"Tax the AI companies to pay for it" is off by a factor of 882

The data-centre excise proposals discussed in Congress this year would raise $17.5bn to $46.5bn over ten years. Take the upper path: $4.65bn a year.

Paying every American $1,000 a month costs $4.1tn a year.

Divide one by the other: 882.

Where the two numbers come from

The $4.65bn is not a back-of-envelope figure. Tax Foundation analysed two paths through the recent congressional proposals this year — a lower scenario raising $17.5bn over a decade and a higher one raising $46.5bn. Both cover removing existing data-centre tax incentives and adding an excise on centres whose workloads are primarily AI compute. ITIF used the same range in its June opposition piece, reaching the opposite conclusion — that long-run economic harm stays below 0.05% of GDP — without arguing the revenue is any larger.

The denominator needs no assumptions either. US resident population is 341.8 million (Census Vintage 2025, as of 1 July 2025). At $1,000 per person per month, that is $4.1tn a year. Total federal revenue in FY2025 was $5.2tn.

So a universal $1,000 a month consumes 79% of federal revenue, and a compute tax covers 0.11% of it.

China’s version: ¥16.9tn, or 77% of revenue

In yuan the shape of the arithmetic barely changes.

China’s population at the end of 2024 was 1,408.28 million. At ¥1,000 per person per month, that is ¥16.9tn a year against general public budget revenue of ¥21.97tn. Seventy-seven percent.

Two countries, two entirely different tax and welfare systems, both landing between 77% and 79% of government revenue. Not a coincidence: “a thousand a month” happens to sit in the same order of magnitude as per-capita fiscal capacity in a middle-income economy.

Another benchmark. China’s domestic VAT raised ¥3,322.3bn in 2024, roughly 2.5% of GDP at a 13% headline rate. Closing a ¥16.9tn gap through VAT would take five times the entire current VAT take.

Only three things can close a gap this size

Rank every possible funding source by the order of magnitude it can raise, and the list of survivors is short.

Consolidating existing transfers. China spent about ¥4.2tn on social security and employment in 2024; US Social Security alone was $1.6tn in FY2025. This is the only existing pool at the right scale. It is also the option with the most direct cost: current benefits are allocated by need, and flattening them into equal payments would likely leave the heavily dependent — disabled people, the chronically ill, large families — worse off. Any plan that closes using this lever has to answer who becomes poorer, first.

Broad-based consumption tax. Doubling China’s VAT burden is worth about 2.5% of GDP. A new 10-point federal VAT in the US is typically costed at 2.5–3% of GDP. This raises real money. It is also regressive — though pairing it with a UBI substantially offsets that, since a flat payment is worth more, relatively, to low earners than the extra tax costs them.

Higher progressive income tax with claw-back from high earners. Mathematically the most effective lever, and the one that most honestly exposes something: if high earners receive the payment and then hand it back in tax, the scheme is already targeted in net terms. “Universal” becomes an administrative form rather than a distributive fact.

A compute tax appears on none of these lines. What it raises is invisible next to any of them.

The people who support basic income should be the ones saying this

The appeal of “make the AI companies pay” is not that it works. It is that it moves the cost onto a specific party who looks like they ought to bear it. The line skips the hardest step in the whole argument: persuading a country’s middle class to accept higher taxes.

The price of skipping that step is that any serious finance official can take the proposal apart in ten minutes. And what comes apart in those ten minutes is not only the compute tax — it is the credibility of the case for basic income generally.

Tax Foundation and ITIF sit in quite different places, one closer to fiscal conservatism and one to industry innovation, and they do not disagree about how much a compute tax raises. They disagree about whether to levy it. When the people arguing against you and the people arguing for you converge on the same number, that number has usually stopped being the thing worth debating.

Alaska’s forty years against AI’s ten

What actually leaves me stuck is not the arithmetic. It is the clock.

Among all the funding designs, a sovereign wealth fund is the only one demonstrated to survive four decades. Alaska’s Permanent Fund Dividend has paid out since 1982 across many changes of government, and Jones and Marinescu showed in the American Economic Journal: Economic Policy in 2022 that it had no negative effect on aggregate employment. It survives because the money comes from an independent fund rather than the annual budget — a politician who wants to cut it faces the entire electorate rather than a small client group.

Ontario launched a basic income pilot in 2017 promising three years. A change of government killed it in 2018, and the complete results will never exist. The difference is not economics. It is institutional design.

The problem is that building a sovereign fund to meaningful scale takes decades, while the gap Stanford Digital Economy Lab has been tracking — employment of 22-to-25-year-olds in highly AI-exposed occupations, 13% below their peers in August 2025 and 19% below in August 2026 — is moving considerably faster than that.

A situation where only the slow design survives, and the problem outruns it, is one I do not currently have an answer to.