The generative AI boom took off through creativity. Tools that could generate images, design assets, and written content were among the first to go mainstream, reshaping workflows for designers, marketers, and creators almost overnight.
Since then, AI tools have made their way into other areas like code and customer service, but its most visible and immediate impact remains in the creative world.
Billions are being invested, and adoption rates continue to rise, but with all this momentum come growing questions: is the current pace sustainable? What happens when regulation catches up? And who’s actually benefiting from all of this?
To dig deeper into those questions, I spoke with Joaquín Cuenca Abela, CEO and founder of Freepik.
As the head of one of the world’s most widely used creative content platforms, he is uniquely positioned to see both the promise and the pressure of AI in action.
Freepik has embraced AI to help designers and marketers move faster, but it’s also navigating complex challenges, from energy use to IP law to the expectations of enterprise clients.
In our conversation, we discussed how Freepik is adapting to fast-moving tech, what’s changing in user behavior, and where Abela sees the biggest risks and opportunities as AI becomes an accepted creative co-pilot.
- I interviewed Astronaut Tim Peake a few days ago and one of the things he told me was "I can't quite see the business model as to who's running these [AI] systems". Freepik is part of that ecosystem. So are we heading for a pop (as in bubble popping) or do you think the current 'fuite en avant' is sustainable for the foreseeable future (and is mostly a marketshare play)
If there is a pop in AI, it will be like the internet pop of 2000. Short term. AI is tremendously useful, and the market for AI products will keep expanding as we find more ways to integrate AI in our workflows.
I actually think that we are underestimating its impact. Companies are getting billions in revenue from real users that are getting a real benefit at a faster rate than any other industry of the past, ever.
- Gemini says that for current state-of-the-art diffusion models, the suggested energy consumption per image varies between 0.01Whr and 0.3Whr. Do you agree with these?
That really depends on the particular model and hardware stack that you’re using. A single B200 GPU (a little monster) can generate 1 high quality image per second, and it consumes circa 1 kW. That means that each image takes on average the energy consumed by an old 60W lightbulb in 16 seconds.
3 months ago the energy consumed per image was 3x that. The energy consumed per generation is dropping very quickly.
- What about video? What sort of power consumption are you seeing and what use cases are the most common on your platform.
A video is around 100x an image (equivalent to a lightbulb on for ~30 min). The most common use cases are product shooting and marketing campaigns.
A traditional product shooting involves flying a crew and equipment to an exotic place, and in some cases that can be replaced by a small team of experts working in an office.
- The quality of generative AI very heavily relies on training data. Navigating between licensing, lawsuits and partnerships is no small business. What are your views on how the market has evolved till now and where it will go next?
Correct licensing is still the single biggest hurdle for enterprises rolling out AI, which is why our end‑to‑end protection and legal indemnity - within our Enterprise plan - have been so warmly received by Fortune‑500 design teams. Clients want certainty that their works are fully compliant with IP laws.
Simplifying, there are three different options regarding the licensing of data: opt-in, opt-out, and all-in. As you mention AI relies on training data, and it needs gigantic datasets, like the whole internet. It’s just impossible to find the author of each piece of data used to generate a modern competitive model. In practice this rules out the opt-in model.
The weakness of the opt-out model is that it requires companies proving that they are not using content that has opted out, disclosing their datasets - today a major part of the secret sauce. Europe has chosen an opt-out model, and the need to disclose the datasets used is one of the major reasons why companies are objecting to the AI Act.
In the US it is currently under litigation. A recent California ruling said training on legitimately purchased books can be “exceedingly transformative,” putting training data under “fair use.”
I don’t know anything about the IP restrictions in the rest of Asia, but all the models I’ve seen from China have clearly been trained with a wide variety of datasets, just like all the competitive models in the US.
My prediction is that many major companies will rather shut down their AI in Europe than disclose their datasets and fully comply with the AI Act.
In the US I hope we will see some consolidation of the multiple federal laws, and have the recent rulings ratified by a higher court.
- ChatGPT launched in November 2022 and its fifth iteration (ChatGPT 5) is said to be imminent. As an external observer (and competitor), what do you think will be the key features of v5?
We are more a user than a competitor of ChatGPT, and we are happily using their image generator on our suite, along with many other great models.
An obvious weakness of ChatGPT is its complexity. It’s difficult for a user to understand which model they should use for which questions.
My guess is that it will be automatic in the future. ChatGPT today understands text, files, audio, and images, so I guess the next obvious step is to understand a video stream, and allow it to be on at all times.
- Everyone right now seems to be racing towards AGI. As one of the biggest players in the AI creative business, do you see any practical applications of AGI for Freepik?
The difficult thing about AGI is finding something where it cannot be applied. I believe it will be all over the place and I see many applications within Freepik, both internally or externally.
- What technology (either hardware or software) excites you the most right now when it comes to AI?
Coding agents. It’s the thing that allows recursive self-improvement of the technology. If we truly reach a point where AI can improve itself without any of the current bottlenecks, we will get unprecedented levels of intelligence to solve our problems.
As with any major progress, it’s equally thrilling and terrifying, but my prediction is that after a few years it will just be the new normal.
We will wonder how we lived in those barbarian times without personalized education, medicines developed for your particular body, and many of the wonders that we can’t just yet think of.
You must confirm your public display name before commenting
Please logout and then login again, you will then be prompted to enter your display name.