Who Paid for AI? The Public Money Behind the Private Fortune

Ask who funded AI and you will usually get a story about garages, dropouts and daring capital — a founding myth in which private genius saw what nobody else could see, risked everything, and was rewarded in proportion to its nerve. It is a good story. It is also, on the evidence of the funding record, mostly wrong about the part that mattered most. For roughly half a century, the expensive, uncertain, unglamorous work that made today’s AI possible was carried by public money: defence research agencies, national science funders, state university departments, publicly backed research institutes. Private capital arrived late, arrived hard, and arrived after the riskiest decades were already behind it.

I want to be careful here, because this argument is often made badly. I am not claiming that companies contributed nothing, or that public funding alone would have produced a product anyone could use. Both of those claims would be false. What I am claiming is narrower and, I think, harder to dodge: the pattern of who bore the risk and the pattern of who collected the reward do not match, and the mismatch was not an accident of history. It is the normal shape of how a technology travels from the commons into private hands.

The record does not look like the myth

The field has a founding date and a funder. The 1956 Dartmouth workshop that gave artificial intelligence its name was a summer research project convened by academics and underwritten by a philanthropic foundation — not a venture round, not a product plan. What followed in the United States was decades of sustained public research money: the Office of Naval Research, then the Defense Advanced Research Projects Agency, then the National Science Foundation, pouring support into a handful of university laboratories at MIT, Stanford, Carnegie Mellon and elsewhere. Public agencies paid for the early work on search, planning, natural language, speech recognition and robotics. They paid for it for years at a stretch, without a product, without a market and frequently without a result.

The same is true well beyond America. Britain, Japan, Canada, France, Germany and later China all ran state-funded AI research programmes of one kind or another. So did the public universities that trained nearly everyone who later became a founder. The graduate student stipends, the departmental compute, the conference travel, the long unproductive years spent on an idea that might not work — a great deal of that was taxpayer-funded, in country after country, for decade after decade.

I will not put a number on it, because honest numbers here are hard. Research funding is scattered across agencies, countries, decades and budget lines that were never designed to be added up, and the figures that circulate are usually assembled to win an argument. But you do not need a precise total to see the shape. The institutions that kept AI alive in its unprofitable years were overwhelmingly public ones.

The winters are the proof

If you want to test whether private capital would have carried this technology on its own, look at the periods when it had the chance and declined.

AI has had at least two long winters. The first came in the 1970s, after an influential and sceptical British government report and a broader collapse of confidence in the field’s early promises; funding tightened, ambitions shrank, and the word “AI” became something you avoided putting in a grant application. The second came in the late 1980s and ran into the 1990s, when the expert-systems boom — which had attracted serious commercial money — collapsed, and the companies built on it went with it. Investors left. They did not leave because they were foolish. They left because the returns were not there and were not going to be there for a very long time, which is exactly the judgement private capital is designed to make.

The work continued anyway, in the places that were not being asked to show a return. Neural networks in particular spent years as an unfashionable dead end, kept going by a small number of stubborn researchers inside publicly funded universities and institutes. The Canadian Institute for Advanced Research is the cleanest example: from the mid-2000s its programme brought together the groups around Geoffrey Hinton, Yoshua Bengio and Yann LeCun and funded precisely the kind of long-horizon, no-deliverable research that no quarterly earnings call would ever tolerate. That work is now the foundation of the industry. It was supported, through the years when it looked like nothing, by public and philanthropic money on the explicit understanding that it might amount to nothing at all.

Private capital did not miss the opportunity. It correctly judged the risk, and someone else was left holding it.

That is the heart of it. The public did not merely co-invest alongside private investors. The public invested in the years when private investors had looked at the same technology, priced the risk accurately, and walked away. Then, once the risk had been absorbed and the results were visible, the money came back — and the returns from fifty years of publicly financed patience landed almost entirely in private hands.

What the companies actually did — and it was not nothing

Now the other half, because an argument that only works by leaving things out is not worth making.

Turning a research result into something a billion people can use is genuinely hard, genuinely expensive and genuinely risky in its own right. Corporate labs produced foundational research too, including some of the architectural breakthroughs the current generation of models is built on. Firms made an enormous bet on scale — that making models much larger and training them on much more data would keep producing gains — at a point when that was a real gamble and many serious researchers thought it would plateau. They built the chips, the data centres, the training infrastructure, the tooling and the interfaces. They absorbed years of losses. They did the unglamorous engineering that separates a paper from a product, and that work is not a footnote.

So the fair account is this: the public took the long, unrecoverable risk on the science; private firms took a shorter, better-informed risk on the products and did the enormous work of commercialisation. Both are real contributions. Only one of them shows up in the story the industry tells about itself, and only one of them is currently being paid.

Mazzucato’s argument, and why it is bigger than AI

The economist Mariana Mazzucato made this case most forcefully in her 2013 book The Entrepreneurial State, and it is worth attributing to her properly rather than absorbing it into the general air. Her argument is that the state is not merely a fixer of market failures but an active risk-taker and market-shaper — that it invests in exactly the uncertain, early, long-horizon research that private capital avoids, and that it then fails to claim any share of the upside it created.

Her most famous illustration is the smartphone. Take an iPhone apart, she argued, and nearly every technology that makes it smart traces back to public funding: the internet from ARPA-backed research, GPS from the US military, touchscreen work supported by public research money, voice assistance descended from a DARPA-funded programme at a research institute, the lithium-ion battery with roots in publicly funded energy research, the web itself from CERN. Apple’s genuine achievement was integration and design — combining those publicly de-risked components into an object people loved. That is a real achievement. It is just not the same achievement as inventing the components, and the difference matters when the profits are divided.

Her thesis has critics, and they are worth hearing: some argue she overstates state causation, understates how much invention happened inside private labs, and glosses over how much value is created in the commercialisation she treats as downstream. I think those criticisms land against the strongest versions of the claim and not against the modest one, which is all I need here — that public money systematically funds the phase where failure is likely, and systematically collects nothing when it succeeds.

This is the same pattern I keep running into everywhere: technologies that begin as broad collective efforts and end as private property. It is worth reading alongside how technology gets captured, because the funding story is one of the clearest mechanisms by which capture happens, and alongside the composite man, because the myth of the lone founder and the myth of the lone funder are the same myth wearing different clothes.

Socialised risk and privatised reward is not a slogan. It is an accurate description of the cash flows.

Why funding history is a moral argument

Here is why any of this matters beyond the history seminar.

Finance has a very clear principle about who deserves returns: those who bear the risk. It is the justification for every venture capitalist’s carry and every founder’s equity, and it is not a bad principle. I am simply asking that it be applied consistently. If bearing risk earns you a claim on the reward, then the public — which bore the largest, longest and least recoverable risk in the entire history of this technology — has a claim. Not a favour. Not charity. A claim, on the industry’s own stated terms.

At present that claim is settled with tax revenue at whatever rate the most mobile companies on earth can be persuaded to pay, and with the general benefit of the products existing. That second one is real, and I do not want to wave it away; useful things being available to people is a genuine return. But it is the return you would get if the public had contributed nothing at all. It is not a share.

The deeper point is one Daron Acemoglu and Simon Johnson make about direction rather than quantity: what a technology does, and for whom, is decided by who holds power over its development, not by the technology itself. I have written about that in Acemoglu on power and progress. Funding is one of the oldest and most direct forms of that power — and a funder who takes the risk and then asks for nothing has given away the steering wheel along with the money.

What a public stake could reasonably look like

None of the options here are exotic. They are ordinary instruments, already used elsewhere, and it is only familiarity with the current arrangement that makes them sound radical.

• Equity or royalty on publicly funded research. When public money funds work that becomes a commercial product, the funder takes a small stake or a royalty stream — and reinvests it in the next round of risky research, making the fund revolving rather than one-directional.
• Conditions attached to support. Mazzucato’s own preferred remedy: if you take public money, public subsidy, public land, publicly built energy infrastructure or publicly funded research, conditions come with it — on pricing, on access, on reinvestment, on where the gains are booked for tax.
• Public compute and public models. Publicly funded compute capacity and open models, so that researchers, universities and smaller firms are not entirely dependent on renting from the few companies that own the infrastructure.
• Retained rights. Keeping enforceable public interest rights in publicly funded inventions, including the ability to act when a publicly funded result is priced or withheld in a way that defeats the purpose of funding it.
• Honest accounting of the inputs. The subsidy did not stop in the 1990s. Data centres today draw on public electricity grids, public water, public land deals and tax concessions, and the models are trained on a collective inheritance of human writing and images. That is a continuing public contribution, and it should be on the ledger.

Reasonable people will disagree about which of these is wise. A badly designed public stake could easily pick losers, entrench incumbents, or hand politicians a lever they should not have. Those are real objections and I would rather argue about the design than pretend the risks are imaginary. But the question of whether the public has standing to ask is, I think, settled by the funding record — and once you accept the standing, the rest is engineering.

The myth is doing work

Founding myths are never just stories. They are arguments about desert, told in the past tense so they cannot be contradicted. If private genius built AI alone, then every rupee and dollar of the returns belongs where it currently sits, and any claim on it is confiscation. If the public carried the technology through five decades that private capital declined to finance, then the current distribution is a choice that was made — recently, by people with names — and choices can be made differently. That is why the story is told so insistently, and it is why the question of who owns AI cannot be separated from the question of who paid to make it possible.

Both things are true. The public de-risked it; private firms built and shipped it. What I object to is not the profit. It is the amnesia — the confident retelling of a history in which one of the two contributors simply does not appear, and which therefore concludes, effortlessly and without argument, that only one of them is owed anything at all.