The Intelligence Dividend
A proposal for a permanent public royalty on artificial intelligence
Version 1.0 — July 2026. A living document: the strongest objections received will be incorporated into future versions, with credit.
The Bet
In July 2026, Neil Rimer — co-founder of Index Ventures, one of the most successful venture firms of the last three decades, and an investor in Anthropic — told TechCrunch that a redistribution of AI wealth is coming. “It’ll either be voluntary or it’ll be involuntary,” he said, adding that he hopes the industry’s leaders choose the voluntary path before history chooses for them.
Rimer named the fork. He did not name the mechanism. This document is the mechanism — the voluntary path, with an instrument attached.
The short version: Every large AI datacenter pays a few dollars per megawatt-hour into a permanent public fund — about half a cent per GPU-hour, too small for any business model to notice. If AI profits ever reach the heights their own builders predict, the winners pay a share of the windfall into the same fund. The fund pays every citizen an equal dividend, forever. If you use AI, you get a cut. If AI takes your job, you get a cut. Either way, you hold equity in the machine. The rest of this document is the engineering.
The leaders of the frontier AI laboratories — Sam Altman at OpenAI, Dario Amodei at Anthropic, Demis Hassabis at Google DeepMind — describe a future in which artificial intelligence performs most cognitive work, compresses decades of scientific progress into years, and generates economic surpluses measured in trillions of dollars. These are not idle claims. Capital is acting on them: hundreds of billions of dollars in datacenter construction, gigawatt-scale campuses, national industrial strategies reorganized around compute.
This proposal takes those predictions at face value and asks a single question: if they come true, does humanity hold equity in the outcome?
Today the answer is no. The Intelligence Dividend is the mechanism for making it yes — negotiated now, while a permanent stake costs almost nothing, rather than later, when the AI economy’s cap table is final and every claim on its surplus has already been allocated.
Where the Royalty Is Placed
Not on API calls. Not on tokens. Not on any attempt to measure “value created.” The royalty attaches in two places — one physical, one financial.
Layer 1 — the meter. A small per-unit royalty on the physical substrate of artificial intelligence: a few dollars per megawatt-hour of facility load, collected at compute facilities above a scale threshold (illustratively, 25 megawatts) whose primary equipment is AI accelerators — the specialized chips that run AI. It starts immediately.
This placement is chosen because it is the only base that survives the industry’s own predicted future. A datacenter is the least mobile asset in the modern economy: it sits behind multi-year grid interconnection queues, gigawatt substations, and export-controlled chips. Intellectual property fits in a Dublin mailbox; a gigawatt does not. The meter already exists — the utility installed it. The chips are already classified and tracked. And because the levy sits beneath every model rather than inside any company’s billing system, it is universal: an open-weight model forked to a private cluster still runs on accelerators drawing power somewhere. Closed or open, domestic or foreign, incumbent or challenger — anyone operating at scale pays the same trivial rate, and everyone below the threshold pays nothing. A gaming PC, a university lab, a startup’s rack never sees it.
Two rules keep the base honest. “Accelerator facility” is defined by chip inventory, not self-declaration — the accelerators in question are already export-classified, so the inventory exists. And affiliated facilities aggregate toward the threshold, the same anti-splitting rule every oil-and-gas royalty regime uses, so no operator escapes by building 24.9-megawatt cells.
The rate is set to be immaterial by design. At two dollars per megawatt-hour, the royalty on a single high-end accelerator running for an hour is roughly two-tenths of a cent — well under one-tenth of one percent of the cost of that compute. No business case changes. No deployment decision moves.
Layer 2 — the windfall. A binding commitment, signed now: a graduated share of a firm’s total profits above thresholds denominated in gross world product — the combined output of the entire world economy, roughly $110 trillion today. Illustratively: one percent of annual profits above 0.25 percent of GWP (about $275 billion today), ten percent above 0.5 percent of GWP (about $550 billion), and twenty-five percent above one percent of GWP (about $1.1 trillion). Below the first line, nothing is owed — ever.
No company on earth is near that first line. The most profitable firms in history earn roughly $100 billion a year — barely a third of the entry threshold. Crossing it would itself be the evidence that transformative AI has arrived. And because the schedule reads total profits, it never has to answer the unanswerable question of which profits are “AI profits.” The apportionment game that broke the corporate income tax is designed out, not litigated.
Layer 2 does not need to wait for legislation. It can begin life exactly as the Windfall Clause — proposed by Cullen O’Keefe and colleagues at the Centre for the Governance of AI in 2020 — was designed to: a voluntary, binding commitment signed by firms and their investors. Commitments are cheap before windfalls and impossible after. This is Rimer’s “easy way,” with an instrument.
In one sentence: the royalty sits on the watt today and on the windfall tomorrow.
Who Pays — A Worked Example
The check is written by the operator of record: whoever holds the utility account for the facility. Not the landlord who owns the building. Not the lab renting the compute. Not the person typing the prompt. One entity remits, and the cost flows downstream in slivers — the way a fuel tax is remitted by the gas station and paid, invisibly, by every driver. The design never has to understand anyone’s org chart. Exactly one name is on the meter.
Take the most famous AI campus on earth: Stargate Abilene, in Texas. Its ownership stack shows why this matters. A consortium of Crusoe, Blue Owl Capital, and Primary Digital Infrastructure financed and owns the buildings, on land leased from Lancium. Crusoe built and operates the campus. Oracle holds a fifteen-year lease and supplies the chips — a reported 400,000 of Nvidia’s most advanced GPUs at full build. OpenAI rents the compute from Oracle. The output flows to ChatGPT and API customers everywhere.
Under this proposal, the campus operator remits. The landlord consortium files nothing. Oracle files nothing. OpenAI files nothing. The user files nothing.
Now the size of the check. At full build — roughly 1.2 gigawatts of capacity across eight buildings — the campus draws about nine to ten terawatt-hours a year. At two dollars per megawatt-hour, the royalty comes to eighteen to twenty million dollars a year. Set that against tens of billions of dollars of chips, an electricity bill around half a billion dollars a year, and compute contracts reported in the tens of billions annually: the royalty is about four percent of the power bill, hundredths of a percent of the compute being sold, roughly half a cent per GPU-hour. No spreadsheet in Abilene changes.
And here is why the meter — not the tenant — is the right anchor. In the two years since ground broke, nearly every logo above that meter has changed. xAI considered the site and walked away to build its own. Oracle and OpenAI shelved their planned expansion of the campus amid financing pressures and shifting demand. Microsoft stepped in to lease remaining capacity. Nvidia put down a deposit while brokering talks to bring Meta onto Crusoe’s books. Through every reshuffle, the watts never left Abilene. Tenants churn; meters don’t. A royalty attached to any one company’s API would have chased four different billing systems by now. The one attached to the meter never moved.
From there, ordinary economics does the distributing. The operator folds the royalty into the lease, the leaseholder folds it into compute rates, the labs fold it into token prices, and the final cost reaches end users at a few cents per hundred dollars of AI spend. The founding intuition survives intact — every use of frontier intelligence contributes a microscopic amount to humanity’s future — but the collection moves from millions of self-reported billing systems to one utility meter that cannot lie.
Where one company owns the whole stack — xAI’s Colossus in Memphis, where a single firm is landlord, operator, and lab at once — it simply remits for itself. The design is indifferent to corporate structure. It follows the watt.
Why There — and Not on Usage or Value
Three reasons.
Value cannot be measured, so the design refuses to measure it. Most of what AI creates arrives as consumer surplus — the conversation that changes a decision, the diagnosis reached a year early, the product a small business could never have shipped. None of it appears in any ledger. Economists have documented how badly national accounts miss free digital goods; Erik Brynjolfsson’s GDP-B research program is the standard reference. Any royalty base that requires valuing AI’s output collapses into permanent arbitration. This design reads only numbers that already exist: a utility meter and an audited income statement.
Usage deflates and disperses. The price of a unit of intelligence at constant capability has been falling by roughly an order of magnitude per year, and the industry’s stated destination is Sam Altman’s phrase — intelligence “too cheap to meter.” A per-call or per-token fee is therefore self-liquidating. Hold it fixed and it becomes a prohibitive multiple of a collapsing price, driving traffic to anything unmetered; scale it with the price and its revenue collapses alongside. Meanwhile, inference is dispersing — into open weights, onto devices, into self-hosted enterprise stacks — and away from any small set of metered APIs. Ten trillion API calls a year at a tenth of a cent yields ten billion dollars; the surplus this proposal addresses is measured in trillions. Per-unit fees on transactions can never reach value-scale money. Only one quantity grows monotonically in every version of the predicted future: power drawn by accelerators. So the floor is placed on watts, and the payload on profits.
Profits are where meaningful taxation belongs. Public finance has understood since Diamond and Mirrlees (1971) that taxing intermediate inputs distorts production while taxing profits does not. The design honors the principle: the input-side layer is priced to be invisible — a royalty for the commons, like the fee oil states charge on every barrel pumped, not a policy lever — and the consequential money sits on the profit side, exactly where the economics says it should.
One honest corollary, stated before a critic states it. Existing datacenters cannot move, but the next gigawatt can be sited anywhere — and that is precisely why the meter rate is trivial and must stay trivial. At two dollars per megawatt-hour, the royalty is twenty to fifty times smaller than the ordinary difference in power prices between candidate regions; no siting model registers it. The corollary is a designed ceiling: the meter layer can never quietly grow into the payload without reactivating the siting question. It is plumbing and precedent, permanently. The payload lives in Layer 2.
Honest Arithmetic
No proposal deserves to be taken seriously without numbers, so here are the unflattering ones.
Layer 1, at two to five dollars per megawatt-hour across the global AI buildout, yields on the order of one to five billion dollars per year by the end of this decade. Spread across a national population, that is not a dividend; it is a rounding error. This is by design, and it should be said plainly: for its first decade, the Intelligence Dividend is an institution, a ledger, and a precedent — not a check.
Layer 2 pays nothing until a single firm’s annual profits cross roughly $275 billion — nearly three times the most profitable companies in history. In the abundance world the founders describe, where a leading firm’s profits reach the neighborhood of one percent of world GDP — one trillion to one and a half trillion dollars a year, from one company — the schedule yields on the order of $60 to $160 billion annually from that firm alone, and low hundreds of billions across an industry with several such firms. If those profits never materialize, Layer 2 never fires, and no innovation was strangled by a payment nobody made.
Notice what this structure does to opposition. It cannot simultaneously be true that the royalty is too small to matter and large enough to kill innovation. Critics must pick an attack, and each attack concedes the other.
What This Means for Your Bank Account
The predictions in this section are not ours. They belong to the people building the technology. Dario Amodei describes a compressed century — decades of scientific progress arriving in a handful of years, with growth rates no industrial economy has ever recorded. Demis Hassabis calls the destination radical abundance. Sam Altman has published his own dollar figure for an AI-funded citizen dividend. Taken at their word, here is what this proposal means for a household, in three honest phases.
This decade: nothing. Say it plainly, because anyone promising you AI checks by 2027 is selling something. Layer 1 accumulates low billions per year — enough to build the fund, publish the ledger, and prove the institution; not enough to mail. Your account is unchanged. What you own is a claim.
The transition — if the predictions hold. Sometime in the 2030s, the first firm crosses the windfall threshold and the first distributions arrive: a few hundred dollars per person, per year. Financially modest; institutionally enormous. It is the moment the claim becomes a check — and the moment the fund becomes politically untouchable, because now everyone owns it.
The abundance world — the one the founders describe. Windfall inflows in the low hundreds of billions per year, layered on a corpus that has been compounding for a decade, support distributions on the order of $1,000 per person per year — $4,000 for a family of four — and rising, because an endowment under a spending rule never stops growing.
One ceiling deserves honest labeling. Altman’s published figure — $13,500 per adult, per year — would require roughly $3.5 trillion in annual distributions. That is the output of his much broader proposal, which would tax equity and land across the entire economy. This mechanism is deliberately narrower. Read his number as what the founders believe the AI economy can ultimately bear; read ours as the conservative floor a single, buildable instrument can credibly deliver.
If you believe the founders, this is your equity in the future they are building. If you do not, the claim costs you nothing to hold. Either way, it must exist before anyone knows which world arrives — because afterward, it will never be granted.
The Fund
Royalties flow into a permanent, independent public-benefit trust — the Intelligence Dividend Fund — that invests like a sovereign endowment and spends only its expected real return. The model is Norway’s: the Government Pension Fund Global, built from per-barrel petroleum revenue, is now the largest sovereign fund on earth and distributes roughly three percent a year while the principal compounds untouched.
Distribution is a direct, universal, equal dividend, visible in every citizen’s account. This is not sentiment; it is the security model. A permanent pool of public capital is the largest capture target a society can construct, and the countermeasure is not a mission statement — it is millions of shareholders. The Alaska Permanent Fund has paid a dividend to every resident since 1982 and has survived four decades of raid attempts for exactly one reason: every citizen has standing, and politicians who touch the principal lose. Diffuse benefits create diffuse defenders. Funds earmarked for abstract goods are quietly redirected; checks are defended.
And the dividend is not welfare; it is a royalty paid to owners. No one asks whether an Alaskan deserves an oil check, or whether a shareholder deserves a dividend.
Beyond the dividend rule: every inflow and outflow published on an open ledger, an independent board with staggered terms, and constitutional-grade protection of the principal wherever the jurisdiction allows it.
Yes, It Is a Tax
Proposals in this family often insist they are “not a tax.” This one does not. Layer 1 is a per-unit fee on a physical input — the same shape as a gas tax; Layer 2 is a contingent profits share. The claim is not that the Intelligence Dividend isn’t a tax. The claim is that it is the best-designed tax in the code: broad-based, physically unavoidable, priced below the noise floor where it touches activity, meaningful only where it touches windfall, and dedicated by law to a fund every citizen owns. Societies have always charged a royalty when private actors monetize a commons — per barrel of oil, per band of spectrum, per acre of federal land. This is that, for the infrastructure of intelligence.
The Real Argument
Two arguments carry this proposal. A third gives it its meaning.
Displacement. Artificial intelligence is the first technology whose explicit ambition is general substitution for cognitive labor. Every prior general-purpose technology displaced tasks; this one is aimed at occupations, by its builders’ own description. When a technology’s success is defined by how much human labor income it replaces, the compensating institution cannot be improvised afterward. It has to exist in advance. The dividend is that institution.
Social license. The greatest threat to AI diffusion is not a levy measured in hundredths of a percent. It is political backlash in a world where the technology’s benefits visibly concentrate while its disruptions visibly spread. And that backlash is not waiting for 2040 — it is arriving on this year’s utility bills, as households in datacenter regions watch electricity prices climb to power a buildout they hold no stake in. The Intelligence Dividend attaches to the very meter driving that anger, which gives the proposal a constituency now, not only in the abundance world. A dividend in every account is the cheapest insurance the industry will ever buy. The Intelligence Dividend is the price of the social license to automate.
Inheritance. Beneath both sits an older truth: AI is not built merely on humanity’s accumulated knowledge but of it — the mathematics, the publicly funded research, the open internet, the language and culture of billions, ingested nearly in full. Standing alone, this argument proves too much; every industry inherits. Joined to the first two, it explains why the stake should be permanent rather than remedial. Civilization is not a claimant on AI’s success. It is a founding stakeholder.
Anticipated Objections
“You are taxing an intermediate input, and AI is becoming an input to everything.” Conceded, and priced accordingly: the input-side royalty is roughly a twentieth of a percent of compute cost — a small efficiency cost traded deliberately for enforceability. The meaningful layer sits on profits, the base public finance approves.
“It’s premature. The frontier labs lose money.” Layer 2 collects nothing until a firm’s profits reach triple the most profitable companies in history. Layer 1 sits below the noise floor of any capital model. What is actually premature is waiting: a stake negotiated before the windfall costs a signature; after, it costs a war.
“Compute will move offshore.” Built assets cannot move — interconnection queues, substations, export-controlled silicon. Future siting is a choice, which is why the meter rate is designed to stay twenty to fifty times smaller than ordinary regional power-price differences: no siting model registers it. Begin in the jurisdictions where frontier compute physically sits, which today is a handful of countries, and harmonize outward. Every resource-royalty regime in history walked this path.
“The inheritance argument proves too much. Pharma is built on public biology — why doesn’t it owe this?” Because inheritance is not the load-bearing argument. Displacement and social license are, and they are specific to a technology whose ambition is general substitution for labor. Inheritance explains why the stake is permanent, not why AI is selected.
“A permanent trillion-dollar fund is the largest capture target ever built.” True — which is why the universal dividend is the security architecture, not the marketing. Universal checks create universal watchdogs; Alaska’s fund has outlasted every legislature that coveted it. Add a spending rule, an open ledger, and a staggered independent board.
“Corporate taxes already exist.” Single-digit effective rates achieved through intellectual-property shifting are the proof that the existing instrument fails for precisely this class of firm. A physically anchored royalty cannot be relocated to a tax haven, and Layer 2’s total-profit thresholds delete the apportionment game rather than replaying it. Earmarking does political work that general revenue cannot: no one marches to defend general revenue.
“A frontier-AI levy entrenches the oligopoly that can afford it.” The design never defines a licensed category and never names a firm. It thresholds on physical scale alone. Anyone at scale pays the same trivial rate; everyone below pays nothing. An open-weight challenger faces no paperwork a utility bill doesn’t already contain.
“Why not let the winners give it back voluntarily?” Because philanthropy is discretionary, revocable, and directed by donor taste. It produces gratitude, not standing. Carnegie built libraries; Alaska mails checks — only one of those is a social contract. But the voluntary instinct is right about sequence: Layer 2 begins as a signed commitment, not a statute. The choice is not between generosity and taxation. It is between a voluntary instrument now and an involuntary one later — the exact fork the industry’s own investors say is coming.
“Why should people who contribute nothing get a check?” Because this is ownership, not welfare. The premise of the proposal is that every citizen already contributed the commons the technology is built of — and a royalty paid to owners requires no means test, no work requirement, and no gratitude. Alaska has never asked whether a resident deserves their oil check.
Where This Begins — The American Implementation
The inheritance is humanity’s; the meter is in Texas.
A royalty requires a sovereign, and humanity does not have a treasury. The architecture in this document is universal — any nation can attach it to its grid and its tax code — but implementation begins in the United States, for three reasons.
The compute is here. The large majority of frontier accelerator capacity sits on American soil, draws American power, and runs on chips governed by American export controls. Collecting a royalty from Abilene requires no treaty; it requires a statute.
The precedent is here. The Alaska Permanent Fund — created under a Republican governor and defended for four decades by every ideology since — is proof that a citizen dividend funded by a commons royalty is durable American politics, not an imported experiment. This proposal is Alaska’s model, applied to the resource of the century.
The leverage is here. Whoever builds the first Intelligence Dividend writes the template the world copies, just as Norway’s fund became the global standard for resource wealth. An American fund makes American citizens the first shareholders of the AI economy — and makes the American social contract the export.
This is national leadership, not national exclusion. The meter layer works unilaterally today; the windfall layer will eventually want harmonization among the handful of countries hosting frontier compute, so that no signatory is disadvantaged — the path every resource-royalty regime has walked. Chapter one is American. The book is not.
Open Questions
The proposal leaves genuinely open: the exact meter threshold and rate (25 versus 100 megawatts; two versus five dollars); the windfall schedule’s entry point and top bracket; whether the fund’s first decade should return something visible to ratepayers in datacenter regions ahead of the corpus, or compound untouched; submetering standards for separating accelerator load where facilities are mixed-use; the detail of affiliation rules; sequencing between national adoption and international harmonization; and eligibility rules for the dividend. These are engineering questions. The architecture — watt today, windfall tomorrow, dividend forever — is the proposal.
The Central Question
If artificial intelligence becomes the most economically valuable technology humanity has ever created, should humanity hold a permanent royalty on that success?
This proposal answers yes — and insists on the corollary the timing implies. Equity is negotiated before the exit, not after. The Intelligence Dividend exists to be built now, while it costs almost nothing, so that it exists later, when it is worth almost everything.
Author’s note: I’m not an economist or a philosopher. I’m a product manager writing from Luzerne County, Pennsylvania — anthracite coal country, where a century of energy wealth was carted out and no dividend was ever mailed back. I live in what happens when a commons is monetized without a royalty. This document is a product spec for a piece of the social contract, assembled from parts built by people smarter than me — Bostrom’s windfall clause, Alaska’s dividend, Norway’s fund — and shipped because a spec left in a drawer helps no one. If you can break it, please try. Every objection that survives contact goes into the next version, with credit.
Prior art and references: Cullen O’Keefe et al., The Windfall Clause: Distributing the Benefits of AI for the Common Good (Centre for the Governance of AI, 2020) · Alaska Permanent Fund (est. 1976; dividends paid since 1982) · Norway Government Pension Fund Global · Sam Altman, Moore’s Law for Everything (2021) · Erik Brynjolfsson et al., GDP-B (NBER) · Peter Diamond & James Mirrlees, production-efficiency results (1971) · Connie Loizos, Neil Rimer thinks the AI money is coming back out (TechCrunch, July 2026)