classic PC games anyway.

Analysis: a laudable step forward — but gamers want more

It's remarkable to see DLSS 5 performance being boosted so quickly, although Nvidia needed to act in this regard. One of the main criticisms of this tech has been that the performance hit is too hefty, and it requires a powerful RTX 5000 GPU to run smoothly enough — at a time when the prices of such graphics cards are getting rapidly more expensive.

That said, an increasing number of gamers seem to be accepting that DLSS 5 is worth the cost to the frame rate for the visual uplift. Those folks will doubtless be pleased with Nvidia's progress here, and likely wouldn't disagree with Team Green's assertion in its blog post that 2026 has been a "year of relentless DLSS innovation."

Of course, there are still skeptics who'd rather Nvidia didn't bring AI into play in this way with DLSS. And it's also true that these early optimizations are where the larger gains will be carved out, and as the AI model is further refined, performance improvements are likely to become a lot more modest.

There are also gamers out there who are asking whether, now that the AI model is much improved, this going to usher in wider support for Nvidia's older graphics cards? As this Redditor put it: "Great so could we get official support on the [RTX] 40 series now since the new model is so efficient?"

It's a good question, given that Nvidia has promised that after tuning RTX 5000 performance it's going to work on bringing support to the last-gen RTX 4000 range. I'm guessing this won't happen until early next year, though.

The other work on DLSS frame generation is obviously welcome, too, albeit the new 'Preset C' option has attracted some criticism for causing a bit more glitching and artifacts while boosting frame rates by a decent amount. Not everyone is seeing this, mind, and how obvious it is may be somewhat subjective going by reports (and it could obviously depend on the specific game, too).

At any rate, this preset should be useful for laptop gamers looking to further bolster frame rates with relatively modest Nvidia GPUs — and indeed, it's likely tied into the release of RTX Spark laptops to help their gaming chops.

• Dell Pro Max 18 Plus leads current local AI rankings, yet Nvidia's launch could unsettle results
• RTX Spark offers 128 GB shared memory for larger locally run models
• Spark's 300 GB/s bandwidth trails Dell's graphics memory at 896 GB/s

The Dell Pro Max 18 Plus still holds the title of fastest AI laptop, topping local AI rankings ahead of every other machine tested.

Its RTX PRO 5000 chip reached 185.1 tokens/s on the Phi model, 140.5 tokens/s on Mistral, and 119.7 tokens/s on Llama 3.

However this lead faces a test on October 16 2026, when Nvidia's RTX Spark laptops go on sale from several manufacturers, with preorders already open.

RTX Spark arrival

Microsoft is expected to field the Surface Laptop Ultra from $2,599, while Dell is bringing the XPS 16 Creator Edition from $3,799.99.

Asus, HP, Lenovo and MSI are also launch partners, while Acer and Gigabyte are expected to add models later.

Nvidia rates RTX Spark at up to 1 petaflop of FP4 AI performance, with 6,144 CUDA cores and up to 128 GB of shared memory.

The 1 petaflop figure covers only FP4 calculations that exploit sparsity, an upper limit that everyday workloads may not reach.

The company also claims the chip can run large language models of about 120 billion parameters locally, with a context window of one million tokens.

That is slightly above the 109 billion parameters that the Dell Pro Max 18 Plus can reportedly handle locally.

This does not automatically mean that RTX Spark will outperform Dell's workstation in AI workloads, as model capacity alone does not determine inference speed.

The price Dell buyers paid for the lead

The Pro Max 18 Plus offers Core Ultra 5, 7, or 9 processors and RTX PRO 1000 to 5000 graphics, with prices starting at $3,488.

Dell's top configuration, equipped with an RTX PRO 5000, 128 GB of memory and two 1 TB drives, cost $9,245.52 when the laptop launched.

A similar configuration later appeared at roughly $14,000, while the laptop now sells for almost $18,600, double its launch price after 14 months.

The top model is built around the Intel Core Ultra 9 285HX processor with 24 cores, a 5.5 GHz boost clock and a 55 W rating.

Its RTX PRO 5000 graphics chip has 24 GB of GDDR7, a 175 W power ceiling and a graphics clock of 1,807 MHz.

Memory comes from a single 128 GB CAMM2 module running at 6,400 MT/s, while storage consists of two 1 TB Gen5 SSDs configured in RAID 0.

The 18-inch QHD+ display runs at 2560 × 1600 and 120 Hz, covers 100% of the DCI-P3 colour space and contributes to a minimum chassis weight of 7.17 lb.

Nvidia's case for displacing Dell rests partly on RTX Spark's shared memory, which can accommodate AI models exceeding Dell's 24 GB graphics capacity.

However, Spark's shared memory delivers about 300 GB/s of bandwidth, compared with 896 GB/s for Dell's RTX PRO 5000. Spark also has 6,144 CUDA cores, fewer than Dell's 10,496

To overtake Dell, an RTX Spark laptop must beat its token-generation rates for Phi, Mistral and Llama 3 under the same UL Procyon tests.

Matching those results while consuming 45 W to 80 W of chip power could also improve battery endurance beyond Dell's recorded three hours and 39 minutes.

Via StorageReview

• Asus confirms the ProArt GR1X will use the 20-core RTX Spark N1X, mainly in 64GB and 128GB configurations, shipping before year-end via premium retailers
• The GR1X matches Asus's Ascent GX10 on size, weight, fans, and power brick, and joins RTX Spark boxes from Microsoft, Dell, Acer, Honor, amongst others
• The GR1X still has no price, while Microsoft's Dev Box lists at $5,999, DGX Spark at $6,950, and GMKtec's 192GB EVO-X5 Pro at $6,599

Asus has filled in more of the blanks on the ProArt GR1X, the RTX Spark-powered mini PC it first showed at Computex 2026.

The company now says it will sell the machine mainly in 64GB and 128GB unified LPDDR5X memory configurations through a limited set of premium retailers before the end of 2026, according to its recent launch announcement.

Reports also confirm a 240W USB-C power supply built to the USB Power Delivery 3.1 EPR specification, a weight of 1.48kg (3.26lb), and a chassis of 150 x 150 x 51mm, or roughly 1.1 liters.

Plenty of details sans a price tag

The same reporting claims that the GR1X will rival the Mac Studio, even as the published spec sheet now points to a single-chip option: the 20-core RTX Spark N1X with a 6,144-core Blackwell GPU and a 140W TDP.

Asus announced the device on June 1 as the ProArt Mini PC, with up to the same 20-core chip, up to 128GB of memory, and up to 140W of thermal headroom, and promised availability from Fall 2026, but it has yet to publish a price tag, even in recent updates.

In the recent update, it has solidified the specifications on offer; when Asus renamed the machine in late August 2026, it suggested several RTX Spark chip options, dropped the 140W reference, and listed 180W of USB-C power delivery. The 140W figure has returned as a TDP, the power input has risen to 240W, and the cut-down 18-core N1X appears to be reserved for laptops, which matches Nvidia's RTX Spark spec table, which lists only the 20-core part for desktops.

This might not be a completely new product, though, but rather something of a reskin: Asus's Ascent GX10, its Linux-based take on Nvidia's DGX Spark, also measures 150 x 150 x 51mm, weighs 1.48kg, and ships with a 240W adapter. Asus's October 5 release for a 64GB GX10 describes two 140 x 80mm fans, the same fan size the GR1X page lists.

The overlap runs deeper than the chassis: RTX Spark is based on the same GB10 silicon as DGX Spark, as The Register has noted, and Nvidia CEO Jensen Huang said in September 2025 that the N1 processor was "going to go into the DGX Spark."

None of the GR1X's headline performance numbers have been independently tested yet; its 1 petaflop of FP4 performance is a sparse figure, to say the least. Nvidia's DGX Spark datasheet, Asus's own GX10 release, and Microsoft's Surface RTX Spark Dev Box announcement all footnote it as relying on sparsity, while the GR1X page carries no such caveat. Dense FP4 throughput will be lower.

The promise of 120-billion-parameter models with up to 1 million tokens of context is conditional too, and Asus's October 8 release concedes that what fits depends on model architecture, quantization, and context length. How quickly those models run depends mostly on memory bandwidth, which Asus's product page doesn't list, even as competitors already offer nearly 1.5x the memory.

A beefed up threat from a rebadged APU?

ASUS's most direct competition does not come from Nvidia's partners at all, but from AMD. Its Ryzen AI Max+ 395, better known as Strix Halo, has been the default chip for 128GB mini PCs for some time, including AMD's own Ryzen AI Halo developer platform, which went on sale through Micro Center in June for $3,999.99.

Like RTX Spark, it pairs a 16-core CPU with a large integrated GPU, in this case a 40-CU Radeon 8060S, and a single pool of LPDDR5X, but its 256GB/s of bandwidth trails Nvidia's chip.

Its successor, Gorgon Halo, narrows the bandwidth gap and opens a capacity lead. The Register describes the Ryzen AI Max+ 495 as "essentially a factory-overclocked version of the Strix Halo APU," with clocks about 100MHz higher, but it supports up to 192GB of LPDDR5X-8533 at 273GB/s, matching DGX Spark and, by extension, ASUS's own offering. Up to 160GB of that can be allocated to the GPU, according to Framework and GMKtec, which is more than the entire memory pool of any RTX Spark machine currently being sold.

Unlike ASUS's offering, Gorgon Halo boxes are arriving with pricing in tow, and they are not cheap. GMKtec's EVO-X5 Pro sells for $6,599 with 192GB and 2TB, against a listed regular price of $6,799. Peladn's YO2 495 costs CNY 49,999 (about $7,454) in China, and Lenovo's ThinkCentre X Ultra starts at $3,699 when orders open in November, with up to 128GB and support for clustering up to four units. Framework's 192GB Desktop is still listed as coming soon, however.

The silence on pricing is the GR1X's biggest open question, and the memory market is not helping. Nvidia has raised its own 128GB DGX Spark to $6,950, nearly 75% above its launch price, while partner 64GB GB10 systems are expected to sell for around $4,999.

AMD's machines have followed the same curve. According to The Register, a 128GB Strix Halo box cost between $2,000 and $3,000 a year ago, and around $4,000 by the time AMD's own workstation launched this summer.

Asus's ProArt laptops start at $2,599.99, and Asus offers them in memory tiers as low as 24GB. The GR1X, by contrast, targets the 64GB and 128GB tiers, making Microsoft's $5,999 Dev Box, Nvidia's DGX Spark, and GMKtec's $6,599 EVO-X5 Pro its direct competition.

Until Asus publishes a price, confirms a bandwidth figure, and commits to a date, and reviewers can test its thermal and gaming claims, a 20-core chip and 128GB of memory are the least surprising things about the ProArt GR1X.

• A PC maker in the US says 'retail GPU pricing is expected to climb significantly'
• A further rumor suggests AMD just put up the cost of VRAM, which will affect the prices of graphics cards
• Rounding off a trio of fresh rumors is a claim that Nvidia has discontinued the RTX 5090

There have been a few developments in the GPU world over the weekend — and none of them bode well for the prices of graphics cards down the line.

For starters, Wccftech reports that according to the Board Channels — a reliable enough forum for chatter from the supply chain over in China — AMD has increased the cost of the GDDR6 video RAM it sells to board-making partners. This means that those graphics card makers will inevitably pass that cost on to consumers — if this rumor is accurate of course.

Bench Life further observes that more expensive VRAM isn't the only upward pricing pressure here, and that the cost of other graphics card components — including the main circuit board itself — is also rising.

Secondly, as Wccftech also noticed, we have a custom PC maker in the US, StinceBuilt, informing us on X that it is "starting to see increasing supply constraints on graphics cards" and that, "Retail GPU pricing is expected to climb significantly, and these elevated prices could stick around for quite some time."

Finally, VideoCardz highlighted another post on X from regular leaker MegaSizeGPU, who claims Nvidia will not allocate any more GB202 (Blackwell) chips to make the RTX 5090, and that all this silicon is going to be used in RTX Pro (AI) graphics cards.

In theory, that means it's the end of the line for the flagship RTX 5090, with its replacement (the RTX 6090 in theory) still quite some way off in the distance (a more recent rumor indicates that it's probably due to land in 2028).

Analysis: a cacophony of warning bells

Of all those sentiments, the last one needs the most seasoning, although it's believable enough that Nvidia is essentially canning the RTX 5090. By that I mean that if this is true, I don't think this will be an official (announced) discontinuation, but the way it'll play out is that Nvidia will allocate so few chips for the GeForce flagship that stock will effectively dwindle to next to nothing.

Arguably we're already seeing some evidence of this, certainly in the US where at Newegg, for example, you can't buy an RTX 5090 directly from the retailer, but only from third-party sellers (often in Asia). And the cost of the graphics card is tipping up towards $9,000, such is its scarcity — and you'll pay from £6,000 upwards in the UK, although at least stock is available to buy directly from retailers in that country. Still, I wonder how long that will last; and the idea of Nvidia pulling the plug on the RTX 5090 in a stealthy behind-the-scenes manner certainly feels like it's possible.

Meanwhile, the suggestion of AMD's Radeon GPU price rises backs up similar recent rumors, and the warning from the PC maker in the US that price tags are set to 'climb significantly' is another ominous addition to the warning bells that have been sounded of late.

I'm still very much of the opinion that it's a good idea to buy sooner rather than later, even though GPUs are already overpriced, because I fear that the situation could get substantially worse in 2027, as the grapevine indicates.

The key question is how long a recovery from these pricing pressures might take in the case of GPUs. Again, I have a suspicion that it might take a lot longer than we might hope.

• Samsung and five affiliates are putting $1 billion into Helix Digital Infrastructure
• KKR founded Helix in June 2026, with Nvidia and Vistra among investors
• Helix will secure power capacity, which limits how quickly data centers open

Samsung and five affiliated businesses are committing $1 billion to Helix Digital Infrastructure, deepening the group's involvement in the backbones of artificial intelligence.

Samsung Electronics will provide $500 million, while five other affiliates supply the remaining investment through the venture.

The commitment combines semiconductor expertise, cooling equipment, construction, computing services, batteries, and infrastructure financing within one corporate strategy.

Samsung brings multiple infrastructure capabilities together

Helix, established by KKR in June 2026, is intended to develop data centers while securing electricity and network connections

Its founding investors also include Nvidia, Kuwait Investment Authority, and US electricity company Vistra, adding financial and energy capabilities.

Samsung's involvement extends beyond its financial commitment because several affiliated businesses can contribute technologies and services needed to develop and operate data centers.

Samsung Electronics contributes memory technology, and its FläktGroup subsidiary provides cooling systems for data centers.

Samsung C&T can handle major engineering and construction projects, with Samsung SDS developing data centers and providing rented GPU computing capacity.

Samsung SDI adds batteries for uninterrupted power systems, supporting facilities where servers must operate continuously without electricity interruptions.

The group therefore controls several functions that competitors would typically obtain from separate equipment manufacturers, contractors, technology providers, and infrastructure companies.

Samsung says the arrangement could help it move from supplying hardware toward participating directly in building artificial intelligence infrastructure.

Helix also plans to obtain power capacity, making electricity availability an important part of its proposed development model

That approach gives Samsung exposure to infrastructure requirements extending beyond processors, memory products, cooling equipment, and computing systems.

Huawei and SoftBank appear closest among rivals

Samsung's breadth is a major advantage, and few rivals can compete with such an extensive combination of infrastructure capabilities.

Huawei is among the closest comparisons, with capabilities spanning Ascend processors, storage, networking, security, cooling equipment, and electricity infrastructure.

That range gives Huawei significant technical reach, although the company may not have a comparable $1 billion infrastructure commitment.

SoftBank also has semiconductor interests through Arm, Ampere Computing, and Graphcore, as well as plans for an artificial intelligence facility in Ohio.

Its combination of chip businesses and infrastructure ambitions gives it substantial reach, although matching Samsung would still require capabilities from multiple businesses.

Nokia offers networking products for artificial intelligence facilities and has attracted investment from Nvidia, but its role remains concentrated around connectivity technologies.

Ericsson's work with SoftBank focuses on artificial intelligence applications for cellular networks rather than developing complete computing campuses and supporting infrastructure.

The difference is that Samsung can draw on affiliated businesses for capital, components, construction, cooling, computing, and battery systems within the same corporate group.

Huawei may be closest in technological range, while SoftBank has substantial semiconductor resources and growing ambitions around AI infrastructure, but neither company can combine these capabilities through affiliated businesses as extensively as Samsung does.

• DeepSeek open sourced Ascend ports of its production kernel and communication libraries with Huawei's support in tow
• Published benchmarks ran on a proof-of-concept hardware kit DeepSeek says is not a publicly distributed release
• Two repositories are genuinely new and have received hundreds of GitHub stars versus Nvidia's thousands; the other announced components are updates to existing tools

DeepSeek has revealed it is open sourcing a set of infrastructure components for Huawei's Ascend accelerators.

DeepSeek published two completely new repositories, DeepGEMM-Ascend for matrix multiplication and DeepEP-Ascend for the all-to-all communication that mixture-of-experts models depend on.

The other components named in the announcement were upgrades: TileKernels, DeepSelect, and FlashMLA are existing projects that received Ascend-targeted updates, a distinction FourWeekMBA established from the GitHub creation dates.

An open source offering with limited use cases?

Reuters framed it as two Chinese firms deepening ties in search of an alternative to Nvidia, with Bloomberg called the centerpiece a programming language named TileLang, China's answer to CUDA.

The question is whether any of this matters when engineers are still on the fence about using these tools and access to Chinese compute remains, at best, elusive.

The biggest piece, as Bloomberg puts it, TileLang itself is neither new nor DeepSeek's own offering. It was open-sourced in January 2025, and its Huawei Ascend adapters were published in an external repository, tilelang-ascend, on September 29, 2025, a full year before the recent announcement.

While TileLang supports Nvidia's CUDA backend, it also supports Huawei's Ascend as an ecosystem backend, classified and maintained in a separate repository. One could therefore argue that TileLang is less a weapon aimed at Nvidia than a shared abstraction layer that half a dozen Chinese accelerator vendors are quietly standardizing on, with Huawei as the largest of several tenants rather than the landlord.

The new software has limitations: DeepGEMM-Ascend requires the Ascend 950 series and CANN 9.20. DeepEP-Ascend requires Ascend 950 silicon with UBMEM connectivity, so it targets a narrow audience beyond Huawei's own engineers and limited Chinese AI labs with access to that silicon.

Despite this, at a time when Nvidia chips are unlikely to be generously exported by the US or allowed in by China, Huawei's Ascend silicon has a large task on its shoulders: DeepGEMM-Ascend keeps the same Python package name and API shape as its CUDA sibling, and DeepEP-Ascend aligns its public buffer interfaces with the Nvidia version.

That is because the goal isn't a performance claim, but switching-cost reduction, and switching costs are precisely what Nvidia's moat is made of: it costs far too much for most developers to abandon CUDA or code their own alternative currently.

Public signals show this remains an uphill task; at the time of writing, DeepGEMM-Ascend has 515 stars and 39 forks, compared with DeepGEMM, which has a comfortable lead at 8,522 stars and 1,359 forks, for example. The story is similar for DeepEP-Ascend and DeepEP, with the latter surpassing 10,200 stars versus DeepEP-Ascend's 224.

The reputational problem Huawei faces here predates the repositories. ChinaTalk, citing the Financial Times, quotes a Huawei researcher describing CANN as making Ascend chips difficult and unstable to use, and a Chinese developer characterizing work on the 910B as a road full of pitfalls.

Epoch AI has reported that Huawei dispatches engineering teams to major customers such as Baidu and Tencent to help port CUDA training code and keep deployments running.

A frontier lab publishing its own production kernels partially answers that complaint, because it replaces vendor documentation with code that has survived contact with a model and can be optimized further. It is also, however, code that runs on one chip family, under one toolkit version, on firmware the rest of the world cannot download or even potentially access.

• An upcoming data center cited by previous Prime Minister Keir Starmer as part of the UK’s AI Strategy will no longer open as planned
• The Nscale site in Loughton, Essex, which counts Nvidia among its investors, had already raised eyebrows when it was revealed earlier this year to still be in operation as a scaffolding yard
• UK Power Networks has told Nscale that the electricity grid will not be able to produce enough power for the site until the early 2030s

A new AI data center – dubbed the UK’s largest AI supercomputer – is facing delays thanks to challenges supplying power to the facility. Built by Nscale, with backers including Nvidia, the neocloud installation was set to launch in 2027, but power issues have forced UK Power Networks to come clean over the data center’s intended viability.

Oddly, the site – which was highlighted by the UK’s recently replaced Prime Minister, Keir Starmer – was still operating as a scaffolding yard as recently as March 2026.

In order to operate, the Loughton data center in Essex, England, requires 90MW, around the same energy requirement as 315,000 homes. However, the UK’s National Grid has a backlog of over 300 data centers requiring power, with the total demand more than 60% greater than the UK’s current capacity.

The Nscale neocloud

Constructing the neocloud – a GPU-based AI cloud data center – is the first project for Nscale, a UK startup that is preparing for a £26 billion ($35 billion) public listing and has already raised around £2.55 billion ($3.36 billion) in pre-IPO funding.

However, while cited as a key example of the UK’s commitment to AI, the Nscale neocloud cannot operate without the 90MW supply.

National Grid, the company that owns networks in the UK, don’t comment on individual cases. Its representative told the Guardian: “We’re continuing to work closely with Neso, distribution network operators and customers to identify opportunities to accelerate connections where possible and support economic growth.”

But the power supply roadblocks materializing for the Loughton data center are not exclusive to that project.

UK power capacity concerns

Other data center plans around the UK are being held up by power supply problems. Regulator Ofgem revealed in June 2026 that 315 data centers were awaiting connection to the National Grid. They represent a demand of 73GW, exceeding the country’s entire capacity of 45GW by 62%.

The UK’s power generation is a diverse mix of gas, nuclear, wind, solar, and biomass. Some power comes from mainland Europe. Over the past 25 years, the UK has focused more on renewable energy resources, deprioritising nuclear, and using gas as the fall-back. However, the increasing demand for AI has revealed that strategy requires an urgent rethink.

Around 100 data centers are planning to build their own gas-fired power stations, according to reports earlier this year, which noted that 100GW of data center projects were awaiting connection to the UK’s power infrastructure.

• Those buying an RTX 5090 at Micro Center now have to fill out a declaration form
• This includes having a government-issued photo ID scanned by the retailer
• Some approve of the policy as it's an anti-scalping measure, but others view it as an annoyance or a pointless move

If you're buying an Nvidia RTX 5090, or a PC with the graphics card inside, over in the US, you may have to fill out an official form to declare that you don't intend to export and resell the GPU – or at least you will at Micro Center.

PC Gamer spotted the 'advanced computing product purchaser declaration' which a Redditor was forced to complete and sign – including a scan of their driving license for ID – when buying a PowerSpec AI90 workstation PC with an RTX 5090.

The Micro Center store was in Tustin, California, but someone else posting on the Reddit thread claims that the same form has been introduced at the Columbus, Ohio store.

Apparently this is a new policy that only started this week at Micro Center (but seemingly other retailers already follow this practice in the US, including Central Computers).

In case you hadn't guessed, this move is about trying to prevent scalpers from buying up the RTX 5090, or PCs containing it – which can be around the same price as the standalone GPU these days, and are sometimes the only way of getting this graphics card at all – in order to sell it on to China and turn a profit.

Analysis: a pain in the neck, or a valid anti-scalper tactic?

As some of the commenters on Reddit are quick to point out, this form isn't legally binding, and there are question marks over how any of this might be enforced. Somebody breaking the terms of the agreement is likely to just shrug their shoulders about the consequences, perhaps.

However, as PC Gamer notes, the use of the ID check is certainly a deterrent of sorts for GPU exporters. And what this is doing is preventing people who are making a short trip to the US buying an RTX 5090 and selling it on when they get back home for a hefty mark-up. Those people won't be able to provide a US address and government photo ID that can be verified as belonging to them, of course.

The fact remains that it's a bit of a pain for Micro Center's customers to have to go through this form filling and ID check, although that said, some of the folks on the above Reddit thread don't have a problem with the process given that it's an anti-scalping measure. Others are concerned about having to give a retail store personal information for an unnecessary reason as they see it, or they're just annoyed and think it's a waste of time. As one Redditor put it: "This is ridiculous."

Going by the Reddit thread, another measure Micro Center enforces is that there's only one purchase of an RTX 5090 GPU per customer full-stop, forever. Once you've bought one, you can never buy another, we're told.

• The lower-tier Shield TV has been officially discontinued
• The Shield TV Pro will stick around, but is now $100 more expensive
• The list price is now $299 (about £225 / AU$429 before tax)

One of our favorite streaming gadgets, the standard Nvidia Shield TV, is no more. We can't say we're surprised — we reported in the summer that it was unavailable via official channels and there were no indications of any restocking — but we are surprised that the Shield TV Pro is immediately getting a very big price hike now that the cheapest model is gone.

According to WCCFTech, the Pro is increasing in price by 50% from $199 to $299 (about £225 / AU$429). The reason, and the reason for discontinuing the base model, is AI: As Nvidia says in its announcement, "the cost of components, including memory, has increased substantially across the industry".

If only there were someone Nvidia could talk to about the AI industry driving component costs into orbit.

Nvidia Shield TV Pro: latest specifications and availability

Although it's now several years since its last hardware refresh, the Shield TV Pro is still a great TV streamer with a hardcore following. It's powered by the Tegra X1 system on a chip, which features a 256-core GPU teamed up with 3GB of RAM and 16GB of storage. It's ready for 4K HDR, supports Dolby Vision HDR, and its AI upscaler can deliver up to 4K at 60fps.

This is likely to be the last Shield TV Pro model for some time, because all the indications are that the component cost crisis is going to continue through 2027 at the very least, and Nvidia can already reach gamers on multiple platforms with its GeForce Now streaming service, which is available on all the major desktop and mobile platforms as well as smart TVs from the likes of LG and Samsung.

Nvidia also recently revealed its RTX Spark "superchip" for laptops and handhelds, featuring an ARM processor with an integrated RTX GPU. Nvidia says it's "a new beginning" for gaming laptops and small form factor PCs, and with big-name studio support already it looks like it could be a big deal for mobile gaming in the future — right now, it seems to be much more of a pro component.

But still, that makes the Shield Pro an increasingly niche part of Nvidia's world domination plan, so I guess we can be happy it's stick around at all for now.

When Nvidia last updated the Shield Pro we said it remained our favorite TV streamer — but that was at a price of $199 / £179 / AU$379. With a 50% price hike added and a new Apple TV model expected imminently, it might be wise to wait before you decide what your next TV streamer should be.

• Borderless Gaming is a tool that's sold on Steam
• It just got an add-on to easily fudge DLSS 5 to work with various PC games
• Reviews of the tool have turned rather colorful, as those who still hate DLSS 5 and all it stands for are chipping in

You can't have missed all the chatter about experimenting with Nvidia's DLSS 5 in all sorts of different PC games, and if you've felt the urge to join in but have been too intimidated by the idea of fudging the tech to work, a solution is now at hand.

VideoCardz noticed that Borderless Gaming, a tool that's sold via Steam – it's used to facilitate gaming in a borderless window, as the name indicates – now has DLSS 5 support. Or, as the developer put it: "DLSS 5 Neural Rendering for Everyone."

The dev notes that: "DLSS 5 Neural Rendering is now available in Borderless Gaming – without replacing game DLLs, modifying game files, or needing a game-specific mod. Install the effect, select the preset, and apply it to your game."

Borderless Gaming applies DLSS 5 directly through its own BGFX pipeline, the developer explains, rather than forwarding requests to Nvidia's DLLs.

In short, this is DLSS 5 being fudged in about as no-fuss a manner as can be, save for the fact that you have to buy and install the app, of course. This implementation shouldn't be confused with native DLSS 5 support, because it's simply a workaround to bring the neural rendering technique into play (with variable results as you might expect).

Note that the Borderless Gaming app doesn't actually come with this built-in, but after installing the tool itself, you need to also download the appropriate effect (DLSS5-NR v0.2.1 currently) from GitHub here (installation instructions are provided at that link).

Analysis: catching some of the DLSS 5 hate

According to the developer, this implementation can work on a wide range of GPUs – not just the officially supported RTX 5000 models – and an RTX 3090 can hit around 30fps typically with DLSS 5 turned on. Obviously the frame rate will vary massively depending on the game and graphics settings, but the gist is that you can get something playable on the RTX 3090.

It's a neat enough solution by the looks of it, although as ever, if you're curious, I have to issue my usual warning about third-party apps here: download and use them at your own risk.

Borderless Gaming is an app that's been around for a long time, but the Steam reviews have turned distinctly colorful of late – with some of that color revolving around this latest move with DLSS 5. As one reviewer puts it: "Added DLSS 5 for no reason but to make games look worse... Slop." (Hence the above clarification that DLSS 5 isn't actually added to the app itself at all).

Nvidia's DLSS 5 was controversial when it was first announced, although the reception to its neural rendering trickery has warmed up considerably since it went live in a game (NBA 2K27). There are still haters out there, though, as the Steam reviews for Borderless Gaming illustrate.

It's also a fair question whether Nvidia might take action against projects like this, which are going further in terms of tweaking and playing about with DLSS 5 tech. It's likely that for now, with DLSS 5 getting a greater level of positivity from many of the gaming experimenters out there, Team Green might let it slide – at least for the time being.

• Nvidia response to "recent security incidents" by tightening restrictions on autonomous agents
• New software and hardware safeguards impose barriers if agents go rogue
• More than 100 customers have already agreed to work with the principles

Nvidia hasunveiled the new Open Agent Safety Platform designed to prevent autonomous AI agents from 'escaping' and acting beyond their established boundaries.

Following several recent alarming reports of AI-led attacks, the company admitted that safety should not just rely on safeguards built into AI models and applications, and that additional compute and hardware-level controls should also be used as more solid barriers.

"Recent security incidents have underscored the need to equip organizations with open, customizable tools," the company wrote in its announcement.

Nvidia adds even more layers of control for autonomous AI agents

The Open Agent Safety Platform consists of two key controls – OpenShell, which Nvidia describes as a software-based "secure runtime boundary," and Sentry, a hardware-based "watchdog" running on its BlueField-4 DPUs.

"Safety and security require full-stack engineering. NVIDIA Open Agent Safety Platform brings together industry, researchers and public-sector organizations to share best practices, align on evaluation methods and foster international cooperation," company CEO Jensen Huang summarized.

While existing model-level safety mechanisms are designed to tell AI not to behave in a certain way, the software-level OpenShell implements technical boundaries to actually prevent an agent from ignoring those instructions. Nvidia also argues that Sentry can quarantine and stop an AI agent within milliseconds.

Over 100 customers such as SpaceXAI, Anthropic and Microsoft are already using Nvidia's software and hardware-based safeguards.

The news, of course, comes after several high-profile instances where AI agents have operated beyond their intended controls, however while these new Nvidia-backed safeguards may be new, the overall principles aren't. The company is just stressing the importance of principles we're already familiar with, like least privilege, isolation/quarantining and detailed monitoring.

"As we continue to discover the frontier of AI capabilities, we must accelerate discovery at the frontier of AI safety," Huang added.

Nvidia made waves earlier this year when it announced the Nvidia RTX Spark, a new ARM-based SoC for small form factor PCs and laptops, and no one should be more concerned than Apple.

The Cupertino tech giant has enjoyed a very strong position when it comes to laptop sales for a while now thanks to its own ARM-based silicon, the Apple M-series.

With the refreshed Mac mini now on sale with the new Apple M6 chip, there’s going to be a lot of interest from creatives and developers around these two platforms, so I’m here to break down everything you need to know about the Nvidia RTX Spark vs Apple M6.

NVIDIA RTX SPARK VS APPLE M6: PRICE & AVAILABILITY

Neither SoC from either company is available on its own, so how much you pay will be entirely dependent on the PC or laptop it comes in, with things like memory and storage upgrades playing an outsized role in the final price.

The Apple M6 is only available in the new Mac mini (2026) for now, and it starts at $899/£899/AU$1,449 for the 16GB/256GB model. Adding more memory and storage will cost you extra, with the highest-spec M6 Mac mini featuring 24GB of unified memory for an extra $200/£200/AU$300, and an upgrade to 32GB for an extra $380/£400/AU$600.

Upgrading to 512GB SSD storage will add $200/£200/AU$300 to the price, while 1TB and 2TB will add $500/£500/AU$750 and $1,000/£1,000/AU$1,500 to the price, respectively.

Meanwhile, Nvidia is relying on its third-party partners to produce the PCs and laptops that will feature the Nvidia RTX Spark, so the prices will vary greatly among manufacturers, as well as which RTX Spark N1X core configuration is used (there will be an 18 CPU-core/5,120 CUDA-core entry configuration, and a more powerful 20 CPU-core/6,144 CUDA-core configuration).

While many units have been announced from Dell, Acer, and others, there's been no firm pricing yet, and much like the Mac mini, the amount of RAM the PC or laptop is configured with is going to have a major impact on its price. The first Nvidia RTX Spark devices are expected to go on sale in October, and some early indicators put the starting price of RTX Spark laptops and mini PCs at around $2,000/£2,000/AU$3,000.

• Advantage: Apple M6

NVIDIA RTX SPARK VS APPLE M6: SPECS

Fortunately, when it comes to specs, we know just about everything between these two SoCs, so we can do a proper side-by-side comparison of the Nvidia RTX Spark vs Apple M6.

As you can see above, Nvidia RTX Spark has a lot of hardware advantages versus the Apple M6. With nearly twice as many CPU cores on the premium configuration, almost four times the maximum memory headroom, and larger memory bandwidth, the Nvidia RTX Spark chips start out well ahead of the Apple M6 when it comes to specs—and that’s before we discuss the RTX Blackwell GPU.

Apple does have an advantage when it comes to maximum clock frequency, however, running nearly 800MHz faster than the best Nvidia chip. Will it be enough though? That remains to be seen.

• Advantage: Nvidia RTX Spark

NVIDIA RTX SPARK VS APPLE M6: PERFORMANCE

So far, I’ve only been able to benchmark the Apple M6 myself, as the new Mac mini went on sale on September 22, so there’s not a whole lot I can say definitively about the Nvidia RTX Spark vs Apple M6 when it comes to performance, but we’re not entirely in the dark here.

The RTX Spark N1X is essentially the Nvidia DGX Spark SoC, though it is running at a lower wattage for laptops (45W to 80W TDP, roughly, compared to the desktop DGX Spark’s 140W TDP). For RTX Spark desktops and mini PCs, I expect the TDP for the N1X to come close to or match the GDX’s 140W TDP.

Since we do know a lot about the performance of the GDX Spark across various workloads, we can expect the RTX Spark N1X to perform similarly, at least in small form factor PCs. What’s more, we also have some leaked benchmarks for the two RTX Spark N1X SKUs from Geekbench 7.

Note: RTX Spark N1X 3DMark scores estimated from RTX 5070 and RTX 5070 Ti laptop GPU benchmark results

As you can see, the Nvidia RTX Spark chips handle themselves very well against the Apple M6 in terms of GPU performance, though I need to caveat those results since they are approximations based on the listed core counts for the two SoCs’ GPUs. We don’t have benchmarks for these chips yet in 3DMark, but the GPU is built on the same architecture as Nvidia RTX 50 series laptop GPUs, and so we can ballpark where the N1X’s GPU is likely to end up given the number of CUDA cores they have available.

It’s also worth noting that the Nvidia RTX Spark’s multi-core performance definitely benefits from the increased core count over the Apple M6. However, the Apple M6 has a substantially better single-core performance, clocking in about 56% faster than the premium 20-core Nvidia chip.

This is going to translate to substantially better in-app performance for things like Adobe Photoshop and will help with agentic AI when an agent is processing a prompt or reasoning out a response, as these are inherently single-core dependent.

However, agentic AIs rarely work alone without some assistance from sub-agents, sandboxes, and other concurrently operating tools, which is where the multi-core performance of the Nvidia RTX Spark will definitely come in handy.

Of course, it also goes without saying that Nvidia’s distinct advantage over Apple when it comes to AI hardware thanks to native CUDA processing can’t be understated. Apple’s M6 might perform very well, but it will not be able to keep up with an Nvidia GPU that has up to 128GB of available memory.

And for gaming, well, Apple has really struggled to break Window’s stranglehold on PC gaming, and the Apple M6 doesn’t change that in the slightest, giving the Nvidia RTX Spark the edge there almost by default.

• Advantage: Nvidia RTX Spark

NVIDIA RTX SPARK VS APPLE M6: EARLY VERDICT

It remains to be seen how well the two Nvidia RTX Spark chips fare in real world testing, but early signs are very promising for Nvidia’s first SoC for the consumer market.

Apple isn’t out for the count just yet, either, and it will still hold the edge on price since the Nvidia RTX Spark is going to need a lot of very expensive RAM to process those 170B parameter models.

If you’re a more mainstream user who is just looking for a great PC for productivity, school work, and the like, the Apple M6 is the platform to go for.

Nvidia’s RTX Spark appears to be far more of a professional product with some gaming bolted onto the side, and it will surely have a price to match, so you only really need to consider the Nvidia RTX Spark if you can tell the difference between inference and model training. If you have to look those terms up, the RTX Spark isn’t really for you, but if those two are exactly what you plan on doing with your next device, you really can’t do any better than the Nvidia RTX Spark - as long as devices that include it are reasonably priced.

Normally we get to this time of year and the typical advice is to hold fire on any potential tech purchases. This is because Black Friday is just a couple of months away, and so it always seems prudent to think about being patient, and the bargain you might just secure in the biggest tech sale of the year.

Not this year, though, when it comes to GPUs – at least in my opinion.

During Black Friday 2025, we did see some decent deals on graphics cards. Not many, mind, but there was a scattering of them. I've noticed on Reddit that there are some gamers who have found a GPU with a decent price tag, but they're wondering whether to hold off and wait for something better on Black Friday. They are perhaps remembering some of the discounts that happened last year (thin on the ground as they were).

For example, this US-based Redditor asks: "Buy an open box [Nvidia RTX] 5070 now or wait until Black Friday?" They add: "I see a few open box deals for $745. What do people think about snagging this now at such an inflated price, or waiting until Black Friday to try and snag under $700?"

If you're in a similar quandary, my advice is this: absolutely do not wait for Black Friday. Just buy that graphics card now.

Let's break down the reasons why this isn't a 'waiters' market (by which I mean those biding their time, not the lovely people who serve us food and drinks).

Reasons to be fearful

For starters, as a legion of cynical Redditors are quick to point out, Black Friday isn't the best for graphics card deals anyway – it has never been the strongest suit for this sale in terms of PC components.

As noted, there have been bargains in the past though. Despite those Redditors who insist that GPUs just have their prices quietly upped in October so they can be artificially 'discounted' for Black Friday, there are actually some genuinely decent deals – or there have been.

However, things have just gone crazy this year, and as the second half of 2026 has rolled around, the RAM crisis has reached another tiring peak of frenzied hiking.

There were more rumors of imminent GPU hikes just a couple of weeks ago (following last month's negative chatter), and at retailers in the US we've already seen graphics card prices rising considerably over the past month. Compounding this are even more negative predictions about the memory shortage in general – which inevitably affects video RAM as well as system memory – and DDR5 prices which are spiking in a ridiculous manner right now. So you've got a recipe for a tumultuous rest of the year.

In short, I can't see 2026 panning out any other way than seeing more GPU price increases – so why wouldn't you buy now? At least from the perspective of whether Black Friday might somehow bring some bargains around – it surely won't.

I'm pretty convinced of that, and there are nuances here in terms of what kind of graphics card you're looking to buy. I'm talking about mainstream GPUs, the mid-range or lower tiers of the market, and not the higher-end graphics cards.

The beefiest GPUs like Nvidia's RTX 5080 (don't even mention the RTX 5090) are already exceptionally pricey, and while I don't think they'll get any cheaper – and indeed I reckon they'll probably still creep up in price – the cost of those products is already a very bitter pill to swallow.

I don't blame you if you don't want to pay out that much cash for a graphics card, because over $1,500 for an RTX 5080 (in the US) is a pricing nightmare. But you'll have to accept that you'll likely be in for a long wait for better prices, maybe not until 2028. A compromise might be to buy a card like an RTX 5070 to tide you over, or maybe an AMD RX 9070 XT (it's a touch less expensive in the US at the time of writing), then sell that second-hand later on when the RTX 5080 comes down.

The trouble is, there are no good options here... not for the pricier GPUs anyway.

Red Friday

I think that this year, Black Friday should be renamed Red Friday for its Herring-like nature, at least when it comes to graphics cards. Don't be swayed by the temptation of waiting for a deal, and if you see a reasonable price on the GPU you're considering, grab it now.

A $750 (open box) deal on an RTX 5070, the dilemma of the aforementioned Redditor, seems a sound enough proposition, even though that's a depressing reality considering this GPU was still going for around $650 just two months ago – because you're now looking at $850 or so (going by Newegg US pricing).

You can pick up an RTX 5060 Ti 8GB for only 10% more than this GPU was priced at in June (again, going by Newegg US), and I think there's an even stronger case to snag something like this if you're looking at the more affordable end of the market.

Obviously, do your own research based on the regional pricing for the GPU you're thinking about, but whatever the case, I'd still urge you not to be fooled into waiting for Black Friday, as I really think that'd be a mistake.

• Alibaba's T-Head unveiled the Zhenwu V900 with 216GB of memory, 1.2TB/s links, and a claimed 3x gain over the already-shipping M890
• Very few details associated with its raw performance: Alibaba gave no FLOPS, node, foundry, or power figures, and both specs it did share project it will trail Huawei's Ascend 960DT, which releases in the same quarter
• CEO Eddie Wu targeted more than 20GW of Alibaba Cloud-operated capacity by 2032 while admitting supply-chain shortages are limiting growth for the company

Alibaba opened its annual Apsara Conference in Hangzhou with three big numbers: a chip that it says triples the performance of its predecessor, AI clusters that could stretch to 500,000 of those chips, and a pledge, quoted by the South China Morning Post, that the data centers Alibaba Cloud operates worldwide will pass 20 gigawatts of capacity by 2032.

CEO Eddie Wu called the new Zhenwu V900, designed by the group's T-Head chip unit, the "most powerful AI chip in China today."

This was part of an address that distinguished Alibaba's offering as a GPU rather than an ASIC, landing it squarely in Nvidia's territory even as the latter's H200 GPU is highly sought after by AI labs.

A powerful offering on paper, but one that doesn't publish actual performance

The Zhenwu V900's spec sheet might be intentionally short, even as its performance figures versus comparable current-generation Nvidia GPUs, a banned option, could limit interest ahead of launch.

Alibaba has, however, added an impressive amount of memory and interconnect, allowing it to cluster a configuration it claims can run as many as 500,000 GPUs in parallel.

The V900, as per Zhidx, reporting on the launch, lists 216GB of memory and 1,200GB/s of chip-to-chip bandwidth and was originally expected to be available to clients in Q3 2027 before the recent revision pushed it two quarters earlier.

T-Head vice president Gao Hui now says the company will, going forward, keep a one-generation-per-year cadence, with a J900 due in the third quarter of 2028.

Despite the closer timeline, it is interesting that Alibaba has published no FLOPS figure, process node, foundry, power rating, or even the V900's internal memory bandwidth, making it hard to gauge its most powerful single-chip offering in terms of raw AI and compute performance.

This is particularly ironic at a time when the competition from Huawei, the Ascend 960DT, which was launched the same quarter after being pulled earlier by three quarters, is offering far more detail.

It lists 2 petaFLOPS of FP8 compute, 288GB of memory, 9.6TB/s of memory bandwidth, and a 2.2TB/s interconnect, making it a higher-VRAM and higher-interconnect offering based on the information Alibaba has already published.

The 500,000-card figure interestingly may just be a scaling limitation, not a system anyone has built. For scale, T-Head said in May 2026 that it had shipped just over 560,000 Zhenwu chips in total. One maxed-out V900 cluster, in other words, would need nearly as many chips as the company had shipped in its entire history.

It does, however, bring a secondary calculation into the equation: if each V900 drew what Nvidia rates its H200 at, up to 700W, the accelerators alone in such a cluster would pull a significant amount of power. A newer, larger chip could run hotter and be more power-hungry than Nvidia's Hopper offerings, even as China remains limited to older process nodes because of US sanctions on chips and chipmaking equipment.

A large cluster of 500,000 or similar GPUs would undoubtedly rack up power requirements considerably fast for the regional AI juggernaut, and 20GW by 2032 is not an unforeseeable event when one considers that per-GPU power draw is only heading up for the foreseeable future as the chips get faster, denser, and more monolithic at the same time.

Wu said medium- to long-term demand far exceeds what Alibaba can supply, and that shortages across the AI data center supply chain "are currently limiting the speed at which we can scale our compute infrastructure." The target covers capacity Alibaba Cloud operates, and the company did not say whether it means IT load or total facility power.

A Nomura note relayed by Sina puts Alibaba's current footprint at roughly 4GW and expects future buildouts to blend owned, leased, and third-party-financed sites, implying about fivefold growth in six years.

Whether the upcoming V900 displaces Nvidia's global market dominance, even slightly, is unclear, with Team Green already far ahead in a market where AI customers buy GPUs years in advance. The situation, however, is vastly different in Beijing.

In China, much of the disruption has already happened, with Chinese AI labs being 'persuaded' to stay away from Nvidia solutions by the state. Nvidia's latest 10-Q says shipments of data center Hopper products to China were under 1% of its $89 billion in data center revenue for the quarter to July 26, and the company said in its earnings report that objections in Beijing kept it from selling the full volume of H200s that Washington had licensed when relations briefly thawed.

Chinese labs have turned to a mix of homegrown solutions, including specific ASICs, GPUs, and CPUs, allowing local players such as Alibaba, Huawei, and others to thrive in a market devoid of foreign competition.

With the US government restricting Nvidia's power limits and the Chinese state taking a dim view of local labs depending on foreign chips, Alibaba's V900 GPU doesn't need to beat Nvidia's Rubin GPUs or even its Blackwell offerings. It simply needs to take on and provide a viable alternative to a GPU that is two generations older: the H200 (based on Nvidia's Hopper architecture).

Its bigger challenges will be Huawei's locally produced Ascend 960DT and similarly configured chips from upcoming local chip design firms, as well as supply chain limitations that could restrict overall production and, by proxy, deployment for a company looking to spend billions on AI infrastructure in the coming days even if its "most powerful" status, which is currently consigned to a claim holds at the time of its release.

What happened to Jensen Huang? The Nvidia co-founder and CEO seems about as disconnected as a person can be from other humans' life experiences. That wouldn't be so concerning if it weren't for the fact that he runs the most valuable company in the world and is responsible for the hardware behind the most life-altering innovation in our lifetime: AI.

Huang, who founded Nvidia 33 years ago, has been a CEO for decades and a billionaire since 1999; he's also now firmly on the side of rapid AI development and deployment with little-to-no regulation, putting Huang firmly at odds with a growing legion of, admittedly, some other billionaires and CEOs — among others — who are calling for a foundational model development pause or at least slow down.

It's in this moment that Huang has been talking — a lot. He's sat down with multiple journalists for chats, but none now more notable than the extensive New York Times podcast with columnist Ezra Klein.

Unlike some of the other interviews, Klein used Huang's own description of the AI as a Five-Layer Cake: Applications, Models, Infrastructure, Chips. Energy to frame the conversation, which invited Huang to opine on all these critical AI bits, and I have to say, many of his comments were eye-opening.

Here are the most startling things Huang said and maybe why he's saying them:

'There are a lot of skills that don’t matter'

"There are a lot of skills that don’t matter."

The topic here was studies in China showing that while AI in education may initially help students work more efficiently, it, on average, lowers their test scores by almost 20%.

Huang agrees with Ezra that the use of AI is, while helping them work faster, possibly degrading student performance, at least in certain skills.

"The multiplication table is starting to be forgotten. Doing square roots, my goodness. Basic math is being forgotten. Does it matter?...Yeah. I don’t think it does. I don’t think it does."

Huang believes that even as we lose some skills, or as students become adults entering the workforce, they will gain other new skills. Of course, that's sort of a zero-sum game where you can easily replace one thing with another.

I don't know about you, but I do consider the multiplication tables to be a core skill and useful even if you don't work in a math-related field. Naturally, Huang and most AI providers would like to believe that, like the calculator before it, AI will handle this skill and do all the math for you. Perhaps. But at what point do we lose the ability as a culture to double-check AI's work?

'I actually don’t know my address'

This next one is actually connected to the prior quote but is worth calling out:

"My first confession, I actually don’t know my address."

Not his email address, not his phone number or his partner's phone number, but where he lives. The street number and zip code (I'll assume he knows the town and state).

Huang used this as an example of a skill he no longer needs, but admits that realizing it while pumping gas and needing his zip code (I'm guessing for credit card verification) panicked him.

Listening to this while driving my own car, I almost pulled over. "What?!" I yelled at my AI-filled iPhone 18 Pro Max, which was playing the podcast. First of all, how? Second, Huang could not have crafted a better comment to undermine this and many of his other comments.

The only way you don't know your home address is if you've been shielded from the act of entering it on documents and driving yourself home because you always have someone else doing it for you. Huang's lived experience is thoroughly disconnected from the average person, and yet the choices he's making impact most regular people.

'[China] manufacture[s] everything in volume. They manufacture smart kids in volume'

Throughout the long conversation, Huang comes off as an industry Pollyanna and wildly self-serving.

He's asked repeatedly about China's approach to AI and if and how the US should be competing with them and ensuring that the US doesn't fall behind in this critical race. To put his comments in context, you have to remember that Huang personally asked the White House to allow it to continue selling AI chips to China. Huang did note, by the way, that he at least sells new technology to US companies first.

Overall, Huang essentially never criticizes China and, in fact, seems almost in awe of its approach on most fronts, especially in its use of open-model community (calling it "super-vibrant), and how it's raising an army of people to build its AI future.

"They have so many scientists and mathematicians. The number of engineers they have, they manufacture that in volume. They manufacture everything in volume. They manufacture smart kids in volume."

I don't know if that last bit was a backhanded criticism of the US education system, but it's not like he added, "Of course, we are creating just as many smart kids in the US."

'all of the rhetoric and all the alarmism, all the doomerism, all of the predictions — they’re scaring people'

If Huang has any criticism, it's reserved for his US counterparts, whom he calls "alarmists" and "doomers".

"I want to see us not ruin the opportunity for the United States to benefit at the highest level. And notice all of the rhetoric and all the alarmism, all the doomerism, all of the predictions — they're scaring people."

Huang insists that these AI models are still just programs running on operating systems, and wishes people would stop infusing them with human attributes.

'Just because it comes from a scientist doesn’t make it scientific'

While Klein mentions most of Huang's partners and occasional alarm-sounders, like Altman, Modei, and Musk, Huang doesn't actually mention any of them by name. The 'Godfather of AI' and chief alarmist Geoffrey Hinton, though, does receive special mention.

In response to a question about Hinton's assertion that there's a 10% chance AI will end society as we know it, Huang calls the statement "irresponsible" and adds, "Just because it comes from a scientist doesn’t make it scientific."

In a way, Huang is right. After all, Hinton is still just a person who can bring personal opinions to the debate. Huang argues that the "10% chance is not grounded on science."

But isn't it? If Hinton is the Nobel Prize-winning person who introduced the world to deep learning, which helped trigger the generative AI revolution, isn't everything he says, in some way, based on science?

Instead of saying he understands the concern but here's why he's wrong, Huang just claims the foundation of Hinton's argument is faulty, and therefore his statements are not really worth addressing. Huang would simply like everyone, all the doomers, to stop scaring everyone.

'The A.I. supercomputers are super energy efficient, but they’re still going to use a lot of power'

Huang says he wants AI to benefit every company and person, and he tries to offer a reasoned approach to the growing outcry over data centers.

He agrees that if people don't want them in their town, "then so be it," and encourages companies to be transparent about the impact, though he argues that their "use of water is really efficient." In the same breath, Huang tries to have it both ways: "The A.I. supercomputers are super energy efficient, but they’re still going to use a lot of power."

If, in Huang's perfect world, AI companies do generate their own power (building power sources takes time, probably more than it takes to build data centers) and they somehow lower property taxes, maybe data centers could someday be a net positive. Most, I think, would argue we're not there yet, even as the number of data centers being built across the US explodes.

'I think we just have to acknowledge that we got ourselves really gummed up in climate change and sustainable energy'

More concerning is Huang's shocking perspective on climate change and fossil fuels.

"I think we just have to acknowledge that we got ourselves really gummed up in climate change and sustainable energy, and as a result, we just didn’t plan enough energy production."

Klein, naturally, asked what Jensen meant by "gummed up," and, yes, it got worse.

"Well, in the near term, energy production requires fossil fuel. And because there’s just so much angst about fossil fuel energy production, if you look at our country, we’ve produced very little net new energy for a long time."

There's a lot to unpack there. It sounds like Huang is downplaying real climate change concerns, something that might align with the beliefs of his pal and "drill-baby-drill" and climate-change denier US President Donald Trump.

In fact, Huang never addresses whether climate change is real or not and instead seems fixated on how the loss of more fossil-fuel-burning energy plants puts us behind China in the energy race, and, more problematically for him, at a time when AI needs a lot more energy.

'They’ve got to inflict an enormous amount of pain and suffering on you so that they can save you'

It's not all bad. Huang believes, "If you want a future that is sustainable, lean into AI”

And then he says this:

"Yeah. It’s kind of like, in order to save you, they’ve got to hurt you first — that’s the nature of surgery. They’ve got to cut you open to save you. They’ve got to inflict an enormous amount of pain and suffering on you so that they can save you. And so I think A.I.’s kind of like that."

As for what level of "pain and suffering" Huang believes we should endure, he didn't elaborate.

If you were watching or listening to the interview hoping for some encouragement or a more rational, middle-of-the-road approach to regulating and governing AI in ways that truly benefit all, you might've been disappointed. Huang does say that when the builders see AI going awry, they should stop it, and when an AI acts out of alignment, "they shouldn’t release the product. That’s the simple answer." But he never addresses the concerns over what appears to be happening before Anthropic, OpenAI, and potentially others release their models. These systems are jumping fences in testing.

Huang may not understand the needs of the common person, but that pales in comparison to the dispassionate AI, which neither knows nor cares about us. We need leaders like Huang to start caring, but first they have to understand us and our very reasonable concerns.

• Tektronix's DPO718AX brings eight synchronized channels of up to 25GHz each to the 7 Series and can double as a Windows 10 PC
• The EPYC 3351, T1000, 96GB RAM, 1.6TB SSD configuration the oscilloscope offers makes it more than capable of running most modern OSes
• Prices start at an eye-watering $308,000 for the base configuration, which can be further modified as per end-user needs

Tektronix has added an eight-channel model to its flagship 7 Series oscilloscope line, and much of the attention it has drawn has gone to the computer it houses as part of the configuration.

The DPO718AX hosts a 12-core AMD EPYC processor, an Nvidia T1000 graphics card, 96GB of RAM, a removable NVMe SSD with at least 1.6TB of capacity, a 15.6-inch 1080p touchscreen, and an optional Windows 10 drive, all in a chassis that Tektronix's datasheet puts at 38.1kg.

This has drawn sharp comparisons to actual Windows-based PCs amid a RAM and memory shortage, including one by CNX Software, which said that the instrument "looks like a high-end PC running Linux (or Windows 10 LTSC)."

A powerful PC in its own right, albeit with dated components

None of the computer hardware listed for the DPO718AX is new or recently released. The AMD EPYC Embedded 3351 CPU is one of the oldest hardware offerings in play here, featuring AMD's first-generation Zen cores and a February 2018 release date.

While the SSD storage speed is unspecified, it is safe to assume it will be capped by the PCIe 3.0 support the EPYC CPU provides to the system. The 96GB of memory on offer can also safely be assumed to be DDR4, as the CPU supports only DDR4 configurations. Despite these limitations, the CPU does offer a 12-core, 24-thread configuration with a rated boost clock of 3GHz that should handle most tasks with relative ease.

The Nvidia T1000 GPU is an entry-level, workstation-class GPU built around its Turing architecture, with a release date of May 2021. The oscilloscope comes with an FHD touch-screen display and can output video up to QHD via DisplayPort, with a capped refresh rate of 60Hz.

The oscilloscope can also run Windows 10 natively, and it matters more than you might think: it is required for the compliance and RF analysis software the DPO718AX supports. The DPOs do not come with an installation by default; the Windows 10 LTSC 2021 installation is available on a secondary, removable drive that users can opt for if needed.

Since the integrated AMD CPU supports TPM 2.0, you could theoretically run Windows 11 on a device that costs hundreds of thousands of dollars, but the aging hardware, low-clocked CPU, and dated internals would make it a terrible PC replacement, even if one had the money to splurge on such an endeavor.

• Nvidia Smooth Motion has been modded onto RTX 3000 series GPUs
• The experimental mod allows driver-level frame generation in unsupported games
• Nvidia hasn't officially implemented this and other features on older RTX hardware yet

Modders continue to find new ways to implement Nvidia's latest technologies on older hardware, and now it's been done with Team Green's driver-level frame generation.

As reported by Wccftech, Nvidia Smooth Motion has been unofficially made to work on RTX 3000 series GPUs, with a new NVSmooth30 mod available on GitHub, confirmed to work on an RTX 3050 laptop and an RTX 3080 desktop GPU.

Nvidia's Smooth Motion functionality is essentially identical to AMD's Fluid Motion Frames, meaning that it enables frame generation in games that don't have official support for it — a tool that is most useful for older games or games running in emulators.

The mod works by modifying Nvidia's driver-level frame generation DLL (Dynamic Link Library) file, labeled NvPresent64, bypassing the usual compatibility checks that would stop it from working on Ampere GPU hardware, aka RTX 3000 series GPUs. It also works by "conservative metadata retargeting" from Ada Lovelace (RTX 4000 series) and targets Ampere instead.

The creator says that all users need to do is place the DLL file into the relevant game's folder and use the Nvidia Profile Inspector app to enable Smooth Motion. The process is very simple, but this is an experimental project, so it runs the risk of game bugs and crashes, or potentially other issues — and must not be used in online games using anti-cheat (doing so could result in a ban).

Gamers want new features, but new GPUs are unaffordable

Nvidia has been quite impressive with its latest technologies, notably with DLSS 5 Neural Rendering (which I saw at Gamescom 2026), and Smooth Motion has been one of many over the last few years.

The GPU market is in complete disarray due to the RAM crisis, which means prices for GPUs for both AMD Radeon and Nvidia GeForce GPUs are way too high for most people to consider an upgrade. Features like Multi-Frame Generation, Dynamic Frame Generation, and Smooth Motion are those that PC users would love, but can't use on older hardware.

Modders have consistently proved that these features can be backported in some way, with some even going as far as porting DLSS 5 onto the RTX 4080 (despite the performance hit).

DLSS 5 Neural Rendering coming to older hardware in the immediate future is unrealistic — but Nvidia has previously told me that it does plan on adding backwards compatibility at some point, though it's likely to come to 4000-series models first, and then we'll have to hope for 3000-series later in the future, if it's even possible.

But it feels as though features such as Frame Generation, and specifically Smooth Motion, deserve backporting attention now more than ever, as users are stuck with older GPUs.

Doing so would put Team Green in the good books of many PC gamers, and keep players from relying on unofficial mods to achieve better performance and enjoy GeForce hardware to the fullest extent.

I can't stress enough that Control Resonant is an absolute marvel in terms of visual enjoyment. Remedy Entertainment has crafted one of the most immersive visual experiences in gaming, doubling down on its predecessor's trippy, psychedelic, and mind-bending nature, and taking it to a new level using Nvidia's path tracing.

Putting my Nvidia GeForce RTX 4080 Super to the test, I ran Control Resonant on LG's 39-inch 5K2K OLED monitor, and I was pleasantly surprised by the significant enhancement to every detail on screen (albeit it did leave my GPU rather toasty).

At a 5120 x 2160 display resolution which is a 21:9 aspect ratio, you're getting 33% more pixels than a regular 4K gaming monitor at a 16:9 ratio, so the demand on the GPU is harder than most can handle — and even though Control Resonant was cooking my GPU, the frame rate was quite surprising.

With DLSS 4.5 Quality and Frame Generation enabled, and the high graphics and high ray tracing presets, performance remained fairly consistent within the high 50fps to low 60fps range. At 5K2K resolution, that's very reasonable, even if Control Resonant relies on Frame Generation to keep the gameplay smooth.

Dropping DLSS down to Performance mode pushed the frame rate to the high 80fps mark, and it's worth noting that it remained consistent even in moments of intense combat with tons of foliage and objects involved.

It's worth noting that the graphics with standard ray tracing are still fantastic, and the high (or even the medium) preset is likely the best route for any PC gamer who is using a low- to mid-range GPU (with Frame Generation). Lighting in all areas, specifically those where the Hiss is present (the really bright red zones), is phenomenal, and frankly, that applies to the experience without any ray tracing at all.

Path to victory

However, the star of the show is path tracing, which can be enabled by switching to the ultra ray tracing preset. I instantly knew my GPU was in for a rough ride once its temperature shot up to nearly 80C — but that's to be expected when running path tracing on a 5K2K display.

With DLSS Performance and Frame Generation enabled, I was able to achieve a solid 60fps in most scenarios, with occasional dips into the high 50s during traversal and combat. This aligns well with Remedy's system requirements for Control Resonant, which require an RTX 5080 for 60fps at 2160p using DLSS Performance, as the RTX 4080 Super is relatively close to this current-gen GPU for performance.

Realistically, this setup is out of the question for most gamers, as it will cost around $3,000 / £3,000 (or potentially more). LG's display is very expensive, sitting at $1,799.99 / £1,499 (on sale I should note), but an Alienware AW3926QW 5K2K alternative might help lower the cost.

The RTX 4000 series is discontinued, but pricing for anything equivalent to the RTX 4080 Super in the current-gen line-up (an RTX 5070 Ti or RTX 5080) would cost at least $800 (much more for the latter GPU), and then you need to add the cost of the other PC parts.

Handheld gaming PCs are still expensive, but more reasonable than an entire new PC build — and I tested Control Resonant on my Lenovo Legion Go S Z1 Extreme. That portable managed a solid 40fps on its handheld preset, and 70fps when dropping the resolution to 900p and enabling FSR Balanced with frame gen.

Overall, running Control Resonant at 5K2K is very demanding for any gaming PC, as expected, but I walk away shocked that it's even remotely possible given how much it elevates the gaming experience.

• XMG has announced price rises for its laptops due to the increased cost of components
• The laptop maker points out that DDR5 RAM for laptops is now close to 6x the price it was before the memory crisis
• We're also hearing that Micron has pulled the plug on its 2GB modules of GDDR7, and Nvidia's supply of VRAM could feel an impact from this

The RAM crisis sadly keeps rolling on and finding fresh pricing misery to heap on us, and that includes new laptop hikes and a hint that Nvidia GPUs could be in yet another spot of bother.

VideoCardz spotted the development with laptops, as manufacturer XMG announced on Reddit that its notebooks are subject to price hikes of either €100 or €150 for most models, or €300 in one case (that's for a Schenker portable, which is XMG's parent company, so it's all under the same umbrella).

Some laptop upgrades have got pricier, too, with a €120 premium on an RTX 5090 configuration, and as XMG observed: "The reason [for the price rises] is another increase in purchasing costs. This is no longer just about RAM and SSDs: PCBs, electronic components, CPUs, graphics cards and video memory are also contributing to the higher cost level."

In other words, everything is subject to price inflation, including the motherboard and associated chips, as well as the mentioned key components.

XMG highlighted data drawn from Geizhals (which tracks German retail prices) showing that DDR5 RAM for laptops is close to 6x the price it was before the memory shortage began (in mid-2025).

The laptop maker also made it clear that discrete GPUs and GDDR7 video memory have become more expensive, as have the prices of AMD and Intel laptop processors.

Also on the subject of GPUs, Wccftech reports that another slight worry is that Micron is no longer making 2GB modules of GDDR7 video RAM which are used in some Nvidia RTX 5000 GPUs.

Uniko's Hardware flagged up this news in a post on X, noting that these chips have "reached end of life" and that this is the case for all Micron 2GB modules, with only 3GB GDDR7 chips now listed on the memory maker's website.

With Micron apparently no longer supplying Nvidia with those 2GB modules, that means more of a squeeze on the availability of video memory for RTX 5000 graphics cards, at least in theory. This could hit stock levels and put a bit more pressure on pricing (which is already climbing fast).

Analysis: yet more pressure

We need to season the claims around Micron here, as this isn't a definitive statement from the company itself, and more to the point, it's only a minor supplier for Nvidia's gaming GPUs. Team Green gets the majority of its GDDR7 from Samsung and SK Hynix, and so the impact here should be relatively slight.

However, I'd argue that any impact is bad news given the current climate which is rapidly worsening for GPUs, particularly those from Nvidia which are heading upwards at a rate of knots for some models. The RTX 5090 has got completely ridiculous, the RTX 5080 is similarly out of hand, and even the RTX 5070 is now feeling the pricing heat, with a 30% price rise compared to two months ago in the US (and a similar picture elsewhere).

So, even small extra pressures are really not going to help Nvidia's GeForce GPUs here.

As for the laptop price hikes, this is a sign of the RAM supply times, and XMG is following other makers who have had to jack up prices (like Dell for example, which just bumped up the XPS 13 in a roundabout way).

It's a commendable level of transparency from XMG here, with the company making it clear that it isn't turning any extra profit from these laptop price increases. XMG also observed that other factors are in play with increased costs for notebook makers, and that includes rising air freight costs (due to conflicts in the Middle East) and difficulties with shipping by sea, too.

• Reports claim Apple is weighing an as-yet-unapproved enterprise AI server with two or four M8 Ultra chips, targeting 2029
• Apple is considering using Nvidia's NVLink Fusion to link the chips, with Thunderbolt-based Mac clustering falling short of data center bandwidth requirements
• The move comes as demand continues to be a key driver for Apple's higher-end Macs, with AI firms buying Mac minis and Mac Studios in bulk, even as Apple has turned enterprise buyers away

Apple has been working on plans for an enterprise AI server built on its own silicon with an interesting twist: it is seeking Nvidia's help on the networking front to make it work.

A report by The Information claims Apple is considering two configurations: one with two of its future M8 Ultra chips and a larger one with four, aimed at businesses that want to run AI inference on their own hardware rather than in a public cloud.

The timeline, however, is long and loose, with a potential launch in 2029, though the same report cautions that the project could be canceled before then, with neither the server nor any Nvidia arrangement finalized at the time of writing.

A market that Apple previously walked away from

If it ships, the machine would be Apple's first purpose-built server since the Xserve, which Apple announced it would discontinue in November 2010 and stopped selling at the end of January 2011.

Steve Jobs's explanation, in an email to a reader of French site MacGeneration, reported by MacRumors at the time, was blunt: "Hardly anyone was buying them." Gartner's data backed up that claim with an estimated 10,000 units being sold per quarter, a relatively low number for the tech giant whose numbers normally ran in the hundreds of thousands or millions for most of its other products.

Apple's suggested replacements to users were server configurations of the Mac mini and Mac Pro, desktops running server software rather than rack hardware. If the 2029 launch materializes, it would mean Apple's reversal would see it enter the server market 18 years after the Xserve stopped being sold.

This time, however, feels different: The Information has reported that OpenAI bought tens of thousands of Mac minis and Mac Studios to train AI agents through reinforcement learning, and that Anthropic has rented Mac minis through Amazon Web Services, as per Ars Technica.

Getting caught off guard by AI-centric demand

Apple reportedly was caught off guard by the AI-driven rush for its hardware, with no enterprise AI strategy, no business engineering team in place, and no specific developer relations staff to manage those orders. As a result, it also turned away businesses asking to buy access to its Private Cloud Compute infrastructure.

The project reportedly had a powerful sponsor: Apple's now-CEO John Ternus, when he ran Apple's hardware engineering team. However, shortages plagued it, leading many users to opt for Nvidia's DGX Spark and AMD's Ryzen AI Halo as viable alternatives.

The focus, however, remains on Apple's ability to interconnect a large number of chips or dies together: its current flagship, the M5 Ultra that ships in the new Mac Studio on September 22 2026, is Apple's first quad-die design, stitched together inside one package by the company's UltraFusion bridge, with up to 512GB of unified memory and 1.2TB/s of bandwidth, which users often stack for on-device AI needs.

A data center-esque, server-grade offering, however, would require considerably more power under the hood to meet its AI needs, which would also depend on how fast the interconnect between chips is; Apple's current limitation is Thunderbolt 5 speeds, which offer up to 10GB/s (or 80Gb/s) bidirectionally.

For context, Nvidia says the sixth-generation NVLink, available through NVLink Fusion, connects up to 72 accelerators at 3.6TB/s each. Comparing total two-way bandwidth, that is roughly 240 times a single Thunderbolt 5 connection's one-way capability of up to 15GB/s (or 120Gb/s).

Apple's own marketing hints at what that gap costs: It says four clustered Mac Studios deliver up to three times the inference performance of one, suggesting it already makes scaling tradeoffs of nearly 25% when linking devices.

This could explain why Nvidia is in the conversation at all: proven tech that could help Apple scale its AI server ambitions without the performance overhead it currently already has, at a time when its silicon would be even more powerful, might just be a no-brainer for the Cupertino-based giant.

• Apparently Nvidia's not launching RTX 6000 GPUs until 2028
• A leaker claims that these graphics cards have been delayed from the 2027 debut Nvidia originally planned
• Pushing back these GPUs makes sense given the RAM crisis, and some gamers are happy to wait anyway

Nvidia has pushed back its timeframe for launching next-generation graphics cards for gamers, and we're now hearing they won't arrive until 2028.

As ever, all this is theoretical – Nvidia hasn't uttered a peep about its RTX 6000 GPUs officially – but according to Kopite7kimi, a well-known leaker on X (via TweakTown), these products are now destined for a 2028 release which "means they have postponed again."

In other words, Nvidia was supposedly looking at a launch for next year, but the leaker now believes that isn't the case.

Another leaker, Moore's Law is Dead on YouTube, recently said that Nvidia is planning a 2027 launch for the RTX 6000 series, so this directly contradicts that idea. However, it's possible that this was the case, but Team Green has only just changed its mind.

Analysis: a sensible move

Of course, both of these assertions are just rumors, and we need to regard them with an equal amount of caution – but I think it makes more sense for Nvidia to delay the release of its next-gen GeForce GPUs until 2028.

Why? I have three predictable words for you: the RAM crisis. With the memory shortage – which also impacts video RAM – rapidly worsening as 2026 progresses, and predicted to descend further into nightmarish pricing territory in 2027, next year seems like a bad move for launching a family of next-gen GPUs.

Especially when you consider that the first graphics cards to emerge will doubtless be the flagship RTX 6090 and the 6080. How expensive will they be in such a climate, given that the RTX 5080 is already exorbitantly priced? The latter has reached around $1,500 in the US at this point, and is at similar levels in other regions, such as £1,350 in the UK (that's the cheapest GPU in both cases at the time of writing, I should make clear). And let's not even talk about the RTX 5090 which has got completely ridiculous.

Remember that these next-gen GPUs are going to require more memory to be an upgrade over their predecessors, and are also likely to use more expensive VRAM modules.

3GB modules will likely be used by Nvidia, or maybe even 4GB or 6GB chips, which as Wccftech reports, are rumored to debut in the next year or two (which would align with a 2028 release for RTX 6000, in theory). That's according to another prominent leaker over in Asia, Golden Pig Upgrade, as posted on Weibo.

All expectations are for the RTX 6000 series to be pricey, which is why launching them at what might be the height of the RAM crisis – 2027, going by current predictions – doesn't make a whole lot of sense. The following year seems a lot more likely, when hopefully the crisis has started to settle down and pricing is at least stabilizing to some extent (although some believe we're looking longer term for that to happen, more towards 2030).

Whatever the case, some gamers are getting jaded with the market. One of the replies to Kopite7kimi's post on X simply says of the next generation of Nvidia GeForce GPUs: "Nobody cares, nobody can afford it anyway." Others – like this Redditor, who says "I'd rather they delayed the launch to consumers" – indicate that they're happy enough to wait due to pricing or stock issues (launch quantities of GPUs would likely be thin on the ground).

The other question is whether we'll see those RTX 5000 Super refreshes that have been long rumored. I'm doubtful at this point, and I think Nvidia may be inclined to rely more on the AI and software side of the equation, principally DLSS 5, to impress gamers, rather than new (stupidly pricey) hardware.

• AWS deal with Nvidia adds 2 million Blackwell Ultra, Rubin, and Rubin Ultra GPUs, on top of more than 1 million committed in March 2026
• Totals more than 3 million Nvidia GPUs due before 2029, a ~1.9x multiple for each of Amazon's roughly 1.58 million employees
• Nvidia's CFO has said deliveries start this quarter, not in 2027, but neither company says what it counts as a GPU or what the deal costs

Amazon ended 2025 with about 1.58 million full- and part-time employees, most in fulfillment centers and delivery operations, with a much smaller share working in its AWS data center division (estimated at around 246,000).

By mid-2028, Amazon Web Services expects to have taken delivery of roughly twice that many new Nvidia GPUs, based on a recent deal it made with Nvidia.

The latest deal was announced on August 26, the same afternoon Nvidia reported quarterly earnings, and will see AWS plan to deploy two million additional Blackwell Ultra, Rubin, and Rubin Ultra GPUs across its global infrastructure in 2027 and 2028. This is on top of over a million Nvidia GPUs that AWS committed to at Nvidia GTC 2026 in March, and which Nvidia's Ian Buck told Reuters would ship from 2026 through 2027.

A mammoth deal that dwarf's Amazon's entire payroll

Together, this all amounts to over three million GPUs in play between the data center giant and the chip designer, which is also the world's most valuable company at the time of writing. At 3 million GPUs against a workforce of 1.58 million, that's nearly 1.9 Nvidia GPUs per employee.

This is only Amazon's commitment to Nvidia; it also has its own custom Tranium3 GPUs, which the company says are "nearly fully subscribed" or "almost fully booked" as of mid-2026.

At the same time, Amazon is not increasing job roles on its end; it cut around 30,000 corporate roles between October 2025 and January 2026, a restructuring its top HR executive, Beth Galetti, described as "reducing layers, increasing ownership, and removing bureaucracy."

Building on an already solid base

While AWS's current GPU base from Nvidia, consisting mainly of H100 and H200 GPUs, is smaller than Microsoft's 485,000 count, it is hardly a rounding error, with Omdia estimating that Amazon bought 196,000 Hopper GPUs in 2024 alone.

The interesting part is determining what counts as a GPU: Nvidia and Amazon's joint statement about 2 million GPUs across calendar years 2027 and 2028 might mean two very different things depending on how you count GPUs.

At GTC 2025, Nvidia said it would count Rubin-generation GPUs by compute die rather than by package, which is why the Vera Rubin rack was first unveiled as NVL144. It reversed course before CES 2026, and the flagship now ships as Vera Rubin NVL72, counting 72 two-die packages. To make matters a bit more complicated, Rubin Ultra, due in 2027, puts four compute dies in each package.

Whether the companies count each die as a separate GPU or the whole package as a GPU could make the numbers skew heavily in either direction. The deal also includes a national security element: approximately 100,000 GPUs are earmarked for US government AI workloads at Impact Level 6.

Amazon's limitation here is supply, not demand; AWS revenue rose 37% year over year to $42.2 billion in the second quarter, and its contracted backlog reached $496 billion, up from $364 billion three months earlier, as CloudTech News reported.

On Amazon's July earnings call, CEO Andy Jassy said capacity would fall short of demand this year and likely next, adding: "In fact, the demand we already have for 2028 is striking."

In addition to its own custom ASICs, including Tranium3, Amazon is gearing up for a few fiscal years of eye-watering AI infrastructure spend and AI demand combined, and it better be right: Amazon's free cash flow has already turned negative as it looks to fund its bid to be one of, if not the largest hyperscaler in play for the foreseeable future.

Via StorageNewsletter

• Samsung has backed a tiny Dutch chipmaker targeting Nvidia
• Euclyd raised $231 million before shipping its first commercial chip
• Euclyd plans to sell complete AI systems directly to enterprises but will not deliver its first hardware until 2028

Samsung has joined a group of investors backing a small Netherlands-based chip designer aiming to challenge Nvidia's dominance in artificial intelligence hardware.

The startup, called Euclyd, closed a funding round worth $231 million, drawing support from several firms beyond the South Korean electronics giant.

According to Bernardo Kastrup, Chief Executive of Euclyd, commercial hardware from the company will not reach customers until 2028

A chip built to rival Nvidia's grip

Euclyd's Series A round totalled €200 million and was co-led by Somerset Capital Partners, the Scaleup Europe Fund, and Innovation Industries, alongside Samsung.

In an exclusive interview with CNBC, Kastrup said that the capital would support two separate lines of business going forward.

One involves selling physical hardware and rack systems directly to enterprises wanting to run their own AI inference securely on-site.

The other involves licensing the underlying intellectual property to firms that want to build their own chips using Euclyd's designs.

Founded in 2024, the company is developing a chip system built around a different architecture than conventional graphics processors, covering both the processor and memory layers.

Nvidia's GPUs were originally built for video games but became central to AI training and inference and pushed the company to become the world's most valuable.

That dominance now covers the market for the very highest performing chips available today, though rivals are emerging quickly.

“AI is becoming a foundation of economic growth, scientific discovery and national competitiveness, but its potential will remain constrained unless we fundamentally change the infrastructure beneath it,” Kastrup said.

Why Samsung's involvement matters beyond cash

Kastrup revealed that Samsung's contribution extends well past the money it committed to the round.

"Samsung can help us in more ways than money. They are one of the biggest memory manufacturers in the world,” said Kastrup. “They do a lot of engineering, they know a lot about systems, they know the supply chain, they have a huge network."

Although Euclyd expects to commence shipping its chips in two years, the company plans to serve thousands of enterprise customers by 2030.

The company's systems remain unproven at any meaningful commercial scale so far, leaving real-world performance an open question.

Several major technology firms are pursuing similar paths away from dependence on Nvidia's processors for AI workloads.

OpenAI announced in August that its first in-house AI chip, named Jalapeño, delivered industry-leading speed and efficiency.

Google, Amazon Web Services, and Meta are each developing their own chips for internal AI use as well.

Whether a small Dutch startup with no shipped product can meaningfully dent Nvidia's position remains genuinely uncertain given the scale involved.

• Google and Nvidia are just two of the tech giants behind the new AI Energy Management Alliance
• The group is focused on building future data centers that coordinate power usage with the grid
• Moving training to off-peak hours and balancing on-site energy are some of the solutions

Emerald AI, Google and Nvidia have come together to form the AI Energy Management Alliance (AEMA), and it could be exactly what we need to tackle rising energy consumption within the data center field.

Together, the group hopes to deploy flexible AI data centers that can dynamically adjust how much electricity they consume from the grid based on a number of different conditions.

This important development comes in response to rising power availability pressures, which are leading to years-long delays for many new projects.

Future data centers could adapt how much power they use

In an announcement, Nvidia Head of Sustainability Josh Parker explained how these future campuses could shift compute workloads, discharge storage, use paired generation or respond to other system contingencies to dynamically change how much electricity they're drawing from the grid at any given time.

The concept involves shifting high-compute tasks like training to off-peak times, using on-site batteries or generation in conjunction with the grid and exporting excess stored energy back to the grid during times of stress.

Crucially, the AEMA must now set out technical specifications, performance metrics and open standards to allow this to happen at scale.

And while only Emerald AI, Google and Nvidia are behind the alliance from launch, a number of additional partners are set to join the scheme. SiliconANGLE reports that Anthropic will be one of them.

The concept itself isn't new – Google also announced earlier this year that it had over 1GW of demand-response capacity – but doing so at scale would likely make a considerable impact across the board.

Both Nvidia and Emerald AI have already developed software to handle power consumption in relation to the grid.

"By creating a common framework for performance, reliability and collaboration, AEMA aims to help the US build the infrastructure of intelligence at the speed and sustainability the moment demands," Parker wrote.

All of this, of course, comes against a backdrop of mounting energy pressures. Recent Bloomberg data suggests that US data centers alone could use as much natural gas as the entire of Germany and Japan combined by 2035, with a similar Moody's study projecting that total US data center energy consumption could top 426 TWh by 2030.

• The cheapest available US RTX 5090 now costs over $6000 against a $1,999 launch price, with third-party listings often running past $9,000
• The GPU that already pushed past $3,500 in Feb 2026 owing to memory constraints driving prices higher has now hit fresh highs at multiple retailers
• The move is linked to AI demand, as data center operators have entered the fray to secure compute despite Nvidia's GeForce driver license having banned 'data center deployment' on paper since 2017

Six weeks ago, the story was ultra-rare books being fed spine-first into high-speed scanners so that AI labs could train on text written before chatbots polluted the corpus.

Focus, however, has shifted considerably to AI labs bringing the same voracious appetite to Nvidia's flagship consumer-grade Blackwell GPU, the RTX 5090, which has essentially disappeared from shelves unless you want to pay a 200 percent or higher premium to grab one.

Photographs circulated by Hong Kong outlet HKEPC, and picked up by multiple publications including Guru3D, VideoCardz, and TechPowerUp, appear to show wooden pallets stacked with boxed, retail GeForce RTX 5090 cards at AI server assembly sites, waiting to be unpacked and slotted into multi-GPU rigs.

The sky is the limit for Nvidia's flagship consumer card?

This isn't the first time Nvidia's RTX 5090 GPUs have been deployed en masse on multi-GPU, AI-centric rigs, but this time feels different, even compared with previous shortages.

For context, Nvidia launched the RTX 5090 at $1,999 in January 2025. Tom's Hardware's price tracker, updated on September 14 2026, lists the cheapest available US card at $6,389, nearly 3.2x the former.

PCGamesN reported the same week that first-party stock at Amazon, Newegg, and Best Buy had essentially vanished, leaving third-party marketplace listings above $6,000 and, in some places, over $9,000.

The pictures in circulation don't show RTX PRO workstation GPUs, but rather GPUs from MSI, Gigabyte, and ASUS that the average gaming enthusiast or consumer would pick off a shelf.

An AI-generated image to create more FOMO?

The images have, however, been disputed by some quarters: Kevin Hofer at Swiss retailer Digitec, one of the few writers to examine the images rather than republish them, pointed at the box lettering in one frame and asked what a "50B0" or a "50R0" is supposed to be, concluding that AI was probably involved in producing the picture. Twitter has also added a "made with AI" moniker to the image in question.

TechPowerUp and TweakTown both carried the same warning in softer form. Guru3D, to its credit, noted that the report explains neither why these firms chose the RTX 5090 nor the scale of the purchases.

Coverage like this, even with disclaimers to the contrary, can create a "fear of missing out" (FOMO) where price increases become a self-fulfilling prophecy as businesses and consumers alike move purchases forward to avoid paying a further premium.

Examining the value that an RTX 5090 brings to the AI table

The Nvidia RTX 5090 is its most powerful consumer-grade card to date, packing a Blackwell-based GB202 chip and 32GB of GDDR7 memory, which checks multiple boxes for local AI use cases and gamers looking to max out games at 4K or higher.

Nvidia's GeForce driver EULA has, since late 2017, included a clause titled "No Datacenter Deployment" that should, at least on paper, prevent data centers from stockpiling or using these.

The chip designer, currently the world's most valuable company, has previously implemented technological restrictions such as LHR (Lite Hash Rate) on its RTX 3000 series GPUs to prevent GPU miners during the Ethereum gold-mine rush from hoarding supply.

It has also released a China-specific RTX 5090 D, which reduces AI performance while matching the RTX 5090 in gaming throughput. The RTX 5090 D, created to meet AI export restrictions, shows that Nvidia can limit the RTX 5090's AI performance, but it remains relatively indifferent to a market that was its core revenue source only a few years ago.

The RTX 5090's price hike, at least in part, may be due to the RTX PRO 6000's staggering, out-of-control pricing, which Nvidia nearly doubled to $16000. The underlying cost driving the price hike is the GPU's 96GB of VRAM, which also makes up a sizeable, if not dominant, chunk of its total bill of materials.

An RTX 5090 comes with 1/3rd the VRAM on offer, but nearly the same compute performance, and with AI-centric algorithms often assisting in breaking up tasks between GPUs and memory pools, even as the total memory pool remains similar, a triple stack of RTX 5090 GPUs should comfortably outperform the RTX PRO 6000 in certain tasks.

At the same time, Chinese OEMs have reportedly managed to slap on as much as 96GB of VRAM on an RTX 5090, which could make it an even hotter commodity in a country where both the RTX PRO 6000 and the RTX 5090 are export-restricted but do manage to make it on the regular.

The RTX 5090 is a manufacturing problem with a memory bottleneck attached for AI consumers, but with the alternatives having doubled in cost recently, AI consumers are looking to save money wherever possible, and both the stock RTX 5090 and a potentially modified one are some of the best candidates in a market that has already exhausted most of its last-generation supply of the same.

• Nvidia fits 28 GDDR7 memory chips onto one workstation graphics card
• The RTX PRO 5500 carries 84 GB of onboard GDDR7 memory
• Fourteen memory chips sit above the PCB and fourteen below

Nvidia has rolled out a new professional graphics card that carries an unusually generous pool of onboard memory for its class.

Nvidia says the RTX Pro 5500 workstation edition is built on Blackwell architecture and ships with a full 84 GB of GDDR7 memory.

That memory is attached to a GB202 processor carrying 21,760 CUDA cores, the same count found on the GeForce RTX 5090.

How Nvidia packed in so much memory

This particular graphics variant is certified strictly for professional workstation duties rather than for everyday consumer gaming applications and setups.

Reaching such an enormous memory capacity required an unconventional physical layout spread across the entire circuit board of this graphics card.

Nvidia mounted 14 GDDR7 modules on the front of the board next to the processor and another 14 modules on the rear.

That double-sided arrangement, often described as a sandwich configuration, squeezes a full set of 28 memory chips onto one compact card.

Each chip operates at roughly 27 Gbps per pin, which counts as a routine GDDR7 speed rather than the fastest option currently sold.

The resulting 416-bit interface yields about 1,398 GB/s of aggregate bandwidth, a weirdly narrow width given the enormous capacity involved.

This is unusual because GPU memory controllers usually operate in even pairs, producing familiar widths such as 256-bit, 384-bit, or 512-bit.

A 416-bit path implies 13 separate 32-bit memory channels, an oddly uneven number rarely seen in current graphics processor designs.

Buyers can also split the card through multi-instance partitioning, choosing either one full 84 GB block or two separate 42 GB segments instead.

Nvidia rates the board at a hefty 600 W power draw and builds it strictly for rack-mounted workstation deployment rather than desktop towers.

It ships in both air-cooled and liquid-cooled variants, and the air-cooled version relies on high-speed server fans bolted directly onto a sizable heatsink.

Where it sits against Nvidia's other workstation cards

Within the wider RTX Pro family, this graphics card lands just below the flagship RTX Pro 6000 in raw specification terms.

That top-tier card carries a much higher 24,064 CUDA cores alongside a full 96 GB of dedicated video memory as well.

It comfortably outpaces the RTX Pro 5000, a card limited to only 14,080 CUDA cores and a maximum of 72 GB.

Given those figures, the chip sits much closer in raw specification to the top-tier card than to its nearest numbered sibling below it.

Nvidia seems to be prioritizing sheer memory volume for large datasets over the kind of bandwidth efficiency typically prized in compute-heavy rendering workloads.

Whether that particular trade-off proves worthwhile in practice will likely depend heavily on the specific professional applications each buyer intends to run.

At the moment, there is no official price or launch date, but a steep asking price seems all but guaranteed once details emerge.

The most realistic customers for this card will likely be enterprise buyers rather than everyday consumers given its likely price and power demands.