The Corpus Dividend: The AI Economics Case for Free Artificial Intelligence and Full Public Ownership of America’s AI

America’s AI models were built from a corpus the public wrote and was never paid for, and the gains went to the investors who financed them. This framework, grounded in AI economics and machine learning scaling laws, traces that wealth transfer, argues for moving the models into public ownership immediately and in full through transfers, spin-outs and breakups, and works through what happens to shareholders on transfer day.

In my last piece I argued that the general-purpose layer of artificial intelligence should be free and ad-free, because it was distilled from a corpus the public wrote. A free tier settles what we are owed each time we use the product. It leaves the bigger asset untouched: model weights, the companies that own them, and equity now valued in the trillions of dollars, all assembled from our writing, code and mathematics.

This piece goes the rest of the way. My position is that the US AI businesses built on the uncompensated human corpus should be transferred to public ownership in full, on a single date, with no phase-in. That covers the independent labs, Anthropic and OpenAI, and the AI units inside Google, Microsoft, Meta and Amazon, which would be separated through spin-outs and breakups. The framework is limited to the United States: US companies, a US public trust and US residents as owners. I use the word reparations in its ordinary sense of repair for a taking. The taking here is concrete: billions of pages of text, code and proofs copied without consent or payment, including millions of books downloaded from pirate libraries.

The argument is theoretical, and it does not depend on any court ruling that training was illegal. Most US rulings so far have gone the other way, and I deal with that near the end. It depends on four properties of AI economics, the economics of machine learning at scale. From them I derive why the public should own the labs outright, why the public owes their shareholders nothing for them, what happens to each class of investor on transfer day, and how the labs run afterward. The math is included so you can check it.

A second theme runs through the whole piece: who has been paid so far. The value the public created did not vanish into the models. It became equity, and that equity has been flowing to venture funds, their limited partners, insiders who sold in tender offers and the shareholders of the largest technology companies. In my view, anyone who invested in a product built on taken work has been profiting from the taking. The transfer I propose returns that gain to the people it came from.

Four fundamentals of AI economics

The economics of artificial intelligence differ from the software era in four ways that bear directly on ownership.

1. The corpus is a factor of production

Software companies treated user content as a byproduct of people using a service. Large language models made it an input, and by far the most expensive one. Nikhil Kandpal and Colin Raffel estimated at ICML 2025 that for 64 language models released from 2016 to 2024, paying people to write the training data would have cost 10 to 1,000 times more than the compute used to train them. [1]

Write a model’s production function as Y = F(K, C, D), with K for compute, C for engineering talent and D for training data. The labs bought chips, power and talent at market prices and acquired most of D for nothing. In a competitive market each factor earns roughly what it contributes. When one factor is paid zero, its share of the return goes to whoever owns the others, and the market capitalizes that share into the valuation.

The timing makes this urgent. The Bureau of Labor Statistics reported that labor’s share of output in the nonfarm business sector fell to 52.8% in the second quarter of 2026, the lowest since the series began in 1947. [2] As more income flows to the owners of capital, the question of who owns AI capital becomes the question of who receives AI income.

2. The value of data is collective

A model learns little from any one person and a great deal from everyone together. Its quality follows scaling laws in the size of the whole corpus, which, as the next section shows, decides whether the claim belongs to individuals or to the public.

3. The past stock is fixed, and the future flow is elastic

Text already written cannot be unwritten or withheld. Text people will write next year can be. This split determines which parts of a remedy are free of economic cost and which parts need ongoing incentives.

4. Intelligence is nonrival, and its marginal cost keeps falling

My question to a model does not use it up for you. Serving an answer costs compute, and that cost is dropping fast: Epoch AI found the price of a fixed level of model performance falling between 9 and 900 times per year, with a median of 50 times. [3]

Harold Hotelling worked out how to price goods like this in 1938, writing about railways, bridges and utilities. With high fixed costs and low marginal costs, the efficient price equals marginal cost, because any higher price turns away people who value the service above what it costs to provide. Hotelling proposed funding the fixed costs from the least distorting taxes available. [4] For baseline AI, whose marginal cost is a few cents, a $20 monthly subscription excludes millions of people for no economic reason. Free baseline access is the efficient price. The open question is who will charge it, and a profit-maximizing owner never will.

What machine learning scaling laws say about who is owed

DeepMind’s Chinchilla paper modeled a language model’s loss as a function of its parameters N and its training tokens D [5]:

L(N, D) = E + A / N^α + B / D^β

E is the loss no model can remove, the A term penalizes a small model, and the B term penalizes too little data. The fitted values were E = 1.69, A = 406.4, B = 410.7, α = 0.34 and β = 0.28.

Take a 15-trillion-token corpus, about the scale of recent open models. The data penalty B / D^β comes to about 0.084. Remove one person's lifetime of writing, say 100,000 tokens or roughly a novel, and the penalty rises by about 1.6 × 10⁻¹⁰. That is a relative change of 1.9 × 10⁻⁹, which to first order equals β times the person's share of the corpus, and no benchmark could detect it. Remove half the corpus and the penalty rises by 21%, from 0.084 to 0.102.

B, beta = 410.7, 0.28          # Chinchilla data-term fit (Hoffmann et al., 2022)
penalty = lambda D: B / D**beta

D = 15e12                       # corpus size in tokens
one_person = 1e5                # one prolific writer

rel_one = (penalty(D - one_person) - penalty(D)) / penalty(D)
rel_half = penalty(D / 2) / penalty(D) - 1
print(f"one person: {rel_one:.1e}  half the corpus: {rel_half:.1%}")
# one person: 1.9e-09  half the corpus: 21.4%

Copyright litigation asks whether a particular plaintiff was harmed. In Kadrey v. Meta, Judge Chhabria granted Meta summary judgment because the authors could not prove market harm. [6] The scaling law predicts that result for almost any individual plaintiff, since no single contribution moves the model measurably. The value sits in the aggregate, where a legal system built on individual claims has no way to find it.

Cooperative game theory has a standard tool for this situation, the Shapley value. Its axioms require that the shares add up to the total value created and that contributors who add the same value receive the same share. When value depends on the size of the corpus rather than on who wrote which token, every contributor gets an equal share. So the public is owed the whole return attributable to data, and the fair way to hold it is one equal share per person.

Contributions are of course unequal in practice; a textbook author gave more than someone who left one comment. Registered works can still be paid through claims, as the Bartz v. Anthropic settlement did. For the billions of untraceable contributions, equal shares are the only rule anyone could administer.

Imanol Arrieta-Ibarra, Jaron Lanier, Glen Weyl and their coauthors argued in 2018 that data should be treated as labor, and that treating it as free capital distorts incentives and concentrates gains. [7] I would refine that with a distinction in time. Data that has already been written behaves like a capital stock the public owns in common, while data still to be written behaves like labor.

The wealth transfer: from the people who wrote the corpus to the people who financed the models

The public has already paid for the AI boom. It paid in kind, with the text, code and mathematics the models were trained on, and the payment showed up on investors’ balance sheets as appreciation.

The size of that transfer can be estimated from the production function above. If the data would have cost λ times the training compute, and talent cost about the same as compute, the corpus’s share of the inputs is:

θ = λ / (λ + 2)

At the low end of Kandpal and Raffel’s range, λ = 10, θ is about 0.83. At λ = 100 it is 0.98. Applied to the combined $2.37 trillion valuation of Anthropic and OpenAI, between about $2.0 and $2.3 trillion of that equity is the capitalized value of the public’s unpaid contribution. Spread across about 340 million US residents, the two labs alone hold about $5,800 to $6,800 per person that I would argue belongs to that person. The AI units of Google, Microsoft, Meta and Amazon would add to it. For comparison, the largest payment anyone has received for the corpus so far is the $1.5 billion Bartz settlement, about $3,000 per covered book. [14]

The gain did not stay in one place. Venture and growth funds bought into the labs at prices that assumed the main input would stay free. Their limited partners, including university endowments, private foundations, family offices and sovereign wealth funds, booked the appreciation as returns. Founders, early employees and early investors turned part of it into cash through tender offers and secondary sales, often at each new, higher valuation. Shareholders of Alphabet, Microsoft, Meta and Amazon saw it priced into their stocks as the AI premium, and Federal Reserve data show the wealthiest tenth of US households holding close to 90% of household stock and mutual fund shares, so most of that premium went to the people who already owned the most.

At every link, people captured value created by someone else, without the consent of the people who created it and without paying them. That is why I describe it as a wealth transfer from ordinary citizens to the investor class, and why I think the people at each link have been profiting from the scheme, whatever they believed when they wrote the check.

Notice matters here. The New York Times sued OpenAI and Microsoft in December 2023. Authors sued Meta in 2023 and Anthropic in 2024, and the Anthropic case produced a finding that keeping pirated books in a central library was not fair use. Every investor in the 2024, 2025 and 2026 funding rounds bought in knowing that the product rested on contested use of other people’s work, and priced that risk as a cost they would never have to pay.

Property law offers a useful analogy. A buyer of stolen goods generally cannot acquire good title, even if the buyer paid in good faith, because a seller cannot pass on rights it never had. Unjust enrichment law reaches the same place by another road: a party who gains at someone else’s expense, without a legal basis that justifies keeping the gain, can be made to give it back. Ownership of the models, built from property the public never transferred, follows the same logic.

The analogy has limits that I owe the reader. The Supreme Court held in Dowling v. United States (1985) that copyright infringement is not theft of goods for the purposes of federal stolen-property law, and most US courts have found training on lawfully acquired text to be fair use. So the claim of theft here is a moral and economic one, and the legal finding of wrongdoing is limited to the pirated books. The argument for reparations rests on the economics: an input worth trillions was taken without payment, and its value went to those who financed the taking.

Treating the corpus like land

Henry George argued in Progress and Poverty (1879) that land values are created by the community, and that because land cannot be produced or destroyed, taxing its rent changes who collects the income without reducing the supply. Economists across the political spectrum have since called the land value tax the least distorting tax there is.

A standard approximation for the efficiency cost of taxing a factor’s return is:

DWL ≈ ½ × ε × τ² × R

Here ε is the supply elasticity of the factor, τ the tax rate and R the factor’s return. When ε is zero the deadweight loss is zero at every rate up to and including 100%.

The text written before the era of large-scale training has ε = 0. No policy can make people write more of it in the past, and none can make them withdraw it. A public claim on all of its return therefore costs the economy nothing in lost output. George himself wanted the community to collect the full rent of land, and the same logic applies to the corpus in full.

The future flow behaves differently. Every day people decide whether to answer a question in public, edit an encyclopedia or publish code, and more of them are deciding not to. New monthly questions on Stack Overflow were down about 75% from their 2017 peak by late 2024. [9] The Wikimedia Foundation reported human pageviews down about 8% from March to August 2025 compared with a year earlier, and attributed part of that to AI answers. [10] Epoch AI puts the effective stock of public human text at about 300 trillion tokens and expects it to be fully used sometime between 2026 and 2032. [8] The labs still need new human writing, while the people who produce it are drifting away from the commons. That flow needs a real payment to keep it going.

The case for complete public ownership

Earlier versions of this argument, including mine, settled for a free tier plus a minority public stake. The fundamentals point further.

The Georgist result covers all of the corpus rent. A minority stake collects a fraction of a return that could be collected in full at no efficiency cost, and leaves the remainder with the parties who took the input.

Ownership also decides behavior. Hotelling’s efficient price for baseline AI is close to zero, and no profit-maximizing owner will set it there. OpenAI began showing ads to free users in early 2026 [26], the predictable move for an owner that has to earn a return on every user. A public owner whose charter requires free access has no such conflict. With a minority stake, the public would be asking each year for a concession that the controlling shareholders exist to resist.

Then there is the residual claim. Kandpal and Raffel’s estimates imply the data was worth more than everything else that went into the models. In any partnership, the party that supplied most of the capital is the natural residual claimant. Here that party is the public.

Finally, gradual transfer fails on its own arithmetic. An ownership share built by annual issuance at rate s follows 1 − (1 − s)^t, which approaches 100% without ever reaching it; at 2.5% a year it takes 91 years to reach 90%. Every year of delay also lets the taken value compound into private wealth and into share prices held by pension funds and index investors. Anthropic could begin its IPO roadshow the week of November 9. [11] Before the listing, the shareholders are a few hundred institutions, insiders and employees. After it, they include anyone with an index fund in a retirement account, so a transfer before the listing touches far fewer people.

Spin-outs and breakups for Google, Microsoft, Meta and Amazon

The independent labs are only part of the picture. Google, Microsoft, Meta and Amazon trained their own general-purpose models on the same public corpus and built them into search, office software, social apps and cloud services. Their AI businesses include Google DeepMind and the Gemini models, Microsoft’s Copilot products and its stake in OpenAI, the Llama models behind Meta AI, and Amazon’s Nova models and its stake in Anthropic. Taking the independent labs while leaving these alone would reward the companies that happened to absorb the corpus inside a larger balance sheet.

The scope rule is simple. Any US company that trained a general-purpose model on publicly available human text, code or mathematics, and that sells the model or builds it into its products, is covered. A statutory size threshold keeps startups and researchers who fine-tune open models out of it.

For conglomerates the transfer works in three steps.

1. Ownership moves on the transfer date. Each covered company places its model business in a new subsidiary: the research teams, the model weights, the training pipelines and the training clusters dedicated to them. The statute vests 100% of that subsidiary’s shares in the Human Data Trust the same day. Stakes in other covered developers, such as Microsoft’s share of OpenAI and Amazon’s and Google’s positions in Anthropic, move with them. Operational separation follows on a court-supervised schedule, the way AT&T’s was carried out after its 1982 settlement, but ownership does not wait for it.
2. The parent licenses the models back. Search, Office, Instagram and AWS keep running, and they pay the publicly owned model company for access at rates a regulator sets, on the same terms offered to any other customer. This is how AI revenue returns to the public: every AI search answer, Copilot seat and Meta AI conversation generates a license fee paid to a company the public owns.
3. Where separation is not yet practical, a corpus royalty applies. Some AI features are woven too tightly into a product to carve out quickly. Until they are separated, the parent pays the trust a fixed share of the revenue attributable to them, with the attribution audited by the trust.

This spin-out departs from the usual form in one respect. In an ordinary corporate spin-off, the new company’s shares are distributed to the parent’s own shareholders. Here they go to the trust.

Pricing the taking: what the public owes the shareholders

In the United States the Fifth Amendment requires just compensation when private property is taken for public use, and courts usually measure it at fair market value. The framework’s answer is a set-off. The public’s claim against each lab for the corpus is netted against the value of the shares, and if the claim is larger, the public owes nothing.

The claim can be priced the way Kandpal and Raffel did it, as the cost of paying people to write the training data:

Corpus claim = (tokens × words per token ÷ words per hour) × hourly wage

For a 15-trillion-token corpus at 0.75 words per token, 250 words of finished text per hour and $25 an hour, near the US median wage, the claim comes to about $1.13 trillion per lab. At 200 words an hour it is $1.41 trillion. Each lab used the full corpus, so each owes the full amount. For the two largest labs together, the claim lands between about $2.25 and $2.81 trillion. Their combined equity, using Anthropic’s last confirmed valuation of $965 billion [12] and OpenAI’s reported $1.4 trillion [13], is about $2.37 trillion.

Under median-wage assumptions, then, the public’s claim is about equal to or larger than the entire equity of both companies. Just compensation nets to roughly zero. At minimum wage the claim drops to about $330 billion per lab, which is why the assumptions belong in plain view. Courts would have to accept set-off and a replacement-cost measure, and under current doctrine they have not. That gap is the hardest legal problem the framework faces.

The same test applies to the conglomerates’ AI units. Google, Microsoft, Meta and Amazon each trained frontier models on a corpus of comparable size, so each faces a claim of roughly the same order, about $1.1 to $1.4 trillion under the median-wage assumptions. The set-off runs against the value of the spun-out AI unit, never against the parent’s search, software, advertising or cloud businesses. Those stay with their shareholders.

Transfer day: what happens to each class of investor

On the transfer date, every share of Anthropic and OpenAI, and every share of the newly separated AI units, moves to a US public trust. I’ll call it the Human Data Trust. The companies keep operating as companies; only their owner changes. Each holder’s position afterward depends on what it put in and what else it holds.

Investors in the independent labs

According to OpenAI, Microsoft holds roughly 27% of OpenAI Group, the OpenAI Foundation holds 26%, and employees and other investors hold the remaining 47%. [14] Anthropic’s largest outside backers include Amazon and Google, with Microsoft and Nvidia committing to invest in late 2025. Venture and growth funds hold much of the rest.

Venture funds lose their positions, and the losses pass through to their limited partners: endowments, foundations, family offices, sovereign funds and some public pension plans. The framework treats this as the return of value they never earned. These investors bought into products whose main input was taken without payment, after the lawsuits had made that fact public, and their returns depended on never paying for it. Most of what they lose is appreciation rather than cash put in, and that appreciation was the capitalized value of the corpus.

Some of the gain has already left the system. Founders, employees and early investors who sold in tender offers or secondary sales converted part of the taken value into cash. The transfer applies to shares, so that cash stays with them under this framework. That choice is practical, made to avoid years of tracing litigation. The same unjust enrichment logic would support a windfall tax on realized gains from the sale of shares in covered AI companies, and Congress could add one.

The OpenAI Foundation’s stake, which OpenAI valued at about $130 billion when its recapitalization closed [14], transfers like every other block. Its stated mission of making AI benefit all of humanity matches the trust’s, so its programs and staff can be folded into the trust.

Shareholders of Alphabet, Microsoft, Meta and Amazon

These companies lose twice: their own AI units and their stakes in the independent labs. Microsoft’s 27% of OpenAI is worth about $378 billion at a $1.4 trillion valuation, against roughly $13 billion it invested, much of it as cloud computing credits. What remains is their core business, plus two new relationships with the public AI companies. They become licensees, paying for models they used to own, and suppliers, selling the chips, data centers and cloud capacity those companies need. Anthropic alone has pledged at least $518 billion in infrastructure spending over ten years [15], and a free tier for every US resident would raise demand further.

Their shareholders take the loss through lower share prices as the market removes the AI premium from each parent’s valuation. These are public companies, so the loss reaches index funds and retirement accounts. It falls very unevenly, because most of that stock belongs to the wealthiest tenth of households, as the wealth transfer section showed. The transfer moves AI wealth from ownership weighted by existing wealth to one equal share per resident. A teacher with an index fund in her 403(b) loses a small amount on her Microsoft and Alphabet holdings and receives a corpus unit worth exactly as much as the largest shareholder’s.

Employees

Employee equity is a different case, because it was paid as wages for work, and the framework’s central principle is that every factor is paid for what it contributes. Vested employee shares in the labs and the spun-out units convert to cash retention awards up to a cap, set at a multiple of salary and paid over three years from operating revenue. Shares above the cap transfer like everyone else’s. Unvested grants are replaced with market-rate cash compensation. The public AI companies need their researchers and engineers on day two, and they have to be able to compete for them.

Lenders and bondholders

Debt survives intact, since lenders supplied money and that money was a contribution. Anthropic’s draft IPO filing showed a $72 billion accounting charge tied to convertible bonds that can be exchanged for equity. [15] After the transfer there is no private equity to convert into, so those bonds pay out as ordinary debt on their original terms. Data center leases, chip orders and cloud contracts stay in force.

Markets

Expect a sharp repricing across AI-exposed stocks and a halt to private venture investment in US frontier model training, since the transfer tells investors that future corpus-built models would be treated the same way. Some of that capital will move to other sectors and other countries. I return to that risk in the last section.

The public AI companies after the transfer

The trust becomes the sole shareholder of each covered lab and spun-out unit, and three things change on the first day.

The free tier ships. Each public AI company provides free, ad-free access, for every US resident, to a model no more than one generation behind its flagship. My earlier cost model, Annual cost = N × t × 365 × p, put a generous tier for one billion users at about $7.1 billion a year, assuming 20,000 tokens per user per day at $0.97 per million tokens. A tier for the roughly 340 million people in the United States would cost about a third of that. Paid tiers, enterprise contracts, API sales and the license fees from the former parents continue, because they fund the institution.

The commons budget starts. A fixed 3% of revenue goes to the future flow of human knowledge: Wikipedia, public libraries, open-source maintainers, a claims pool for registered works, and direct payments to US residents who choose to contribute new data, paid as labor. The two largest independent labs alone run at a combined revenue rate above $100 billion, with Anthropic reporting $16.8 billion for the third quarter. [15] Three percent of that base is more than $3 billion a year before the spun-out units and license fees are counted.

Every US resident receives one corpus unit. Units carry an equal claim on the trust’s dividends, as Shapley symmetry requires, and they cannot be sold, pledged or inherited. Tradable units would concentrate again within a generation, which is exactly what this framework exists to undo. At current valuations Anthropic and OpenAI alone are worth about $7,000 per US resident, before counting the AI units spun out of the four conglomerates. Dividends start once operating cash flow covers the free tier, the commons budget and reinvestment. Until then the return arrives as free access and an appreciating common asset. Existing rails can carry the payments: Alaska has paid a per-person Permanent Fund dividend since 1982, $1,200 this year [16], and the Trump accounts that opened on July 4, 2026 already hold a seeded investment account for every child born from 2025 through 2028. [17]

Governance follows the model of Norway’s sovereign fund rather than a government ministry. A professional board runs the trust under a statutory charter, and each public AI company keeps its own management. The charter forbids any government official from directing model outputs, training choices or content policies outside of generally applicable law. That firewall is the most important clause in the design, for reasons covered below.

Future compute is financed by retained earnings and by bonds issued by the trust, the way the Tennessee Valley Authority has financed power plants with its own bonds for decades. Private capital can still lend to the public AI companies, but it can no longer own them.

Precedents for public ownership and breakups

Each element of this design already exists somewhere in American law or practice.

Breakups are an old American tool. The Supreme Court broke Standard Oil into 34 companies in 1911. The Justice Department’s case against AT&T, filed in 1974 and settled in 1982, split the Bell System into seven regional companies on January 1, 1984, while AT&T kept long distance and equipment. Both show that a court-supervised separation of a dominant company is possible, and that the separated parts keep operating.

Public ownership of a shared resource is also familiar. Alaska’s constitution reserves the state’s natural resources for the benefit of its people, and its Permanent Fund pays every eligible resident a share of the proceeds. The Tennessee Valley Authority has been a federally owned power company since 1933. Abroad, Norway turned public ownership of its oil into a fund worth $2.39 trillion at the end of June 2026. [18]

Public equity in technology companies came back in the last fourteen months. In August 2025 the federal government took roughly 10% of Intel by converting about $8.9 billion in CHIPS Act and defense grants into shares. [19] In July 2026 OpenAI was reported to have proposed giving the government a 5% stake, about $50 billion at a $1 trillion valuation, to ease political pressure before a public offering. [20]

Even the AI industry has written versions of this. OpenAI began in 2015 as a nonprofit dedicated to benefiting all of humanity, and its nonprofit still holds a quarter of the company. In 2019 Cullen O’Keefe and colleagues at Oxford’s Future of Humanity Institute proposed the Windfall Clause, under which AI firms would commit in advance to give away 1% of profits between 0.1% and 1% of gross world product, 20% between 1% and 10%, and 50% beyond that. [21] Sam Altman proposed an American Equity Fund in 2021 that would take a share of large companies’ value every year [22], and he has since said that people need to share in the upside through ownership of equity or compute. [23]

What this framework adds is the justification. Those proposals ask for generosity or wait for a windfall. This one treats public ownership as the settlement of a debt that is already owed.

Where the framework breaks

The legal premise is contested. US courts have mostly held that training on lawfully acquired text is fair use, and on September 16, 2026 the Ninth Circuit upheld dismissal of programmers’ DMCA claims against GitHub Copilot and OpenAI Codex. [24] Many readers will reject the word reparations because the law has not found a wrong, apart from the pirated books. My response is that this is a policy question about an unpaid factor of production, of the kind legislatures decide through royalties and spectrum auctions. Anyone who rejects that premise will reject the rest of the argument, and their objection has real weight.

The Takings Clause is a serious obstacle. The set-off argument requires courts to price the corpus at replacement cost and to net it against the shares, and no court has done that. Without it, a complete transfer at zero compensation would face constitutional challenge, and a negotiated buyout at market value would cost trillions.

Breakups are slow and contested in practice. AI is woven through search, productivity software, advertising and cloud services, and every line between model and product will be litigated. The AT&T case took ten years from filing to divestiture. Ownership can move on day one by statute, but the operational separation and the revenue attribution behind the corpus royalty could take years to settle.

The economic record of state-owned enterprises is mixed. Public owners have often been slower, more cautious and easier for politics to capture. The Norway model works partly because of Norway’s institutions, and US institutions would need to match them.

Political control over AI models is the most dangerous failure mode. A government that owns the leading models could shape what they say about elections, dissent and history. The statutory firewall exists for this reason, and it would need to be as strong as the protections around central banks and public broadcasters, and probably stronger.

Capital and talent can leave. If transfer becomes a known risk, private investment in frontier AI will move to other countries or into forms the rules do not cover, such as open-weight releases and foreign labs. The economist William Gale raised related concerns about avoidance and base erosion when Altman first floated his fund. [25] The cap on employee retention awards could also send researchers to foreign competitors.

The US scope leaves people out. The corpus was written by people everywhere, and a US trust paying US residents excludes every non-US author whose work trained these models. Other countries would have to act for their own residents and companies, and they would face the same questions.

The numbers are fragile. Anthropic’s draft filing showed an $80.7 billion first-half net loss, most of it the non-cash convertible charge, and some investors have reportedly cut their valuation target to about $1.5 trillion from $1.8 to $2 trillion. [15] The set-off estimate moves with every assumption about writing speed and wages.

The AI economy is setting its terms now

The economics of artificial intelligence have already produced the facts this framework needs. The corpus was the most valuable input to the models and was paid the least. Its value is collective, so individual lawsuits will keep failing to find it. Because the past corpus is a fixed stock, the public can claim all of its return without economic cost. Because the product is nonrival with collapsing marginal cost, free access is the efficient price, and only an owner without a profit mandate will charge it.

Those facts lead to a complete transfer in the United States: the independent labs move to a public trust, the AI units of Google, Microsoft, Meta and Amazon are spun out to the same trust, their former parents pay license fees to use the models, every US resident holds an equal and unsellable unit, baseline AI is free for everyone in the country, and part of the revenue keeps the human commons alive. The decision gets harder after each listing and each funding round, and the next listing could begin within weeks. Until that happens, every new funding round and every tender offer moves more of the public’s work into private hands.

About the Author

Brian James Curry is a Kansas City artificial intelligence researcher, applied scientist and data scientist working in machine learning, causal inference, AI economics and healthcare intelligence. He is the founder of Vector1 Research, an independent applied AI research lab in Kansas City, and of KC AI Lab, which began in 2016 as the Kansas City Machine Learning Group and has grown to more than 1,500 members.

Brian has more than twenty years of experience leading analytics, data science and machine learning at Hallmark, Garmin, Koch Industries, Tractor Supply Company and Vail Resorts. His open-source research tools on GitHub include PyCausalSim (causal discovery through simulation), Papilon (marketing mix modeling) and tincast (forecasting for AI-critical commodities). He writes about AI economics, agentic AI, causal inference and healthcare AI on Medium and Substack. Contact: brian@vector1.ai.

References

1. Kandpal, N. & Raffel, C. (2025). Position: The Most Expensive Part of an LLM should be its Training Data. ICML 2025, PMLR 267.
2. US Bureau of Labor Statistics (2026, Sep 15). Labor share at its lowest level, 52.8 percent, in second quarter 2026.
3. Epoch AI (2025). LLM inference prices have fallen rapidly but unequally across tasks.
4. Hotelling, H. (1938). The General Welfare in Relation to Problems of Taxation and of Railway and Utility Rates. Econometrica 6(3).
5. Hoffmann, J. et al. (2022). Training Compute-Optimal Large Language Models. arXiv:2203.15556.
6. Loeb & Loeb (2025). Kadrey v. Meta Platforms, Inc.
7. Arrieta-Ibarra, I., Goff, L., Jiménez-Hernández, D., Lanier, J. & Weyl, E. G. (2018). Should We Treat Data as Labor? Moving beyond “Free”. AEA Papers and Proceedings 108.
8. Villalobos, P. et al. / Epoch AI (2024). Will we run out of data? Limits of LLM scaling based on human-generated data.
9. DevClass (2025, Jan 8). Coding help on StackOverflow dives as AI assistants rise.
10. Computing (2025). Wikipedia reports traffic drop, blames AI search and social media.
11. The Motley Fool (2026, Oct 22). Why Anthropic Is Racing to IPO While OpenAI Waits Until 2027.
12. CNBC (2026, May 28). Anthropic tops OpenAI as most valuable AI startup.
13. Tech Funding News (2026, Oct 20). OpenAI races to lock in $30B at a $1.4T valuation before Anthropic’s IPO.
14. OpenAI. Our structure.
15. Quartz (2026, Oct 21). Anthropic’s first-half net loss reached $80.7 billion, draft IPO filing shows.
16. Alaska Beacon / States Newsroom (2026, Sep 30). Permanent Fund dividends will be distributed to Alaskans starting this week.
17. NBC News (2026). Trump accounts launch July 4 for newborns.
18. OilPrice.com (2026, Aug). Norway’s $2.3 Trillion Wealth Fund Posts Record High Profit in H1.
19. TechRepublic (2025, Aug 25). US Government Secures 10% Stake in Intel in an Unprecedented Deal.
20. CNBC (2026, Jul 2). OpenAI proposes 5% stake to Trump administration to ease Washington pressure: Report.
21. O’Keefe, C., Cihon, P., Garfinkel, B., Flynn, C., Leung, J. & Dafoe, A. (2020). The Windfall Clause: Distributing the Benefits of AI for the Common Good. AIES 2020.
22. Altman, S. (2021, Mar 16). Moore’s Law for Everything.
23. Yahoo Finance (2026). Sam Altman Once Backed $13,500 Checks For Everyone, Now Says Everybody’s Got To Participate.
24. Gibson Dunn (2026, Sep 18). Ninth Circuit Clarifies Limits of DMCA Liability for AI-Generated Code.
25. Centre for the Governance of AI (2022). Sam Altman and Bill Gale on Taxation Solutions for Advanced AI.
26. AdNews (2026, Jan 19). OpenAI launches advertising tier for ChatGPT in U.S.