AI’s Power Race: How to Tell a Project from an Announcement

A large AI power deal can describe several different realities: an equipment order, a future connection, a contract for electricity, or a site already doing useful computing. Those are different stages. Treating them as interchangeable makes it harder to understand who is getting ahead—and who may carry the cost.

The Financial Times reported on November 8 that Anthropic had made roughly $3.4 billion in generator commitments covering 1.7 gigawatts of heavy-duty gas equipment from OnePWR Solutions. Reuters relayed that reporting and said it could not independently verify it. That is one reporting chain, not two independent confirmations. The report alone does not establish that all the equipment has arrived, received permissions or begun supplying usable power. Reuters reporting via StreetInsider

For the wider AI race, the useful question is what happens between a commitment and a working facility. Three checks help: identify the route to usable power, find the enforceable responsibilities, and look for evidence of delivery.

1. Follow the route to usable power

Imagine a completed server hall with racks and chips in place. Outside, the feeder connection remains unfinished. The building can look ready while its useful computing capacity remains constrained. An attractive photograph of the hall does not answer the electrical question.

Computing is physical. Servers draw electricity and release heat; cooling and other facility systems also consume electricity. A capacity figure for a generator is therefore different from the electricity available to the computing equipment. Power describes a rate; energy describes an amount used or produced over time. Operating hours, outages, reserves and facility loads all matter. U.S. Energy Information Administration

The global numbers provide context rather than a site-level answer. The International Energy Agency’s April 2026 update estimated that all data centres used about 485 terawatt-hours in 2025. Its central projection rises to approximately 950 terawatt-hours in 2030. Those figures include data centres beyond AI, and the later number is a forecast. They do not tell us whether one campus can receive dependable electricity next year. IEA, executive summary

Two construction clocks can run at once. The computing project needs buildings, servers and cooling. Its supply project may need generation, transmission, a substation, studies and approvals. Completing the first clock does not complete the second.

Berkeley Lab’s queue analysis illustrates the distinction. In the regions covered by its measurements, generation and storage projects completed during 2025 had spent a median of more than five years between requesting an interconnection and reaching commercial operation. This is a statistic about that completed supply-project cohort. It is not a five-year waiting-time estimate for every AI data centre. Berkeley Lab, Queued Up

Connection also does not end the operating challenge. PJM has described how persistent large-load demand can complicate maintenance seasons and why coordination matters. A facility needs a workable supply arrangement after its opening ceremony, too. PJM, October technical-conference account

For a specific announcement, ask what feeds the site, which unfinished dependency controls its opening date, and how the supply arrangement handles planned maintenance or failures. A grid connection, a dedicated plant and a combination each deserve their own examination.

Supply choices carry different dependencies

Onsite generation can place production closer to servers and reduce reliance on an unfinished external connection. It also adds a power project to the development. Gas equipment needs fuel arrangements, installation, maintenance and permission to operate. An order does not settle those requirements.

Nuclear supply needs a different distinction: an existing operating plant, a possible restart and a proposed new reactor are not equivalent opportunities. A long-term agreement is not evidence that a new reactor is operating. Wind and solar add other considerations because their production varies with conditions. Matching an annual total is different from meeting demand in every hour. IEA, energy supply for AI

Batteries can move energy across time, but they must first be charged and have finite stored energy. Other generation, grid connections and flexible demand may contribute to the same plan. Geothermal offers another possibility for steady supply while retaining drilling and development challenges of its own. U.S. Department of Energy, geothermal and data centres

No single technology removes every question about location, timing, reliability and cost. A credible plan explains how its parts work together at the actual site.

2. Find the enforceable responsibilities

People living nearby may ask a more immediate question: what changes here, and who pays? New demand can require costly infrastructure, but a construction announcement does not by itself determine how the costs will be allocated. Utility arrangements, regulatory decisions and developer obligations matter. Local land, noise, water and emissions effects also depend on the facility’s actual design and operating systems. A global demand forecast cannot resolve those questions for a neighbourhood. Congressional Research Service, September 2026 analysis

In February, Anthropic said it would cover grid upgrades and electricity-price effects associated with its data centres. Its statement distinguishes directly developed projects from capacity leased at existing facilities. The relevant question is what the commitment covers and how it becomes an enforceable arrangement. The statement is not verification that every future effect on other customers’ bills has already been covered. Anthropic’s February 11 statement

Useful questions include who funds the connection, who bears the risk if the projected load never arrives, and what protects other customers. New supply may bring benefits as well as costs. Neither outcome should be assumed solely from the project’s announcement.

The policy process is developing. In June, FERC opened a proceeding involving PJM large-load service and cost-shifting. PJM later described work on coordination and flexible computing demand. These are dated developments, not a universal permission rule. FERC, EL26-67-000, PJM, October 22

Some computing work may move away from peak hours. That possibility does not establish that every training run or user request can be interrupted without consequences. The workload and the service obligation matter.

3. Look for delivery after the announcement

Power constraints could influence competition in several overlapping ways. Companies with financing, powered sites and early equipment commitments could gain earlier access to capacity. That would make infrastructure part of the competitive advantage without proving that the largest plant produces the best model.

Efficiency could change the equation. Better hardware, software and workload management may produce more useful work from the same electricity budget. The evidence should compare performance at comparable quality. Lower task costs may also encourage more use, so efficiency alone does not guarantee falling total demand.

Growth could also change shape: projects might open in stages, relocate or wait for dependencies. Higher costs could change which workloads are worth running. These are possibilities, not forecasts with assigned probabilities.

Watch for sites being commissioned, customers receiving usable capacity, and comparable work taking less energy. Also watch whether completion dates move. Those observations can help distinguish emerging outcomes from a stack of announced gigawatts.

The evidence does not establish an imminent worldwide electricity shortage for AI. It does explain why access to usable power can affect where AI grows, how quickly it grows and who bears the cost. Keep the three checks close: route, responsibility and delivery.

Media and attribution: Jeremy appears as an authorized AI-generated presenter using the channel’s approved identity and voice. Infrastructure scenes are generic synthetic conceptual imagery, not footage of Anthropic, OnePWR or verified facilities. Explanatory animations are not engineering designs or measured site telemetry. Original Run #9 music is combined with licensed Mixkit door/latch Foley and outdoor ambience; underlying stock recordings are retained internally and are not offered for redistribution. IEA material is credited under CC BY 4.0; the recreated Berkeley chart/statistic credits Lawrence Berkeley National Laboratory and GridTracker, Queued Up: 2026 Edition, under CC BY 4.0. No source article photographs or third-party music recordings are reproduced.