This week, OpenAI and Oracle shocked the markets with a surprise $300 billion, five-year agreement, part of a surge of new business that sent the cloud provider’s stock skyrocketing. But maybe the markets shouldn’t have been taken by surprise. The deal is a reminder that, despite Oracle’s legacy status, the company still plays a major role in AI infrastructure.
On the OpenAI side, the agreement was more revealing than the lack of details suggest. For one, the startup’s willingness to pay so much for compute provides a measurement of the startup’s appetite — even if it’s unclear where the electricity to power said compute is coming from or how it will pay for it.
Chirag Dekate, a vice president at research firm Gartner, told TechCrunch it’s clear why both sides were interested in this deal. It makes sense for OpenAI to work with several infrastructure providers, he noted. It also diversifies the company’s infrastructure — spreading out risk among several cloud providers — and gives OpenAI a scaling advantage compared to competitors.
“OpenAI seems to be putting together one of the most comprehensive global AI supercomputing foundations for extreme scale, inference scaling where appropriate,” Dekate said. “This is quite unique. This is probably exemplary of what a model ecosystem should look like.”
Some industry watchers expressed surprise that Oracle was involved, citing the company’s diminished role in the AI boom compared to cloud rivals like Google, Microsoft Azure, and AWS. But Dekate argues that observers shouldn’t be so surprised: Oracle has worked with hyperscalers before and provides the infrastructure for TikTok’s sizable U.S. business.
“Over the decades, they actually built core infrastructure capabilities that enabled them to deliver extreme scale and performance as a core part of their cloud infrastructure,” Dekate said.
Payment and power
But even as the stock market celebrates the deal, key details are missing and questions around power and payment remain.
OpenAI has made a string of infrastructure investment announcements over the past year, each one with an eye-popping price tag. OpenAI has committed to spend around $60 billion a year for compute from Oracle and $10 billion to develop custom AI chips with Broadcom.
Meanwhile, OpenAI said in June it hit $10 billion in annual recurring revenue, up from around $5.5 billion last year. That figure includes revenue from the company’s consumer products, ChatGPT business products, and its API. And while its CEO Sam Altman has painted a rosy picture of its future prospects in terms of subscribers, products, and revenue, the company is burning through billions of dollars in cash each year.
Power is another question, or more specifically where the companies plan to source the energy needed to run this level of compute.
Industry observers have been predicting a near-term boost for natural gas, though solar and batteries are arguably better positioned to deliver power sooner and at lower cost in many markets. Tech companies are also betting big on nuclear.
Despite market moving headlines, the energy impact of OpenAI’s anticipated growth isn’t entirely unexpected. Data centers are anticipated to consume 14% of all electricity in the U.S. by 2040, according to a report the Rhodium Group published yesterday.
Compute has always been a constraint for AI companies, so much so that investors have bought thousands of Nvidia chips to ensure their startups have access to the power they need. Andreessen Horowitz has reportedly purchased over 20,000 GPUs, while Nat Friedman and Daniel Gross rented access to a 4,000 GPU cluster (though maybe Meta owns that now).
But compute is worthless without power. To ensure their data centers remain juiced, large tech companies have been snapping up solar farms, buying nuclear power plants, and inking deals with geothermal startups.
So far, OpenAI has been relatively quiet on that front. CEO Sam Altman has placed several prominent bets in the energy sector, including Oklo, Helion, and Exowatt, but the company itself hasn’t thrown money into the space like Google, Meta, or Amazon.
With a 4.5 gigawatt compute deal, that may soon change.
The company may play an indirect role, paying Oracle to handle the physical infrastructure — something it has extensive experience with — just as Altman invested in startups aligned with OpenAI’s future power needs. That will leave the company “asset light,” something that will undoubtedly please its investors and help keep its valuation in line with other software-centric AI startups and not with legacy tech firms, which are burdened with pricy infrastructure.
Source: https://techcrunch.com/