A small object makes rain or shine in the AI. The fate of a myriad of artificial intelligence start-ups depends on it. Her name ? The H100, Nvidia’s famous AI chip that everyone is chasing. Problem, Elon Musk quipped ironically, these chips are “harder to find than drugs”. A shortage which allows us to better understand the curiosity aroused by the French start-up FlexAI, which announced on April 24 that it had raised 28.5 million euros, from Alpha Intelligence Capital (AIC), Elaia Partners, Heartcore Capital and BPI.
Set up by two former Nvidia employees, Brijesh Tripathi and Dali Kilani, it is tackling a major project: the construction of a “universal AI infrastructure” in Europe. “The wait to obtain computing capacity is measured in months. And those who do not plan to spend millions of dollars per month at this level have difficulty obtaining it at all. This causes a real growth crisis for the sector “, explains Dali Kilani, co-founder and CTO of FlexAI, to L’Express.
He and his sidekick know that many companies have tried in recent years to develop their own AI chip so as not to put all their eggs in one basket. “These initiatives have not generated significant revenue. A chip in itself is of no use to anyone, you need the software that goes with it, the cloud, the tool that will make it usable,” explains Brijesh Tripathi, CEO from FlexAI. The start-up is therefore betting that by providing simple and grouped access to these heterogeneous capacities, mayonnaise can take off. It develops a software layer capable of harmonizing everything which acts as “an intermediary between developers and the infrastructures used to carry out their tasks”. A sort of common language that will allow the customer to use what they need without being bothered by technical complexities.
The generative AI pyramid
This product, inspired by cloud platforms, will be launched in the coming months. “Few companies need to have access to a cluster of 1000 GPUs continuously for a year. Having access to computing capacity on demand is necessary in this market,” analyzes Dali Kilani. And as the duo worked for the giants of the sector, they had no difficulty in securing the appropriate partnerships, with AMD, Intel, Amazon, Google and even InstaDeep. “Nvidia is also part of it of course. Not all companies need chips as powerful as those from Nvidia. But it would be foolish not to offer this option in 2024,” points out Brijesh Tripathi.
If their future product keeps its promises, it is good news for European start-ups. Because access to computing power is the Achilles heel of the EU. Generative AI start-ups have multiplied there over the last two years, particularly in France, the United Kingdom and Germany. And Europe has a very strong talent pool in this area. If we except the Dutch ASML, the undisputed champion of high-precision machines for manufacturing chips, Europe is however poorly positioned in what constitutes the base of the generative AI pyramid.
As pointed out by one recent report from France Digitale, the sector is made up of four layers: chips then infrastructure (data centers, etc.) to which are added the foundation models and, finally, AI applications. And if the EU is doing rather well in the last two, in the field of chips, the situation is alarming. The majority of the heavyweights are American: AMD, Intel and of course Nvidia which takes the lion’s share of the graphics processor (GPU) market, which is very effective in training generative AI. And if China suffers from American restrictions in this area, it is working hard to catch up. It also clearly dominates the market for many raw materials necessary for the manufacture of chips (silicon, germanium, gallium, rare earths).
AI Shovel and Pickaxe Vendors
In terms of infrastructure (data center, distribution services, etc.), the situation is not much better. The United States has cloud giants (Amazon, Google, Microsoft). China too (Alibaba, Tencent, etc.). Europe has no player of comparable scale. These factors can slow down the development of European AI start-ups. “And if the founding models of OpenAI or Mistral capture the public’s attention, it is in reality in the chips and the infrastructure that the greatest economic value lies”, points out Marianne Tordeux Bitker, director of public affairs from France Digitale. In this gold rush, it is ultimately the shovel and pickaxe manufacturers who get the biggest share of the pie.
Europe must strengthen itself on this subject. The field is so complex that you must nevertheless approach it intelligently. Opening state-of-the-art factories can cost tens of billions of dollars and often takes years. TSMC, which has extensive experience in the subject, should take three years to open its factory in Arizona. The expert labor needed to operate these sites is also not easy to find. The established players finally have their thousands of patents in the field on their side.
The sweet dreamers who call for public authorities to orchestrate from the heavens the creation of a “European Nvidia” are therefore mistaken on the subject. “Given the speed at which the existing players innovate, there is a great risk in proceeding in this way that a project of this type will produce obsolete chips when they are ready. And that this will cost a huge amount of money. Even if the result is of good quality, it is also necessary to find commercial outlets large enough for the economic equation to hold. However, countries like the United States will undoubtedly favor their local players”, confides a good connoisseur of the French tech ecosystem.
FlexAI’s strategy, which is attacking the market from a completely different front, by placing itself in a currently unexploited niche, is interesting in this regard. For the two founders, in the field of chips and infrastructure, “Europe must position itself on the next wave, rather than trying to fight yesterday’s battle”. For the EU to catch up in this area, it must also invest significantly in it. In France alone, the committee on generative AI co-chaired by the economist Philippe Aghion and the president of the board of directors of the ENS Anne Bouverot, recommends investing more than 7 billion euros in the next five years to accelerate the emergence of a European sector of semiconductor components adapted to AI systems. And for good measure, a billion more to make France and Europe a major center of computing power.
.