The Buyer Inside AMD’s 2 GW AI Deal

AMD’s agreement with Anthropic pairs Anthropic’s commitment to deploy up to 2 GW of MI450 systems with an AMD investment of up to $5 billion. The seller is helping fund the customer expected to consume its hardware.

The supplier now helps finance the buyer

Chipmakers once sold into demand that arrived already financed. The customer raised capital or generated cash, placed an order and carried the utilization risk after delivery. Cloud rental shifted ownership to a provider, but the supplier still faced a distinct buyer.

AI labs and cloud intermediaries now reserve large infrastructure blocks before they have proved they can keep them busy. Suppliers increasingly answer that uncertainty with reciprocal commitments: they put capital or rental revenue toward the customer while the customer sends a capacity promise back.

The AMD-Anthropic agreement schedules MI450 deployments from the first half of 2027 and couples them to AMD’s equity investment. Anthropic gains capital, and AMD gains a committed deployment path.

Nvidia routed the same risk through a rental commitment. It agreed to rent 10,000 of its own AI chips from Lambda for $1.3 billion over four years, becoming a customer for hardware bearing its name. AMD uses equity; Nvidia supplies rental revenue.

Suppliers and customers create circular AI financing. The contracts, machines and obligations are real. But when a supplier helps fund the buyer, the resulting order no longer proves that demand arose independently of supply.

A supplier funds a customer; the customer commits capacity; that commitment helps an infrastructure developer finance construction. Suppliers, customers and developers benefit while workloads arrive, utilization holds and refinancing remains available. If any one falters, the customer’s credit quality and the supplier’s order book share the damage.

The rack has replaced the chip as the financed unit

AMD frames the Anthropic deployment around MI450 accelerators installed in Helios rack-scale systems. Nvidia likewise describes liquid-cooled racks containing 256 Vera CPUs, each with 88 custom cores and up to 1.2 TB per second of memory bandwidth. Both vendors now define the product at system scale.

A Helios deployment also needs CPUs, networking, electrical gear, liquid cooling and a building that can support them. Financing the accelerator without those dependencies would resemble purchasing turbines without securing a grid connection: the expensive component exists, but cannot operate.

Nvidia says its 72-chip GB200 system can deliver ten times the performance of H200 servers on specified mixture-of-experts workloads. A buyer may need fewer systems for a fixed workload, or lower computing costs may induce more workloads. The available evidence does not settle which effect dominates.

Capacity contracts therefore finance useful workloads delivered by energized systems over time. Hardware performance, facility availability and utilization meet inside the same agreement.

Credit markets now allocate marginal capacity

AI labs must commit racks, substations and cooling before the applications expected to pay for them have matured. Hyperscalers can bridge part of that gap with internal cash flow. For other buyers, lenders increasingly decide which promised capacity becomes a construction project.

AI infrastructure companies borrowed more than $100 billion in 2025, and projected global AI-tied debt issuance is nearly $570 billion in 2026. At that scale, lenders and bondholders join procurement departments in deciding which long-lived commitments can be built and refinanced.

When suppliers book orders and hyperscalers move projects outside conventional debt accounts, the risk moves with them. At Alphabet, Microsoft, Amazon, Meta and Oracle, estimated off-balance-sheet debt has overtaken balance-sheet debt. Leases and special-purpose vehicles can keep an asset off a corporate balance sheet, but the asset still needs utilization, interest payments and electricity.

Smaller AI infrastructure companies still face elevated rates because lenders distrust unproven businesses. A supplier can help a customer commit demand, but it cannot give every buyer the same cost of capital.

A buyer now has to secure power and cooling, match lease terms to workload commitments and preserve refinancing room before signing for accelerators. Suppliers, lessors and lenders must judge whether that buyer can consume the infrastructure long enough to keep every linked contract sound.

Electricity converts commitments into revenue

A lab can sign for GPUs or racks years before they produce compute. The contract becomes productive capacity only after transmission, generation, permitting and cooling arrive in the same place. Data-center operators have begun building on-site power plants to bypass overloaded grids, but those projects face permitting and supply-chain constraints of their own.

PJM Interconnection exposed that boundary when it proposed requiring large data centers to bring their own generation or curtail consumption to avert large-scale outages. A project that must curtail during grid stress may contain every server in its contract and still produce less usable compute than its financing assumed.

Electricity is the operating permission that converts financed equipment into billable capacity. A building waiting for interconnection may be complete in construction accounts and absent from revenue.

Vertiv sells electrical distribution, power conditioning and cooling at the point where these contracts either become operating capacity or stall. Vertiv has no documented agreement to finance an AI lab, underwrite a data-center project or take customer equity. Its exposure here is physical: every multi-year capacity commitment depends on equipment that can energize and cool the rack.

AMD’s deployment commitment and equity commitment meet at the same chilled rack and substation breaker. Until electricity turns that equipment into billable compute, the buyer inside the sale remains a promise in steel.