Anthropic's $3.4B Generator Commitment Before the Grid

Microsoft’s Satya Nadella warned that chips could sit in inventory because the company lacked power to plug them in. Anthropic reportedly committed $3.4 billion for gas-fired generators. The industry can secure the processors and still have to shop for the means to turn them on.

AI data centers were designed around a utility connection that supplies ordinary operations and generators that protect against interruption. Developers are reversing that sequence: when grid delivery sets the campus schedule, they reserve generating equipment as part of deployable compute capacity, taking on power-project execution risk that a backup plan was never meant to carry. The purchases establish a change in procurement; they do not, by themselves, establish that particular generators have replaced the grid in sustained operation.

The utility connection no longer reliably arrives before the servers

A data center is a computing system made of power distribution, conditioning, cooling, hardware and software. Barroso, Hölzle and Ranganathan’s The Datacenter as a Computer describes availability as a property of the whole system, not a promise made by any one component. A server rack without an energized electrical path is installed capacity, not delivered compute.

For years, the operating assumption made the utility connection the starting point. Developers could design the building and its backup equipment around grid service. That assumption weakens when the utility’s delivery date falls behind the construction and hardware schedules. US developers facing grid-access waits of up to seven years have turned to aeroderivative turbines and diesel generators. By October 2025, developers were also building on-site power plants to bypass overloaded grids, even as those projects encountered their own permitting and supply-chain constraints.

Chips and grid interconnections follow different schedules, and the later arrival determines when a campus can use both. Nadella’s inventory warning gave that mismatch a warehouse address.

A generator reservation is different from a backup order

A standby generator is bought to cover an exception. A forward commitment for generating capacity addresses an anticipated condition of operation. Sources told the Financial Times that Anthropic bought commitments for 1.7 gigawatts of gas-fired generators from OnePWR for $3.4 billion, describing it as the largest known purchase of its kind by an AI lab. The commitment reserves equipment; the reporting does not show that 1.7 gigawatts of generation has been installed, permitted or dispatched.

Other developers are seeking shorter routes from an equipment order to electricity. Operators and power developers have been repurposing jet engines to obtain power in as little as two years rather than wait through a multiyear grid-connection process. That comparison explains the appeal of on-site generation without making two years a guaranteed completion time for any campus.

Global Energy Monitor counted more than 97 gigawatts of US gas-fired capacity explicitly earmarked for data centers in the 2025 development pipeline, compared with 4 gigawatts in 2024. Pipeline capacity is proposed capacity, not electricity already flowing to racks. Its value as evidence is narrower and still substantial: developers had begun identifying generation for data centers on a scale that the old standby assumption does not explain.

The engine maker inherits a place in the compute schedule

The Wall Street Journal reported that Caterpillar grew generator sales 22% to $7.8 billion in 2024 and another 28% year over year through September 2025, attributing the demand to AI data centers. Caterpillar’s sales do not establish that it supplies Anthropic’s project. They show why a company that once appeared mainly in the facilities budget now matters to the timetable for usable computing capacity.

Cummins, Eaton and other industrial suppliers have also pursued the data-center power-equipment market. Their positions differ: generating electricity, distributing it and preparing a campus to receive it are separate jobs. Transformer lead times that have stretched from months into years make the distinction consequential. A developer can secure an engine and still lack the equipment needed to deliver its output safely to the racks. Installation, controls, servicing and fuel logistics belong in the same project plan, even where the public reporting does not identify who has contracted to provide each one.

Generac’s reported agreement to supply Amazon with up to $8 billion of backup generators supplies a useful boundary. A large equipment commitment is not proof of primary generation. The buyer’s intended duty cycle, the site’s interconnection status and the plant’s actual run-hours decide what role the machines play.

Lenders are being asked to finance the building and its power

Compute sellers already contract ahead of delivery. OpenAI introduced Guaranteed Capacity with one- to three-year customer commitments, while lenders marketed more than $56 billion in construction loans tied to Oracle’s future data-center leases. Those arrangements make the completion date financially important: a lease or compute commitment needs a working campus behind it.

Power assets are entering that same underwriting conversation. Lenders considering Meta’s Project Walleye in Ohio were asked to fund both a data-center building and its power assets in a proposed $3 billion loan. In Louisiana, Meta agreed to fund energy infrastructure for its data center, including seven natural-gas power plants. These are different arrangements—one brings power assets into a construction financing package; the other funds utility infrastructure—but each places electricity supply inside the economics of the campus rather than leaving it as a bill paid after opening. Lenders in the broader data-center project-finance market must underwrite the link between a future compute customer and a powered physical site.

On-site gas is not the only way to secure the supply. Long-term nuclear power contracts can support grid-facing generation without requiring a developer to assemble a plant beside every campus. The choice is between different construction schedules, delivery arrangements and parties responsible for making the electricity available.

The grid still has answers that a gas plant cannot provide

A UK trial involving National Grid and Nvidia found that AI data centers could adjust consumption when asked rather than draw peak power continuously. If operators can shift some work away from constrained hours, a utility may be able to serve load it could not support as an inflexible peak demand. The trial does not establish how much flexibility every AI workload can offer, but it challenges the assumption that each proposed campus requires generation sized for uninterrupted peak use.

PJM made a different correction to expectations. In January 2026, the grid operator reduced its forecast for summer 2027 peak demand to about 160 gigawatts from about 164 gigawatts because some projects, including data centers, lacked firm service or construction commitments. PJM revised a forecast, not delivered electricity. Its decision is a warning against treating every announced campus—or every proposed power plant—as inevitable demand.

Those limits do not erase the wait confronting a developer with a site and hardware ready before grid service. They do keep the roles of bridge supply, permanent on-site supply and flexible grid demand distinct. A generator ordered to cover an interconnection delay may have a different useful life from one contracted as a campus’s continuing source of power.

Operating power brings operating liabilities

A backup machine can spend most of its life waiting. A machine expected to supply ordinary compute operations needs fuel, maintenance, dispatch decisions and permission to run. Amazon’s reported backing of a 7.65-gigawatt gas plant for an off-grid Texas AI data center puts the emissions question alongside its 2040 net-zero goal; project coverage describes the proposed plant as potentially the largest single source of emissions in the United States. Permitting and supply-chain problems already affecting on-site plants can delay the supposed shortcut around the grid.

Public reporting has not established standard terms assigning fuel prices, maintenance, emissions, permitting or availability liabilities among AI labs, power developers, equipment vendors, utilities and lenders. Nor does it document two named AI campuses with generation run-hours, fuel arrangements and operation ahead of grid interconnection. OnePWR’s reported equipment commitment therefore cannot be read as a record of how Anthropic’s generators operate. That missing operational evidence is the line between a procurement reversal and a completed replacement of utility service.

The old campus drawing put the grid connection before the server racks and the generator beside them, ready for an outage. At a campus waiting for the grid, the generator order can come first; whether the racks light up still depends on a working plant, not the order.