The Machine Dividend

The problem

Seven proposals have argued about protecting American workers through this transition: their power bills (№ 024), the measurement of what is happening to them (№ 025), the landing when it happens (№ 026), the right to a person on the other end (№ 027), a way in (№ 028), a public floor under the technology (№ 029), and a supply chain that does not snap (№ 030).

Every one of them is defensive. None of them makes a single American better off if the machine economy succeeds spectacularly.

That is the gap this proposal closes, and it is the one this organisation exists to close. If artificial intelligence delivers what its builders forecast, the resulting wealth will accrue to whoever owns the capital, and the honest question is not whether to slow that down. It is whether ordinary Americans hold a share of it.

Proposal № 001 says they should. The unresolved question has always been the same one: out of what?

Look at what the obvious answers give you. Corporate profits are the most mobile tax base ever devised; a century of transfer pricing says so. A tax on "robots" or "AI" founders on definition — every firm's software becomes not-a-robot within one quarter of the statute passing. Payroll taxes fall on exactly the workers being displaced, which inverts the goal. Data (№ 016) is a real base but a contested and definitionally slippery one.

Now look at compute.

Frontier training compute is physical: it happens in a building with a street address, drawing power off a grid we can measure. It is already metered, to six decimal places, because that is how it is billed. It is concentrated: the five largest US providers have committed $660 to $690 billion in capital expenditure for 2026 alone, roughly three-quarters of it AI-specific, nearly double 2025. And it is hard to disguise, because a gigawatt of load does not hide.

Compute is the most administrable tax base of the machine economy, and nobody is using it.

The proposal

A small levy on large-scale commercial AI compute, deposited as principal in the American Permanent Fund. Not to slow the buildout. To make every citizen a shareholder in it.

How it would work

1. The base. Commercial compute above a high threshold, measured where it is already measured — accelerator-hours or reported training and inference spend, for operators above a floor set well above any research use, university, startup, or public allocation under № 029. Small and mid-scale AI is untouched by design. This is a levy on the frontier, not on the technology.

1. The rate. Low single digits. The number is deliberately unheroic: high enough to compound in a permanent fund, low enough that it does not change a siting decision. If the rate ever becomes large enough to alter where a datacenter is built, it is too high and has stopped doing its job.

1. The destination is locked. Receipts go to the Fund as principal under the two locks of № 023 — never spendable, never borrowed against — and never into general revenue. This is not a revenue measure and must not be allowed to become one, because the moment it funds ordinary spending it becomes a tax people resent rather than a share people own.

1. Credit the public floor. Operators receive a credit against the levy for compute contributed to the Model Commons (№ 029) or committed under the Compute Reserve (№ 030). A firm that helps build the public capability pays less into the Fund, which is the correct trade.

1. Meter it honestly. Reported to the same statistical agency that runs the Automation Ledger (№ 025), audited, and published in aggregate. The grid interconnection data in № 024 is an independent cross-check: a facility's power draw is a floor on its compute, which makes systematic under-reporting visible.

The numbers

Take AI-specific capital expenditure of roughly $450 billion in 2026 among the largest providers, and an operating base that grows as that capital is put to work.

A 2 percent levy on a base of that order yields something like $9 billion a year, rising as the buildout continues.

Deposited as principal and left alone at a 5 percent real return, $9 billion a year compounds to roughly $120 billion of corpus in a decade, throwing off about $6 billion a year, permanently. Per citizen, that is about $18 a year.

Eighteen dollars. We are going to sit with that number rather than dress it up, because this catalog's credibility rests on publishing the unflattering arithmetic.

Three things make it matter anyway. It compounds, and the base grows with the thing it taxes — if AI is a tenth as transformative as claimed, the 2050 figure is not $18. It stacks: spectrum (№ 015), data (№ 016), sovereign equity (№ 020), lapsed war outlays (№ 019), and the grid dividend (№ 024) all feed the same corpus, and a permanent fund is built by accretion rather than by one heroic source. And it establishes the principle while the industry is young, which is the only time such principles are ever established. Nobody has successfully attached a public claim to an industry after it became powerful.

And per № 023, be explicit: while debt exceeds 90 percent of GDP, none of this is paid out. Every dollar it earns goes to the debt. The first actual Machine Dividend cheque is a 2040s event. We would rather say that now than be caught saying it later.

The honest objections

"Training will move offshore. Compute is the most portable thing in this entire supply chain." The strongest objection by a distance, and the one that could sink the proposal. Training runs can in principle be sited anywhere with power and fiber. Four partial answers. The rate is deliberately small relative to the cost differences that actually drive siting — power price, latency, talent, and legal stability dominate a two percent levy. Inference, which is the larger and faster-growing share of compute, is far more location-bound because it sits near users. The credits in item 4 give firms a domestic way to reduce the charge. And the base can attach to compute serving the American market rather than compute located here, which is harder to administer and is the fallback if leakage proves real. If it turns out that a low levy does relocate the frontier, this proposal fails and should be withdrawn rather than raised.

"You are taxing the input to the thing you say will make everyone richer." We are, slightly, and input taxes are genuinely worse than profit taxes on efficiency grounds. The trade is administrability: a profit tax on this sector has been demonstrably avoidable for decades, and a tax that is elegant in theory and collects nothing is worse than a crude one that collects. We would happily trade this for a functioning international profits regime. Nobody is offering one.

"This is a robot tax with better manners." It differs in the way that matters. A robot tax aims to slow substitution and protect jobs by making machines expensive; that is a goal we reject, because it makes the country poorer to preserve the form of work rather than the welfare of workers. This levy is set too low to change any deployment decision, and its purpose is ownership, not deterrence. If it deterred anything, it would be badly designed.

"Compute is a terrible proxy for value. Efficiency gains will collapse the base." A real risk: algorithmic efficiency improves fast, and a base measured in accelerator-hours could shrink even as capability grows. This argues for measuring spend as well as hours, and for reviewing the base on a fixed cycle. It is the objection most likely to require redesigning this proposal within five years, and we would rather flag that than discover it.

"$18 a year is not worth a new tax and a new bureaucracy." If that were the whole return, we would agree. The claim is about a compounding corpus and a principle established early, not about the first year's cheque. Anyone promising that an AI levy will pay your rent is doing the thing we criticise everyone else for.

Sources

• Hyperscaler capital expenditure guidance for 2026: $660–690B across the five largest US cloud and AI infrastructure providers, roughly 75% AI-related, approximately double 2025 levels
• OECD/G20 Base Erosion and Profit Shifting programme, on the mobility of corporate profit as a tax base (oecd.org)
• Proposals № 001 (The American Dividend), № 015, № 016, № 019, № 020 (contributing streams); № 023 (The Debt Covenant, which governs payout); № 024, № 025, № 029, № 030