OpenAI's AMD bet: 6 gigawatts and a path to 10% of AMD

For most of the AI boom there was one chip that mattered, and one company that sold it. Then OpenAI, the most demanding buyer in the world, signed up for six gigawatts of someone else's silicon, and took an option to own a tenth of the supplier. When the buyer becomes an owner, something structural has shifted. The question is what it shifts for everyone downstream.

The deal, in numbers

On 6 October 2025, OpenAI and AMD announced an agreement to deploy 6 gigawatts of AMD Instinct GPUs over multiple years, beginning with the MI450 generation and a first 1 GW tranche slated for the second half of 2026 (OpenAI, AMD Investor Relations).

The structure is the eye-catching part. AMD issued OpenAI a warrant for up to 160 million AMD shares at an exercise price of $0.01, vesting in tranches as deployment milestones and AMD share-price targets are hit, the top target reportedly $600 per share. Fully vested, that's roughly a 10% stake in AMD handed to its own customer (TechCrunch, Tom's Hardware).

Why the buyer wanted to crack the monopoly

Think about it from OpenAI's chair. A single supplier of the one component you cannot operate without is the most dangerous dependency in your business. It sets your price, your delivery date, and indirectly, your ceiling. The rational response of any buyer with enough leverage is to fund a credible second source, and then bind that source to you so tightly that its success becomes your success.

That's what the warrant does. It turns a procurement contract into an alliance. OpenAI doesn't just buy AMD chips; it profits if AMD's stock climbs, which it will if AMD becomes a real challenger. The buyer has paid to manufacture its own competition to the incumbent. That is the clearest sign yet that the market wants, and will pay for, a world with more than one chip.

The downstream consequence: optionality becomes an asset

If the frontier is deliberately splitting its compute across NVIDIA and AMD, then the ability to host either, without re-engineering the building, stops being a nice-to-have and becomes a competitive edge. The facilities that win the next decade won't be the ones tuned to one vendor's reference design. They'll be the ones that treat the accelerator as a variable, and the power and cooling as the constant.

This is good news for anyone who isn't a chip company. The thing you actually control, the building, the power contract, the thermal envelope, is also the thing that survives every change of silicon. Bet on the layer that doesn't go obsolete when the next generation ships.