The AI Race Now Has Two Scoreboards

The unit of measure in the AI buildout changed this summer, and almost nobody announced it. For a year the industry's plans arrived in gigawatts: whole campuses, a large power station's entire output pledged to a single company. Since summer, OpenAI and Anthropic have been signing data-center deals at 20 to 30 megawatts, roughly the peak draw of a big hospital, across the United States, Britain and the Nordics, after power shortages and community opposition delayed the giant projects. The data-center builder Crusoe has raised $3.9 billion for a fleet of small, fast facilities built on exactly that template ^[1]. The cause of the shrinkage is shared, and it is physical.

8.9 vs 21.1 GW new data-center capacity added vs. demand, 2025 — Jefferies counts a 12-gigawatt hole between them, with $2 trillion in signed cloud contracts waiting behind it ^[2].

The market's response was to change its unit of score. DeepSeek set the new measure in May, with a permanent 75 percent cut on its flagship V4-Pro that it attributed to cheaper per-request computing on Huawei's newest chips. The same task now runs at a 12th to a 19th of what OpenAI's and Anthropic's top models charge ^[3]. The cascade ran through August: OpenAI cut fees on its Luna model by 80 percent to hold market share; Anthropic shipped a high-performance model at half the price of its top system; xAI and Meta followed with cuts or low-cost options. DeepSeek itself, its floor established, has since raised some prices ^[4]. Anthropic began rationing Claude like a rush-hour utility, with weekly limits burning down fastest from 5 to 11 a.m. on weekdays ^[5]. Inside the companies doing the buying, boards now demand a quantified return, and the buyers cannot produce the number ^[6][7]. The scale of that failure was measured in August, when a study of the 60 largest U.S. financial firms found exactly one that could report a realized dollar return from AI ^[8].

Three years into the AI buildout, the firms buying the technology still cannot put a dollar figure on the payoff. — Milos Maricic

Analysts have a name for the wall, the Utility Wall, and for the scoreboard that has replaced model size: what a task costs ^[9]. Same wall, same months. Washington's answer is where the scoreboards split. Its AI Action Plan, issued in August 2025, directs agencies to remove barriers to adoption, accelerate the technology and use AI to police AI, with the intelligence community embedding it across the investigative cycle, a plan whose only direction of travel is more ^[10][11]. On October 23, 2025, Energy Secretary Alfie Moon directed the grid regulator FERC to speed connections for any new load over 20 megawatts: 60-day review targets, mandatory cost payments from the connecting customer, third-party vetoes over hydropower permits stripped away. One commissioner called the order the key to winning the AI race ^[12]. The president's framing has not moved either. At the All-In Summit he vowed against regulating AI, dismissed the safety debate as a Chinese psychological operation, and cast the contest as winner-take-all ^[1]. Even the campaign for 100 gigawatts of new capacity a year has an author: it was pitched to the White House in an October 2025 memo by OpenAI's Chris Lehane, urging the country to close the "electron gap" with China — 429 gigawatts of new capacity added there in a year against America's 51 — and to let simple AI projects skip the permitting line ^[13]. The blueprint Washington is still building to was written by a company that has since stopped building that way ^[1]. Washington did add one genuinely new tool this year, and it is worth reading clause by clause, because it is not a second scoreboard. In March, President Trump signed a Ratepayer Protection Pledge with Google, Microsoft, Amazon, Meta, OpenAI, Oracle and xAI, committing any data center of 100 megawatts or more to cover the full incremental cost of the generation, transmission and distribution upgrades it triggers ^[14]. The states wrote their own clauses. Oklahoma's legislature, unanimously, requires facilities of 75 megawatts or more to fund their own plants and substations; Wisconsin's regulators ordered Microsoft and Vantage to cover the full cost of the generators and transmission serving their projects; Oregon put a one-cent-per-kilowatt-hour surcharge on mega-data centers ^[15]. On June 18, FERC voted unanimously to speed large connections on one condition: that data centers pay the full cost of the grid upgrades they require. ERCOT, the Texas grid operator, replaced its overwhelmed one-at-a-time queue with batches that favor projects with land and financing already secured ^[16]. In July, the House Energy and Commerce Committee approved, unanimously and with both parties, a bill writing the pledge into law ^[14]. The sponsors put the rulebook's whole logic in a single sentence.

If America wants to lead the world in AI and strengthen our national security, we have to build the energy infrastructure to support it. But we must do that without passing the costs on to working families and small businesses. — Jon Husted

The first half of that sentence is the race. The second half is the invoice. Read together, the pledge, the tariffs, the queue and the bill keep every megawatt of the buildout moving while moving its cost, off the household ratepayer and onto the contract. The rulebook answers who pays for the wall. It never asks what winning means. Nothing in the set slows anything down, and no one in the administration has picked up the market's new measure of what intelligence costs. The questioning that did reach official Washington came from inside the president's own coalition, and by August it had not moved the growth doctrine at the top ^[17]. The bill's critics mostly argued it shifts too little: Frank Pallone said it does not go far enough on pollution and water, and Food & Water Watch said it still leaves loopholes for the utilities. The one proposal in sight that would have slowed construction itself, a moratorium Pallone floated in the same debate, went nowhere. Outside the committee rooms, 142 coordinated protests against data-center expansion had run on July 18 ^[14]. All of it advanced while the public moved the other way: 71 percent of Americans now oppose a data center in their own community, and 60 percent support limits on new construction ^[18]. None of which means the buildout is ending. The five biggest cloud companies will spend $800 billion on capital projects this year and $1.3 trillion in 2027, by Nvidia's own projection — a scale the market had not penciled in before 2028 ^[19]. What changed this summer is the yardstick, not the budget. Both halves of this divergence come with tripwires. If any administration figure starts scoring AI by what intelligence costs, or attaches an actual slowdown to any of these rules about who pays, Washington's side of the pattern breaks. If the labs go back to signing gigawatt campuses, or roll the price cuts back, the market's side does. Through September, neither has tripped: the new deals still run 20 to 30 megawatts, the price floors still hold, and Washington's instruments still count electrons. The next contract signature or committee vote takes another reading.