Ministry of AI · Dispatch from 2047
Public Funding of AI Research: The Seed We Forgot
Editorial note. The Ministry of AI is a work of disciplined foresight: it describes the year 2047 in the present tense, and treats our own era as history. The institutions are imagined. The economics, the evidence and the historical parallels are real and sourced.
My middle son is fifteen and has decided he wants to work on protein design. He asked me last week who pays for that. I told him the truth — that for the first several years, nobody who expects anything back does — and watched him accept it as an ordinary fact of the world. That is the sentence I want to hold up. In 2047 a fifteen-year-old assumes that a society funds work whose payoff arrives after everyone involved has retired. In the decade I worked in, we had almost stopped believing that, right up until the moment the largest commercial fortune in history was harvested from exactly that kind of patient, unglamorous, publicly financed seed.
I was in the rooms where the harvest was booked. I built growth and measurement systems for companies deploying the first coordinated AI stacks, and I can tell you what our internal accounting recognised as an input cost: compute, licences, salaries, data acquisition. That was the list. Nowhere on it was the four decades of government money that made every item on the list possible. It was not concealed. It was simply not a category. You cannot expense an inheritance.
What the seed actually paid for
The history is documented and dull, which is why it lost the argument for so long. The National Academies published a full accounting of it in 1999 — Funding a Revolution: Government Support for Computing Research — tracing how federal agencies financed timesharing, networking, graphics, and artificial intelligence itself through decades in which no commercial market for any of it existed. DARPA funded machine intelligence through two winters when the private sector had written the field off entirely. And this was not a historical phase that ended when the money got good: by the mid-2020s the U.S. National Science Foundation was still investing over $700 million a year in AI — institutes, instruments, and the shared research infrastructure that academic labs used because they could not afford industrial compute.
The economics of that arrangement were understood too. Work on public research funding found that its returns show up overwhelmingly outside the funded institution — one careful study of NIH funding rules showed public grants generating private-sector patents in firms that had never received a dollar of the grant. That is the whole shape of the thing. Public money buys the option; private capital exercises it. Everyone treated the spillover as a happy accident rather than as a claim.
Here is the part I insist on, because the argument dies without it: the private laboratories were not parasites. The architecture that made everything after 2017 possible was published by a corporate research team, freely, for anyone to read. The scaling engineering was hard, expensive and genuinely inventive. The residual networks that unlocked deep vision came out of an industrial lab. If you need the companies to be villains for this argument to work, you do not have an argument; you have a grievance. What you have instead is a sequencing problem. The state carried forty years of risk at zero return, the market carried five years of risk at enormous return, and only the second stretch had a ledger attached to it.
The public-seed schedule
So the Ministry built the missing ledger. When a system is assessed for its Machine Yield Account, the assessors compile what we call a public-seed schedule — an evidentiary record of the publicly financed inputs its capability descends from, sorted into four bands.
| Band | What it covers | How it is evidenced | What it establishes |
|---|---|---|---|
| Foundational research | Methods, architectures and theory originating in publicly financed laboratories or grant-funded work | Grant records, publication lineage, agency programme histories | That the capability has ancestors nobody billed for |
| Public infrastructure | Networks, shared compute, standards, national research facilities | Facility allocations, network provenance, procurement records | That the training run was physically possible because of collective assets |
| Public corpora | Publicly funded science, statistical series, archives, libraries, government records inside the training data | Dataset composition, licence classes, archive provenance | That the machine’s knowledge includes work the public already paid for once |
| Trained labour | Researchers and engineers educated in publicly financed institutions | Institutional records, aggregate education financing | That the human capital was a public product before it was a private hire |
Four bands, and every one of them is retrospective. A schedule does not tell you what to charge. It tells you the capability did not begin where the invoice begins. The assessment itself still happens where it always happens — at the deployment boundary, where an autonomous system does work a paid human would otherwise have done, measured against the displacement record for that role. The schedule is the reason nobody wins the argument that the yield is purely a return on private capital. It is a lineage, not a bill.
I want to be precise about the difference, because the distinction is the whole institutional design. A bill implies a creditor. The seed schedule has no creditor. There is no agency that can plausibly claim the return on a 1970s networking grant, no university that owns the concept of an educated engineer, no treasury with standing to invoice for a benchmark dataset. What the schedule proves is that the intelligence was grown in common soil. What follows from that is not a royalty. It is a dividend on an inheritance, paid to the population, because the population is the only entity whose claim is not arbitrary.
Why we refused the two obvious alternatives
The first proposal, every time, is sovereign equity: let the state take shares in the companies its research seeded. It is intuitive and we rejected it twice. A government holding equity acquires an interest in the valuation of the thing it is supposed to meter, and a regulator that owns the regulated writes soft rules. Worse, equity returns land in a treasury, and treasuries reallocate. Within one budget cycle an inheritance becomes a contribution to general revenue and the citizen’s claim is gone, unnoticed, in a footnote.
The second is a research royalty — trace the seed, charge a percentage. This one died on measurement. You can establish that publicly financed work is load-bearing in a system’s lineage; you cannot establish that it constitutes eleven percent of it. Any number you produce is a negotiation dressed as arithmetic, and an instrument built on a negotiated number does not survive its first hostile government. We had learned that lesson expensively already: the value estimate is only credible when the error is published with it.
So we kept the meter where the physics is clean and used the seed schedule to answer the only question the meter could not answer on its own — why does this yield belong to everyone rather than to the balance sheet it landed on? Because the seed was ours. All of it, none of it apportionable, which is exactly why the payout is universal and flat.
A Thursday in a building nobody was going to keep
My mother is seventy-three. Nineteen years reading insurance claim files, then a system trained on her own decisions, then redundancy at fifty-eight and a dividend at fifty-nine. On Thursdays now she takes the bus to a small research station on the edge of the city — a soil and water monitoring post that was scheduled for closure in the 2030s because it produced no output anyone could sell. It monitors. That is all it does. It has published nothing anyone has read in eleven years.
The town kept it open out of dividend-funded local allocation, and my mother sits in it two afternoons a week logging readings she does not fully understand, alongside a retired plumber and a woman who used to schedule freight. My son will use data like theirs in thirty years, in work none of them can imagine, and none of them will be paid for it or remembered.
This is what the seed looked like when it was being planted, and it never looked like anything else. Unglamorous rooms, patient people, no return in the lifetime of the funder. The reason the argument of my century failed was not that anyone disputed the history. It was that the seed is always boring and the harvest is always spectacular, and ledgers are built by people watching the harvest.
What it does not fix
Apportionment, and therefore geography. The seed schedule establishes lineage, not shares — so the yield is still collected where deployment is booked. That was an administrative decision made for administrative reasons, and its distributional consequence is ugly: countries whose public universities trained the researchers, whose archives fed the corpora, and whose taxpayers financed decades of foundational work collect nothing unless the deployment happens to be booked inside their borders. Half the seed schedule of any large system is foreign. None of the yield is.
There is a committee working on it, as there has been for eleven years. I have read three of its drafts. Every one of them requires a global apportionment key, and every one of them collapses at the same point: you can prove the ancestry and you cannot price it.
We built the instrument we could defend and we are living off it. My son will go and work on his proteins, funded by nobody who expects anything back, and in forty years someone will harvest that too. I would like the ledger to be ready that time.
FAQ
How much of AI research was publicly funded? Enough that the private breakthroughs are unintelligible without it. The National Academies documented decades of federal financing for computing research before any commercial market existed, and by the mid-2020s the U.S. National Science Foundation alone was still investing over $700 million a year in AI, plus shared compute for laboratories that could not afford industrial infrastructure.
Did companies invent modern AI or did the state? Both, in sequence. The decisive engineering of the 2010s and 2020s was largely private and genuinely inventive — the key architectures were published by corporate laboratories. But the field they were working in, the networks they ran on, the benchmarks they measured against and the people who did the work were financed publicly through forty unprofitable years.
What is a public-seed schedule? The Ministry’s evidentiary record for a deployed system: a four-band classification of the publicly financed inputs — foundational research, infrastructure, corpora and trained labour — that its capability descends from. It establishes lineage for the assessment; it does not price any contribution.
Why not take state equity in AI companies instead? Because a regulator holding shares in the regulated has an interest in the valuation and writes softer rules, and because equity returns land in a treasury rather than with the population. An inheritance parked in general revenue stops being an inheritance within one budget cycle.
Isn’t this just a return on taxes paid? That is the weakest version of the claim and it does not carry the design. Taxes financed the seed, but the training corpus was written overwhelmingly by people who were never funded and are owed anyway, including the dead. The payout is universal because the inheritance has no register of contributors.
What does the argument fail to settle? Apportionment. Lineage can be proven; shares cannot. So yield is collected where deployment is booked, and the countries that seeded capability without hosting deployment collect nothing. No one has yet built a global apportionment key that survives contact with a finance ministry.