Meta’s 2,600 MW Contract Separates Demand From Delivery

More than 2,600 megawatts for 20 years. In the first quarter of 2026, Meta agreed to buy that much output from three Vistra nuclear plants. A campus plan records what a developer hopes to build; the Vistra contract records what Meta intends to consume.

Across the 16 months beginning in September 2024, Microsoft signed for 835 megawatts over 20 years, Google signed for 615 megawatts over 25 years, and Meta signed for more than 2,600 megawatts over 20 years. Together, the three agreements cover at least 4,050 megawatts. Those named volumes and 20-to-25-year terms let generators, lenders and utilities evaluate demand over decades. Lenders must still assess the builder’s credit and ability to deliver the data center.

Twenty years changes the unit of evidence

Each agreement names the buyer, seller, volume and term. Those fields make three superficially similar AI infrastructure claims comparable.

Microsoft agreed to take all output from a revived Three Mile Island reactor. Google’s agreement includes restarting the Duane Arnold plant, which shut down in 2020. Meta contracted with Vistra while separately backing TerraPower and Oklo reactor projects. The buyers and plants differ, but the contracts cluster between 20 and 25 years.

By mid-2024, owners of roughly one-third of US nuclear plants were discussing data-center supply with technology companies. Talks gave plant owners options; later contracts named the buyers and fixed the terms. As the contracted megawatt argues, generators, lenders and utilities can evaluate physical capacity once a contract fixes its buyer and duration.

Meta put both scale and duration under contract—more than 2,600 megawatts for 20 years.

A power contract and a construction loan answer different questions

Developers building Stargate data centers still need tenant cash flows to support construction debt and other investors to absorb concentrated exposure.

Banks marketed more than $56 billion in investment-grade construction loans tied to Oracle’s future leases. JPMorgan and other banks later struggled to distribute risk from billions of dollars lent against Oracle-leased data centers in Texas and Wisconsin. Oracle’s leases strengthened the financing packages while leaving concentration, construction and distribution risk with lenders.

Meta and Vistra made a different bargain. Meta committed to buying power for 20 years, giving Vistra a named buyer and fixed term. A construction lender still asks whether a developer can finish the site and service its debt.

A developer can struggle to finance a site even when hyperscalers sign long-term power contracts. A bank can refuse more construction exposure without doubting Meta, Microsoft or Google’s appetite for compute. The developer’s cost of capital cannot stand in for the buyer’s willingness to consume power.

Tenants sign leases, lenders fund construction, hyperscalers contract for power and utilities control interconnection rights. The broader AI infrastructure commitment stack places each promise at a different execution gate.

Grid rules force hyperscalers to manage compute as load

PJM Interconnection proposed requiring large data centers to bring their own generation or curtail demand during reliability events. Separately, Google signed demand-response agreements with five US utilities covering as much as one gigawatt and committed to reduce data-center consumption during peak periods.

Microsoft, Google and Meta now manage data centers as both computing assets and grid loads. They procure generation and grid access alongside chips, buildings and cooling, then decide which computing loads can yield when power tightens.

Developers facing grid-access waits of up to seven years turned to aeroderivative turbines and diesel generation. Other projects pursued on-site power plants to bypass overloaded grids, while permitting and supply chains imposed their own constraints. Power, cooling, reliability and electrical distribution now determine how much compute a site can use.

Under PJM’s proposal, a continuously available megawatt carries more value than one subject to emergency curtailment. Google’s demand-response agreements recognize that difference by committing the company to curtail consumption during the grid’s most constrained hours.

A signed megawatt still has a delivery date

Before Three Mile Island can supply Microsoft, Constellation plans to invest $1.6 billion to revive and refurbish the reactor. Google’s Duane Arnold agreement also requires a restart. Meta’s support for TerraPower and Oklo reaches further into new reactor development. Every agreement leaves the supplier with work between signature and generation.

PJM later reduced its summer 2027 peak-demand forecast from about 164 gigawatts to about 160 gigawatts because some projects, including data centers, lacked firm service or construction commitments. PJM made capacity lag visible by cutting four gigawatts from its forecast and excluding projects that had not cleared later execution gates.

US developers put more than 97 gigawatts of gas-fired generation earmarked for data centers into the 2025 pipeline, more than 24 times the 4 gigawatts recorded in 2024. The nuclear agreements show hyperscalers securing long-duration supply; the gas pipeline shows developers reaching for power sooner.

Gas developers face construction and connection risk. Plant owners must fund and refurbish reactor restarts; new-reactor developers must clear development hurdles; data-center builders need financing and firm service. A total that adds every announced megawatt mixes incompatible stages of completion.

Meta’s 2,600 megawatts is a 20-year reservation for power. Vistra must supply it; builders must finance and connect the data centers; PJM must protect the grid. The signature assigns each remaining risk to the party that carries it.