Stargate is nowhere near enough to make OpenAI's AMD and Nvidia tie-ups work

Comment AMD has issued OpenAI a warrant for roughly 10 percent of its stock. In exchange, the AI model giant will work with its partners (such as Oracle) to deploy up to 6 gigawatts' worth of AMD GPUs.

But while Wall Street rejoices and AMD's share price rallies, OpenAI faces a very real problem. To make good on either the AMD deal or a similar one with Nvidia, it will need to expand its Stargate datacenter initiative massively. 

Even then, AMD and Nvidia are still getting the better end of the deal here.

Nvidia's agreement to pony up to $100 billion for 10 gigawatts with OpenAI is really just the GPU giant scratching its own back. Nvidia sells GPUs and networking kit at a sizable margin to the likes of Oracle, who in turn rent them to OpenAI at a profit. And because those payouts are tied to specific capacity milestones, Nvidia and its datacenter and cloud partners are guaranteed a return on every dollar invested in OpenAI. It's a bit like giving out a "buy four get one free" coupon.

AMD's tie-up with OpenAI is similar in many respects, but House of Zen lacks Nvidia's war chest. So instead, CEO Lisa Su is essentially making a trade. AMD has granted OpenAI a warrant to purchase up to 160 million shares of common stock at just a penny apiece in exchange for driving up AMD Instinct's share of the GPU market.

Assuming OpenAI hits its milestones and AMD's share price grows as it anticipates, OpenAI will be able to use its equity in the company as collateral to solicit new debt financing, and buy time while it figures out how to extract enough cash from AI developments to justify the investment. And considering that Altman doesn't expect to turn a profit until at least 2029 at the earliest, the company is going to need all the financial help it can get.

Not enough datacenter capacity

While six gigawatts in exchange for the option to buy 160 million shares of AMD common stock certainly got the market's attention — AMD's stock surged by as much as 35 percent on the news — the figures represent an upper bound.

Just like OpenAI's partnership with Nvidia, the deal is structured in such a way that the AI flag bearer can only exercise its options on the first tranche of AMD stock after its partners have deployed a gigawatt of Instinct GPUs. Even then, OpenAI is under no obligation to execute on its warrant — though they'd be crazy not to.

The bigger problem is OpenAI doesn't actually have the datacenter capacity necessary to see either deal through. As of October, OpenAI has only received commitments for about seven gigawatts of Stargate datacenter capacity, and from what we understand, Crusoe and Oracle have only managed to bring about 200 megawatts online at the company's flagship Stargate datacenter in Abilene, Texas.

Stargate's addressable compute capacity is expected to grow significantly over the next year. Crusoe is slated to bring six new buildings online by mid-2026, bringing its Texas campus to about 1.2 GW total capacity.

Along with OpenAI's Abilene centers, Oracle, which also happens to be one of the few companies with experience deploying large quantities of AMD Instinct hardware, has committed to overseeing about 5.7 gigawatts of compute over the next four years.

• OpenAI IP promises ring hollow to Sora losers
• OpenAI and AMD link arms for AI buildout: It's a power-for-equity swap
• Big money is nervous about AI hype, but not ready to call it a bubble
• Microsoft CTO says he wants to swap most AMD and Nvidia GPUs for homemade chips

What does AMD get out of the deal?

AMD expects to begin delivering the first gigawatt of capacity in the second half of next year. So what exactly would that mean for the chip biz? The answer to that depends on a number of factors, notably system power and pricing. At this point, AMD has only teased its MI400-series Helios rack systems and we don't yet have good information on either data point.

But, assuming peak system power consumption comes in somewhere in the neighborhood of 250 kW, we're looking at around 3,200 double-wide rack systems or around 230,000 GPUs per gigawatt of datacenter capacity.

Even a single gigawatt-scale deployment of MI400 would have a meaningful impact on AMD's GPU market share. If we assume the House of Zen can sell its Helios rack systems for at least $3.5 million apiece — roughly in line with what Nvidia is charging for its own NVL72 rack systems today — it's looking at at least $11.2 billion in revenues.

For reference, AMD sold a little over $5 billion worth of Instinct GPUs in the entirety of 2024. ®