Rival assault on Nvidia's fortress: first doubts on the market

There is a well-established misunderstanding in the world of semiconductors dedicated to artificial intelligence. A subtle detail, but one that makes all the difference: a difference that has allowed Nvidia to go from 300 billion to $3 trillion in market capitalisation in less than two years, dominating the AI chip market.

The most frequent narrative is that the giant led by Jensen Huang is the only one capable of producing the processors needed to feed the algorithms of OpenAI (ChatGPT) or Google Gemini. But this is only partly true. Because the computing capacity achieved by competitors' chips (from AMD's to Intel's) is absolutely competitive with the famous H100 that made Nvidia's fortune.

What then is the secret weapon of Huang's colossus? Getting into the bowels of AI, the question becomes more complex, but provides more structured answers. Alongside the power of the H100 (and the new Blackwells, when they are ready), there is in fact a software owned by Nvidia, called Cuda, already used in the field of gaming, which today accelerates AI applications, and links machine and algorithms well.

The crux of this story is that Cuda (which Nvidia has been offering to millions of developers for over 15 years) only works with Nvidia infrastructure. In other words: Cuda-H100 has been, and still is, the combination that has made the Californian company's fortune, because it has created a closed ecosystem somewhat reminiscent of the iPhone-iOS model. And this is where the rivals are moving. From AMD to Intel, up to Qualcomm and Broadcom (but also Microsoft, Meta and Google, who want to produce their own chips for AI).

The attack on Nvidia's fortress is moving towards creating an alternative to Cuda that can pave the way for chips from different manufacturers. A big question mark for Nvidia's investors.