AMD Isn't Just Buying GPU Capacity From Core Scientific. It's Taking Stock Warrants Tied to How Much You Pay to Rent It.

AMD just signed 15-year agreements for roughly 530MW of data-center capacity across five Core Scientific sites (Pecos and Hunt County in Texas, Dalton in Georgia, Muskogee in Oklahoma, and Auburn in Alabama), with Core Scientific's own management estimating the contracts could generate more than $14B in base revenue over their duration. There's an option to scale as high as 2.5GW. On its own, that's a big-but-familiar AI-infrastructure headline: another chipmaker locking down another few gigawatts.

The detail that's actually interesting is buried a few paragraphs into the coverage: AMD also received warrants to buy Core Scientific stock as part of the deal. AMD isn't just selling silicon into these sites and collecting a check. It now has a direct equity stake in whether Core Scientific's data-center business does well.

What was actually signed

The deployment starts in the first half of 2027 at the Pecos and Auburn sites, with the full 530MW online by the end of 2028. The infrastructure runs AMD Instinct GPUs and EPYC CPUs on AMD's ROCm software stack, and the two companies are jointly designing the physical buildout rather than AMD simply leasing finished space. The warrant component gives AMD the option to acquire Core Scientific shares, which means AMD's own stock performance now has a small but real dependency on Core Scientific successfully filling and monetizing that 530MW, and any further capacity it scales to.

This is a variation on a pattern that's been running through 2026's AI buildout all year: Nvidia has done versions of this with cloud and neocloud partners, trading favorable chip allocation or financing terms for equity stakes or long-term commitments. AMD taking warrants in a company it's simultaneously the anchor tenant for is the same idea, just visible this time in a single, disclosed transaction with real numbers attached.

Why the warrant matters more than the megawatt count

A pure hardware sale gives AMD one incentive: sell more chips, once. A warrant-based stake gives AMD a second, ongoing incentive: Core Scientific's data-center business needs to actually perform, keeping that capacity rented out and pricing holding steady, for the equity position to be worth anything. That's a structurally different relationship than a customer buying GPUs off a price list.

It also means AMD is no longer a neutral supplier when it comes to how GPU-hour pricing evolves downstream. If Core Scientific's utilization or pricing softens, AMD's warrant position is worth less. That doesn't mean AMD is going to manipulate pricing: there's no evidence of that here, and it would likely violate securities law in ways neither company would risk over a warrant position this size relative to their overall balance sheets. But it does mean the incentive alignment between "chip vendor" and "compute landlord" is now explicit and financial, not just strategic.

What this means if you rent GPU capacity through anyone downstream

If you're a solo operator paying for inference or training compute through a cloud provider, a neocloud, or a managed API, you're several layers removed from deals like this one. But the deals like this one are increasingly what sets the floor under what you pay. When the chip vendor has equity upside tied to a data-center operator's revenue, that operator's pricing decisions aren't purely a function of hardware supply and demand anymore. They're also shaped by a financing relationship where one party benefits from utilization and pricing staying strong.

The practical takeaway isn't "GPU prices are rigged." It's that treating compute pricing as a simple commodity market (more chips shipped, prices fall, straightforward supply curve) is increasingly the wrong mental model. A meaningful share of 2026's AI infrastructure is financed through leases, take-or-pay commitments, and now equity stakes between the people who make the chips and the people who operate the data centers. That's a market with financing incentives layered on top of physical supply, and it's worth remembering the next time a provider's pricing doesn't move the way pure chip-supply numbers would predict.

The honest counter

A stock warrant is a small slice of a $14B, 15-year deal. Most of AMD's incentive here is still exactly what it looks like: sell GPUs, collect revenue, build market share against Nvidia in the data-center chip fight. Reading a deep alignment-of-interests conspiracy into what's a fairly standard vendor-financing structure (one that's shown up in other AI infrastructure deals all year without anyone treating it as sinister) probably overstates the case. Companies attach warrants and equity kickers to large, long-duration commercial agreements all the time as a way to share upside without pricing it entirely into the contract terms up front; it's a financing mechanic, not evidence of collusion.

What I'd actually do

Don't panic about GPU pricing over one warrant clause in one deal. Do start treating "who has a financial stake in my compute provider's utilization" as a real input when you evaluate long-term pricing risk for anything compute-intensive you're building. That's the same way you'd look at who owns equity in a landlord before signing a long commercial lease. The chip-vendor-as-landlord-and-shareholder structure isn't going away; AMD and Nvidia both have obvious reasons to keep doing more of it, and the next several gigawatt-scale deals will likely look similar.

Author

Lukas

@lukcombinator

Sources

• Core Scientific ties its AI pivot to AMD in multi-gigawatt infrastructure deal (The Block)
• Core Scientific rises 5% as AMD strikes $14 billion AI infrastructure deal with option to reach 2.5 GW (Yahoo Finance)
• Core Scientific secures AMD as customer at five sites (Data Center Dynamics)
• AMD secures up to 2.5 GW of Core Scientific data centre capacity (Cloud Computing News)