Core Scientific announced Tuesday, alongside second-quarter results, that it has signed 15-year agreements with AMD covering roughly 530 megawatts of AI data center capacity across five sites — more than $14 billion of base contracted revenue, with 2.5% annual escalators baked in. The company says the partnership could ultimately support up to 2.5 gigawatts of leasable capacity.
Read the structure, not the headline number. About 380 megawatts goes to AMD directly, under triple-net leases at Core Scientific’s Pecos, Hunt and Muskogee campuses. The other 150 megawatts, at Auburn and Dalton, is leased to an unnamed neocloud customer under modified gross leases — with AMD providing full credit support for the entire 15-year term. AMD is not just renting space. It is co-signing someone else’s rent.
In exchange, AMD received market-priced warrants on up to 30 million Core Scientific shares. Roughly 6.5 million vested immediately; the rest vest against megawatt-delivery milestones. The first Pecos megawatts are expected in the first half of 2027, about 265 megawatts by the end of 2027, and the balance in 2028. AMD also holds reservation rights on up to 1.9 gigawatts of additional capacity through late 2028.
Our take: This is the Nvidia playbook aimed at a different target. Nvidia has spent a year underwriting its customers — guarantees, equity, backstops — so the chips have somewhere to go. AMD is underwriting its landlord, so the chips have somewhere to sit. Same circular logic, one rung down the stack: a chip company puts its balance sheet behind the power and concrete, and takes equity upside as payment for the risk. The tell is the 150 megawatts AMD guarantees for a customer it won’t name. You don’t co-sign a lease for a tenant you’re confident can pay.
Why a chipmaker is leasing power
AMD finally has a full rack to sell — Helios, MI450-series Instinct GPUs on open Ethernet — and Anthropic has committed to up to 2 gigawatts of it. But a rack order is worthless without energized floor space to put it on, and the 2027–2028 power queue is where AI capacity actually gets rationed. Contracted megawatts are the scarce asset now. Silicon is comparatively easy.
Core Scientific’s side is simpler: it is finishing a full pivot out of bitcoin mining into AI colocation, and it just swapped speculative hash-rate revenue for 15 years of investment-grade-adjacent contracted cash flow. That’s a genuinely better business. It also means the company now depends on one anchor tenant for the bulk of its future book.
The market’s verdict Tuesday was split and instructive. Core Scientific opened higher — up more than 5% pre-market — then gave it back, trading down about 3% as investors weighed the capital intensity and the 2027 start date. AMD fell roughly 5%, dragged along in a chip rout that was itself a referendum on exactly this kind of financing.
What to watch
- The unnamed neocloud. AMD’s credit support on 150 megawatts is the highest-risk line in the deal. Who it is determines whether that’s a formality or a real contingent liability.
- Pecos energization in H1 2027. Everything — warrant vesting, revenue recognition, AMD’s deployment schedule — keys off megawatts actually going live. Interconnect delays are the default outcome in this industry.
- Whether the 1.9 GW option gets exercised. Reservation rights are free. Signed leases are not. The conversion rate tells you what AMD really believes about 2028 demand.
- Lease accounting. Contracted revenue for Core Scientific is a contracted obligation for AMD. Watch how much of that $14 billion shows up as an off-balance-sheet commitment rather than debt.
Bottom line: AMD stopped waiting for the AI buildout to make room for it and bought the room outright. That’s a credible answer to Nvidia’s infrastructure lock-in — and a $14 billion, 15-year bet that demand in 2028 looks like demand today.