AI Compute’s $19 Billion Capacity Promise

A reported two-year compute rental worth as much as $10 billion could turn Meta and Anthropic—frontier-model rivals—into landlord and tenant. Set beside Anthropic’s reported 20-year, roughly $19 billion Kentucky data-center lease, the talks expose a constraint no GPU count captures: whether powered racks arrive when promised.

AI companies have needed more GPUs for years. The shift lies in how they bridge the delay between committing capital and energizing racks. Buyers commit years ahead. Operators secure named customers. Investors finance the equipment, site, power and construction needed to fulfill the contract. The resulting procurement resembles infrastructure offtake more than cloud billing.

Long-term contracts now reach every layer of the stack

In 2023, the infrastructure advantage still looked internal. Omdia expected Meta and Microsoft to receive 150,000 Nvidia H100 GPUs each by the end of that year, three times the expected allocation for Google, Amazon or Oracle. In 2024, Meta detailed two clusters containing 24,576 H100s apiece. The visible contest was inventory accumulation. Companies acquired chips, placed them inside proprietary data centers and kept the resulting capability within their boundaries.

By 2025, OpenAI had reportedly agreed to buy $300 billion of Oracle compute over roughly five years, requiring 4.5 gigawatts of capacity beginning in 2027. Anthropic then signed a 20-year, roughly $19 billion lease for a TeraWulf data center in Kentucky expected to provide about 400 megawatts from the second half of 2027.

Anthropic has also reportedly signed more than 12 initial agreements for direct data-center leases. Google may provide a financial guarantee. OpenAI has carried the same structure downstream through Guaranteed Capacity, offering customers assured access to its compute in exchange for commitments lasting one to three years.

These commitments are larger than cloud credits in both duration and function. A credit lowers the price of usage. A capacity commitment pays for certainty before usage occurs. The buyer secures future availability, while the supplier gains contracted demand against which it can organize infrastructure.

Model labs now sit on both sides of the capacity contract. They reserve physical infrastructure from operators, then sell reserved access to customers. The agreements specify a term, counterparty and delivery date. Those terms give operators demand they can show lenders.

Rivalry matters less when the scarce product is delivery

The reported Meta–Anthropic talks remain unconfirmed. They nevertheless fit a wider pattern in which companies buy capacity from operators that compete with them elsewhere.

Microsoft signed an agreement worth up to $14 billion with Nscale to deploy roughly 104,000 Nvidia GB300 chips in Texas within 18 months. CoreWeave supplies cloud-based Nvidia processors to Meta and Microsoft, while Google has reportedly discussed renting Blackwell chips from CoreWeave. OpenAI and Nvidia have discussed a structure in which OpenAI would lease Nvidia chips instead of buying them.

These deals make corporate boundaries porous at the infrastructure layer. Microsoft can buy from a specialist. Nvidia can act as a lessor. Meta can bargain with a model rival without aligning with it elsewhere.

Each participant solves a different bottleneck. Frontier labs avoid waiting for every site to be built internally. Specialized operators combine capital and chip allocations into working clusters. Hyperscalers can monetize capacity they do not immediately need, while utilities decide whether any cluster can be energized.

Meta and Anthropic can compete over models while finding a compute rental mutually rational. The shared signal is not trust but the value of an available cluster when delivery schedules are binding.

A contract coordinates dependencies that ownership once kept inside one company. The customer bears delivery risk. The operator bears demand risk. The financier judges whether both commitments will survive long enough to service the capital.

Long commitments separate the compute user from the asset owner

A 20-year lease changes the financial character of a data center. Instead of building solely against an internal demand forecast, an owner gains an identified payer and a long revenue schedule. The user gains capacity without necessarily owning the site. Investors can finance the gap between those roles.

Meta has moved $30 billion of AI data-center construction debt off its balance sheet through special-purpose vehicles. It has raised $62 billion in debt since 2022, roughly half of it during 2025. The company also says it will spend an additional $40 billion on its Louisiana campus, which is intended to exceed five gigawatts of compute and push total spending there beyond $250 billion.

An SPV does not erase the economic obligation; it makes the project separable. Different investors can own, lend against or assume risk in the physical asset while Meta retains access to the infrastructure.

US data-center credit deals had reached $178.5 billion during 2025 amid a buildout involving thousands of new entrants. TeraWulf, Nscale and CoreWeave sit between semiconductor suppliers, capital providers and large compute buyers. They bundle chip allocations, construction, power and customer commitments into operating clusters.

Long-term commitments give lenders a revenue schedule, but lenders still have to judge whether the customer can pay for decades, whether the operator can finish the site and whether the power connection will arrive on schedule.

Buyers face the mirror image of that checklist. A procurement team now has to negotiate energization dates, curtailment rules, substitution rights and remedies for delayed delivery—not just an hourly GPU price.

Meta’s Louisiana spending shows that long leases will not replace proprietary capital expenditure. Internal fleets, leases, specialist clouds, project vehicles and external power contracts will coexist.

Power and interconnection determine what a GPU is worth

An accelerator becomes useful only when it arrives with power and an operating data center. A contracted megawatt therefore bundles conditions that a GPU purchase order leaves unresolved.

OpenAI says it has signed contracts for 10 gigawatts of US AI compute capacity, including more than three gigawatts added during the preceding 90 days. Its Oracle capacity is expected to begin in 2027. Anthropic’s Kentucky lease also points to delivery beginning in the second half of 2027. A 2027 commitment cannot satisfy a 2026 workload, regardless of the headline gigawatts.

Google’s two 20-year power-purchase agreements, worth $3 billion for up to 670 megawatts of hydroelectric power, apply the same contracting logic one layer down. The data-center lease secures the site. The power agreement secures the input without which the site is an unusually expensive warehouse.

Grid rules now shape the compute market. FERC has approved orders intended to process data-center power requests within 90 days while imposing new requirements on AI hyperscalers. PJM plans to require large data centers either to bring generation or curtail consumption to reduce the risk of large-scale outages. Local backlash has also produced laws restraining data-center development in Arizona, Georgia and other parts of the United States.

Only 32 countries host AI-specialized data centers, and the United States, China and the European Union control more than half of the leading facilities. Companies can ship a chip; they cannot ship an interconnection. Location enters the price of compute even when the accelerator model is identical.

Chips remain one allocated input inside a larger production system. The same constraint-driven allocation visible in advanced memory now extends through accelerators, power, sites and delivery schedules. A buyer that secures Blackwell GPUs but misses its interconnection date still has no usable cluster.

Benchmarks can emerge before compute becomes fungible

Buyers in a capacity market need ways to observe prices, reserve supply and manage exposure to changing rental costs. Those mechanisms are starting to appear even though each underlying cluster remains highly specific.

CME Group and Silicon Data have announced futures contracts for computing capacity based on daily benchmarks for on-demand GPU rental rates. AWS raised Nvidia GPU prices by 20% for EC2 Capacity Blocks, its advance-reservation product, while leaving Trainium prices unchanged.

Spot benchmarks are moving as well. The Ornn Compute Price Index put one hour of Blackwell GPU rental at $4.08 in April 2026, up 48% from $2.75 two months earlier, citing demand from agentic AI.

These instruments do not make compute a commodity. A daily GPU benchmark cannot erase the difference between an on-demand accelerator and a future 400-megawatt data center, or between capacity in two locations with different delivery dates. Hardware configuration, site and surrounding infrastructure still change the product.

They show something narrower: price uncertainty is large enough to hedge. Buyers committing years ahead and suppliers financing against those commitments both care about capacity prices beyond the current billing period. A forward curve can become useful before the physical market becomes standardized.

Bespoke delivery terms will also keep trading fragmented. Standardized hourly benchmarks can coexist with large, custom agreements, but the basis risk stays with the data center. A Blackwell-hour benchmark cannot hedge a delayed Kentucky interconnection.

Oversupply and lock-in create contracts, not a commodity

The strongest counterargument is that AI capacity is not uniformly scarce. In 2025, TD Cowen reported that Microsoft had walked away from US and European data-center projects representing two gigawatts of planned consumption. The firm interpreted the cancellations and deferrals as evidence of oversupply relative to Microsoft’s demand forecast at the time.

Long construction schedules can produce too much capacity in one place or period while another buyer faces a shortage elsewhere. Contracts do not abolish that cycle. They transfer, divide and price its risk.

Internal optimization can also relieve shortages without external leasing. Amazon created Project Greenland in 2024 to manage GPU allocation across its retail unit and later said it had “ample” GPU capacity. Alibaba Cloud says its pooling system reduced the number of Nvidia H20s needed to serve dozens of large language models by 82%.

Ownership will not give way universally to rental. Companies with enough internal scale can gain substantially by pooling, scheduling and reallocating existing fleets. External contracts become rational when internal optimization cannot meet the required location, configuration, volume or delivery date.

Long commitments can also reduce substitutability for an individual buyer. Ofcom found that AWS and Microsoft made cloud switching difficult for UK businesses. A multi-year capacity agreement can deepen that dependence by tying a workload to one provider’s commercial and technical environment.

The market can broaden even as each participant becomes more locked in. More owners can finance assets, more operators can sell access and more buyers can reserve capacity, yet a specific contracted cluster may remain difficult to replace. Portfolios and new agreements can become liquid before individual workloads do.

Nor does a capacity market require permanently rising prices. Oversupply makes owners more willing to rent unused infrastructure. Shortage makes buyers more willing to sign early. Both conditions produce contracts; they move the clearing price in opposite directions.

The $10 billion question is whether the racks arrive

The reported Meta–Anthropic rental would not reconcile the companies; it would coordinate a delivery obligation for two years. Anthropic’s roughly $19 billion Kentucky lease extends the same obligation across 20 years. Those sums price confidence that chips, capital, construction and power will meet on schedule.

In 2023, advantage meant counting H100s. Now it means making one bankable promise: powered compute, delivered on time.