Because without a certain type of semiconductor, graphics cards (or GPUs), difficult to develop generative AI models. Only one company currently produces components of this type powerful enough for this technology: Nvidia.
The founder and boss of the American company, Jensen Huang, bet years ago on these processors popular with video game developers, convinced that, sooner or later, the world would need them for AI.
And he won his bet: today, no company can build generative AI systems without relying on Nvidia's GPUs, especially its H100 chip and associated software.
A harsh reality for the competition, from Amazon to Intel to another champion of computers for video games, AMD, who have since sought to catch up. Catching up will take several years.
This situation is also not obvious for small businesses that must be imaginative to gain access to these valuable components.
"It's getting your hands on thousands of GPUs because tech giants are pouring billions of dollars into stockpiling them," said Fangbo Tao, co-founder of Mindverse.AI, a Singaporean startup.
"There aren't many chips available."
The price to pay
ChatGPT hit the screens just as Silicon Valley was emerging from a dark 2022, marked by the backlash of the pandemic, declining investment and waves of layoffs.
The overwhelming success of this new AI has reinvigorated the entire digital sector. Start-ups embarking on generative AI are very well received by investors.
ChatGPT landed on screens just as Silicon Valley was emerging from a black © 2022 Olivier MORIN / AFP/Archives
But the funds raised immediately go into the pockets of Nvidia or cloud services (remote computing).
"We use a lot of large cloud providers (Microsoft, AWS and Google, editor's note) and they tell us that they have trouble sourcing," says Laurent Daudet, CEO of LightOn, a startup that creates language models.
The problem is particularly thorny for those who want to train their own generative AI models, the most computationally intensive creative phase, as it requires processing mountains of data collected online.
Only a few companies therefore have the means to develop large generalist language models.
Microsoft's ten-billion-dollar investment in OpenAI is widely seen as the price to pay to build the necessary servers, powered by Nvidia GPUs of course.
Google is building its own models and on Monday, it was Amazon's turn to announce the payment of four billion dollars to Anthropic AI, another generative AI startup, competitor of OpenAI.
Model training "is almost completely drying up the GPU market," said Said Ouissal, CEO of Zededa, a company that seeks to reduce AI power consumption.
Amazon announces up to $4 billion investment in US artificial intelligence (AI) startup Anthropic © ALAIN JOCARD / AFP/Archives
"Currently, you have to wait until the middle or even the end of next year to be delivered if you place an order. And the shortage doesn't seem to be going to be solved anytime soon," said Wes Cummins, managing director of Applied Digital, which specializes in AI infrastructure.
The players also point out that Nvidia's essential role in the ecosystem places the group, in fact, as a kingmaker for the future of AI.
The market is "almost totally in the hands of the giants of the sector, Google, Amazon, Meta", who have "huge amounts of data and capital" to buy chips from Nvidia and thus develop AI on a large scale, said Jacopo Pantaleoni, former engineer of Nvidia, in an interview with The Information.
"This is not the world I wanted to build," he added.
In the opinion of some old Silicon Valley connoisseurs, the frenzy around Nvidia's GPUs will not last forever and other options will eventually appear.
Without this, the cost of entering the AI market risks becoming prohibitively expensive, even for the digital giants.
© 2023 AFP