Training AI models demands chips that have high amounts of memory, said Greg Osuri, the founder of Akash Networks, a marketplace for companies to sell access to their GPUs to other firms that need it. “Nvidia happens to be the only company that makes those chips.”

For months, start-ups trying to get into the AI race have competed with big tech companies for Nvidia’s GPUs. At the Newcomer AI conference in March, David Luan, co-founder of AI start-up Adept Labs and a former Google and OpenAI engineer, said once an AI company figures out its business model, they need venture capital money so they can begin “sending checks to Jensen”, referring to Jensen Huang, Nvidia’s CEO.

A set of eight of the most advanced chips can cost $US300,000, Osuri said. Companies buy thousands of them. Twitter owner Elon Musk recently bought about 10,000 GPUs as part of his plans to start a new AI company of his own, according to a report in Insider.

The launch of OpenAI’s ChatGPT, a chatbot that could have complex conversations, pass professional licensing exams and write computer code, in November helped show the promise of the technology, essentially serving as a firing pistol for a new AI arms race.

In a conference call with investors on Wednesday, Nvidia’s chief financial officer, Colette Kress, called ChatGPT’s launch the technology’s “iPhone moment”, marking the point at which the world realised the potential for the new tech.

Loading

“The technology came together and helped everybody realise what an amazing product it can be and what capabilities it can have,” Kress said.

Nvidia’s stock had already more than doubled this year as the AI boom took off, but the company blew past already-high expectations on Wednesday when it forecast that sales in the second quarter would be $US11 billion, compared with the $US7 billion that Wall Street analysts had forecast.

The numbers “blew everyone away”, said CJ Muse, an analyst with Evercore. “No one saw that.”

The dependency on Nvidia is so widespread that big tech companies have been working on developing their own competing chips, much in the same way as Apple spent years developing its own chips so it could avoid having to rely on – and pay – other companies to outfit its devices. Google has built its own “Tensor Processing Units” for several years, and Microsoft and Amazon have programs to design their own as well.

Even so, those alternatives won’t be enough for the biggest companies, Muse said. “I think Nvidia dominance will continue.”

Washington Post

The Business Briefing newsletter delivers major stories, exclusive coverage and expert opinion. Sign up to get it every weekday morning.