Sovereign Compute

Federal Reserve Governor Lisa Cook told an audience in May that companies had announced more than $1.5 trillion of data center plans, and that only a small portion had been realized.¹⁶ That gap, between what's announced and what's energized, is the right way to read every sovereign AI program of the past twenty months. Governments have attached names, land, permits, and in a few cases actual appropriations to AI infrastructure at a scale industrial policy hasn't seen before. The money is mostly private. The construction is mostly ahead of us. And the lever states hold outright, the power grid, has turned out to be the one that decides the pace.

Announced and built

The chart lists state-linked AI infrastructure programs announced or materially updated since January 2025, sized as each government stated them. Amber bars are public or sovereign-fund capital. Grey bars are private capital that a state convened, which is a different thing, and the difference is the point.

Three patterns run through the list. The biggest numbers are private capital that a state convened: Stargate, the Paris pledges, Korea's trillion-dollar drive. The state's contribution is the summit, the permit, and the signature on the press release. Where the money is genuinely public, it's an order of magnitude smaller, and it's spread over years: Japan's ¥10 trillion runs to fiscal 2030, and Korea's ₩150 trillion fund will invest about a fifth of itself in 2026.⁸¹⁰ And deployment is disclosed almost nowhere. Epoch AI had to reconstruct Stargate's site-by-site status from satellite imagery and local permit filings because nobody involved reports it.²

Stargate is the case to study because it's the largest and the best documented. The January 2025 announcement was $500 billion over four years with $100 billion "immediately."¹ As of April 2026, one site is running, Abilene, at about 0.3 gigawatts. Six more sites totalling about 8.7 gigawatts are under construction, and most of them target the fourth quarter of 2028.² No participant has disclosed how much equity has actually been funded against the pledge. None of it is federal money; the federal role is permitting, land, and sponsorship. That's not a criticism of the project. It's a description of what "sovereign" means in this context, and it applies to nearly every row in the table.

The license is a price list

Export control on advanced chips has existed since October 2022. What changed in 2025 is that the United States started pricing access instead of only denying it, and the sequence reads better as a tariff schedule than as a policy debate.

A 15% revenue share became a 25% duty with a volume cap and a testing requirement. The Gulf licenses came with security and reporting conditions attached. The state that controls the supply of frontier compute has decided to sell it, at a price, to the sovereign buyers it chooses, and to charge a toll on sales to the one it doesn't. China's answer, telling its largest firms not to buy, is the other side of the same ledger: a state concluding that access on those terms isn't worth having. If you're modelling chip demand, the practical consequence is that the export line is now a function of two governments' pricing decisions, not one company's supply.

What the central banks count

The institutions whose job is to measure the economy have started measuring this. The Bank for International Settlements' 2026 annual report attributes roughly one percentage point of 2025 US real GDP growth to AI-related investment, notes that the five largest hyperscalers will spend over $1 trillion across 2025 and 2026, and files the episode next to the canal, railway, electrification, and dotcom booms.¹⁵ The Federal Reserve Bank of St. Louis puts AI-related categories at 0.97 percentage points of real GDP growth over the first three quarters of 2025, above the 0.81 points that information technology contributed in 2000 at the peak of the last buildout.¹⁷

The IMF has walked its 2026 global growth forecast through 3.2%, 3.3%, 3.1%, and 3.0% over four releases, and in every one of them AI investment is the main thing holding growth up against tariffs and conflict, while an abrupt rethink of AI productivity is the main thing that could pull it down.²⁷²⁸ The ECB's chief economist put the productivity stakes for the euro area at more than four percentage points over a decade under fast adoption and 1.5 under gradual adoption, with Europe holding about 3% of AI-related patents against 9% in the United States.²⁹ None of these figures is a verdict on whether AI is worth the money. They measure how much of today's output is the buildout itself, and therefore how exposed that output is to the buildout slowing.

The substation

The role states didn't choose is the one that now binds. The Department of Energy's July 2025 reliability assessment says the US grid needs 100 gigawatts of new peak capacity by 2030, half of it for data centers, and that outage hours could rise a hundredfold if scheduled plant retirements go ahead.³⁰ The North American Electric Reliability Corporation documented a single event in which 1,800 megawatts of data center load dropped off the grid at once when uninterruptible power supplies tripped during a routine fault, and it issued a Level 3 alert in May 2026. In July, FERC ordered mandatory reliability standards for computational loads by the end of the year.³¹³² The International Energy Agency expects global data center consumption to roughly double, from about 485 terawatt-hours in 2025 to about 950 by 2030.³³

The response has been to make the state the landlord. The July 2025 executive order defines any data center above 100 megawatts as a qualifying project for expedited federal permitting and federal land.³⁴ The Department of Energy opened four national laboratory sites, and in July 2026 the National Nuclear Security Administration picked Amentum to negotiate a lease at Savannah River for a 1 gigawatt data center with about 2 gigawatts of dedicated on-site generation, gas first and nuclear later.³⁵ The same logic sits under SoftBank's pairing with EDF in France, the Barakah nuclear plant behind the UAE campus, and Korea's 10 gigawatt campus inside an 18 gigawatt national plan.⁷¹¹

This is where the three roles meet. A government can convene private capital, and it can price chip exports, but the buildout still stops at the substation. Whoever owns the land under the substation and writes the reliability standard for the load behind it sets the pace of the whole program, and as of this month that's a small number of governments who've started to act like it. So when the next sovereign AI announcement lands, skip the dollar figure and look for three things: whether any of the money is appropriated, whether there's a megawatt number with a date on it, and who's signed the interconnection agreement. If the third one is missing, the first two don't mean much yet.

@misc{wisniewski2026sovereign,
  author = {Wisniewski, Kyle},
  title  = {Sovereign Compute},
  year   = {2026},
  month  = {sep},
  url    = {https://www.kylewisniewski.com/institutional-ai/analysis/sovereign-compute}
}