Records to 2026-07

Sovereign AI

Announced commitments against government budgets and reported disbursements, by country.

Announced totals exceed what governments will disburse

One screen in plain English: how far announced sovereign-AI money exceeds deployed compute, what the government-only figures are, and how to read a missing figure (an em dash marks a data gap; a true zero prints as $0).

Across 36 national programs, the headline commitments add up to $1.1T. Strip out the private pledges, foreign direct investment, and credit lines that get folded into those announcements, and what governments commit directly (appropriated budgets and state investment vehicles) is $182.0B: at least 17% of the headline is government money, and no more than about 83% is not. Those are bounds. Only the 27 programs that decompose their headline contribute, so the government share can only rise as more of them do.

Of that floor, $21.9B belongs to the 11 programs that report what they have spent, and those have disbursed 23.9% of it. The other $160.1B belongs to 16 programs that publish no disbursement figure, so no execution rate exists for it; 82.9% of that unreported floor is Saudi Arabia and China. Dividing every reported dollar by the whole floor gives 3.1%, a figure that measures disclosure more than spending.

$500.0B of the announced total, 46.8%, is one component the United Arab Emirates record marks unverified: no primary source discloses the amount.

The typical program has disbursed 27.0% of what it announced (the median across programs that report both). Smaller commitments tend to clear at higher rates; the largest announcements show the lowest.

Announced$1.1T

the headline figure, across 27 programs

Government-only$182.0B

what governments commit directly, budgets and state investment vehicles together; no more than about 83% of the headline is private pledges, foreign investment, and credit lines

Disbursed$5.7B

money confirmed spent, across 13 programs that report it

What the numbers do and don’t say

A missing figure is not a zero. Where a figure is not publicly available the record shows an em dash, and the program is never treated as having committed nothing. Several programs (China, Indonesia, Poland, and others) publish no financial figures, and they count as unknown.
Government-only is the metric this rests on. It is decomposed for 27 of 36 programs where the source records permit; the remainder contribute nothing to it, so the government-only total is a floor.
Not all of it is appropriated budget. The floor counts money governments commit directly, which includes state investment vehicles as well as budget lines: Saudi Arabia’s Public Investment Fund and China’s national semiconductor and AI funds are together 72.9% of it, and each record says so in its own classification note. The Funding view keeps those vehicles in a ledger of their own, because an appropriation and an allocation are different kinds of money.
Most headline figures are converted, and most conversions are not yet reproducible. Of the 27 programs with an announced figure, 3 record which currency it was announced in; the other 24 are shown in US dollars with no exchange rate on file, so those conversions cannot be re-derived from the record. Where a rate is recorded it is shown with its source and date.
Disbursed is not deployed. Money confirmed spent is not yet compute in the ground; a facility enters the record only once primary sources place it there.

Continue

The per-country reconciliation, execution tracker, and hardware-dependency reach are on the Programs view; the capital vehicles behind the money on Funding; the anatomy of the gap on Inflation.