AMD's $14B, 15-year lease sets your 2030 token floor

The price you pay for tokens in 2030 is being negotiated right now, in fifteen-year increments. On July 28, Core Scientific announced leases with AMD covering roughly 530 megawatts of US data center capacity, worth more than $14 billion in base contracted revenue over a fifteen-year term. Not a purchase order. Not a cloud commitment. A lease that runs into the 2040s — signed by a chipmaker that is now buying floor space like a hyperscaler.

What actually happened

Per CoinDesk, the agreements cover 529MW across five campuses on 15-year terms, generating more than $14 billion in base contracted revenue. AMD directly leases 377MW across Pecos and Hunt County in Texas plus Muskogee, Oklahoma; an unnamed cloud provider takes 152MW in Auburn, Alabama and Dalton, Georgia with AMD credit support behind it. AMD also holds the right to reserve up to roughly 2GW more, which would push the relationship toward 2.5 gigawatts.

Data Center Knowledge reports construction starts at Pecos with first deliveries expected in the first half of 2027 and deployments continuing through 2028. The deal takes Core Scientific to roughly 1.1GW of leased AI capacity and over $24 billion in potential contracted revenue across all its agreements. The company's COO put fully delivered capacity at about $11–12 million per megawatt — call it $6 billion of infrastructure for the AMD tranche alone. Meanwhile Core Scientific is finishing its exit from bitcoin: it terminated an agreement to buy mining chips from Block and took a $41.9 million charge, with colocation now 83% of second-quarter revenue.

Why 15-year compute leases matter for your business

Two things fall out of this, and neither is "buy AMD stock."

First, the cost floor under your inference bill is being fixed by contracts you'll never see. A fifteen-year lease is a bet that AI demand justifies the power through 2041. When a chipmaker takes that position itself rather than waiting for cloud providers to order chips, it's signing up to sell capacity, not just silicon. That's the structural change here — AMD is moving from component vendor to capacity supplier, which is the only thing that has ever put real pressure on Nvidia's pricing.

Second, that pressure arrives in 2027 at the earliest, and it won't be evenly distributed. Half a gigawatt landing across 2027–2028 doesn't reprice your API bill next quarter. What it does is make the cheapest token in 2029 come from whoever signed the smartest lease in 2026 — and that is genuinely not knowable today. Which is the entire argument for not welding your product to one provider.

So the operator move is unchanged and unglamorous: keep the model layer swappable. Route through an abstraction that lets you point at a different provider without touching feature code. Track cost per completed task, not cost per million tokens — the two diverge fast when a cheaper model needs three attempts. Re-run that number quarterly. When the AMD-backed capacity comes online and someone undercuts your current vendor by 40%, the question shouldn't be "can we migrate," it should be "is the eval green." Everything else is somebody else's fifteen-year bet.

Sources: CoinDesk, Data Center Knowledge.