Sam Altman converted OpenAI’s nonprofit, mission-bound charter into a roughly five-hundred-billion-dollar private equity vehicle.
Dario Amodei founded Anthropic in April 2021 as a Delaware Public Benefit Corporation with charter language where a structural commitment should have been. The two arrived at the same destination by different roads. Altman through conversion: in 2019 he created a capped-profit subsidiary under nonprofit control to attract outside capital, and on October 28, 2025, with the blessing of the Delaware and California attorneys general and Microsoft’s consent, completed the conversion into a Public Benefit Corporation in which the renamed OpenAI Foundation retains 26% of an enterprise now valued at approximately $500 billion — a stake worth on the order of $130 billion in its own right — and Microsoft holds 27%. Around the same restructuring, the word safely was quietly removed from OpenAI’s mission statement, as the 2024 Form 990 filed in late 2025 confirms. Amodei through original sin: Anthropic was constituted from day one as a PBC with an embedded Long-Term Benefit Trust whose disinterested trustees hold voting authority over a portion of the board. The corporate form is a vibes clause in lieu of a structural commitment. Both companies are now structurally committed to delivering investor returns. The thing they were founded to refuse is what they became.
This is the sentence Bruce Schneier and Nathan Sanders are reaching toward, more politely than they quite say, in the column picked up everywhere this week. Their argument is the more interesting one: if the market eventually concludes that the frontier labs cannot produce a sustainable return on equity, the public should be ready. Their answer — preemptive nationalization, structured in the manner of the postwar national labs and the regional supercomputing centres, with an innovation function run like a lab and a compute function run like a utility — is the right shape of argument. The alternative they propose to displace, “let the labs collapse and let the technology disperse to whoever picks it up,” is the same answer the country has produced for every prior case of mission-driven science that outgrew its mission: the Bell System after 1984, ARPANET after 1990, Conrail after privatisation, the S&L cleanup after 1989. The pattern is consistent. The country lets the public-good version of the technology die, then spends the next thirty years arguing about whether the privatised version counts as a public good. It does not.
The compute stack is not yet where Bell or ARPANET was at its equivalent moment. Frontier training runs are extraordinarily capital-intensive. The payback window for any individual model is measured in months, not years, before a successor depreciates it. Enterprise customers are getting disciplined about token spend — as recent coverage of U.S. companies walking away from leading-edge API pricing has documented, the open-weight models out of China and the Western open-source community are catching the closed labs in capability at a lag that has been narrowing for two years. The model layer is commodifying faster than the labs would like.
It is true, in the narrow sense in which arguments about financial viability often are, that the AI labs face existential trouble. Nvidia’s stock has slumped. SpaceX’s freshly public shares tanked within weeks of listing. The trillion-dollar valuations OpenAI and Anthropic each filed for in confidential draft S-1s with the SEC in June 2026 — the same filings whose federal-review disposition is still working through the interagency process — were, by almost any structural reading, a fundraising document before they were a market position. The trouble is that “nationalise them if they fail” addresses only the question of who absorbs the wreckage. It does not address the architecture that produced the wreckage.
The term the situation most resembles is John Kenneth Galbraith’s — from The Great Crash, 1929, the magic interval when an embezzler has the money but the victim does not yet feel the loss.
Cory Doctorow has pressed the term into public service for two years now, and he is right, I think, that the bezzle is the most accurate single description of the current AI investment cycle. The models are not yet profitable. The customers are not yet paying enough to cover the inference and training bills. The equity is not yet marked to market at the moment of reckoning. But the bills will come due. The bills always come due. And somebody is going to be holding the bag.
The bag has been pre-stuffed with publicly funded research. The transformer architecture that the leading labs treat as their own was developed inside Google Brain on the back of Stanford-trained talent whose graduate education was federally subsidised. The DARPA-grant-to-nonprofit-laboratory, corporate-spinout, tech-giant-acquisition pipeline that produced Siri — SRI’s CALO program under a $150 million DARPA contract from 2003 to 2008, spun out, acquired by Apple in April 2010, integrated into iOS in October 2011 — is the canonical extractive pattern. The Katalin Karikó and Drew Weissman mRNA platform work that won the 2023 Nobel is the same pattern in biomedical research: publicly funded for a decade at Penn, commercialised through the entities that could afford to license it. The AI labs are running this pattern faster than any preceding wave, and the bill is correspondingly larger.
Cui bono — the question the framework forces every analyst to ask — is straightforward. The value produced by twenty-odd years of public investment in the computational, mathematical, and biological substrates of machine learning has flowed to NVIDIA shareholders on a capex binge funded by depreciating GPU lifecycles; to commercial real-estate firms leasing warehouse-scale data centres near cheap power and population centres; to utilities issuing ratepayer-backed bonds for the Microsoft–Constellation twenty-year power purchase agreement that restarted Three Mile Island Unit 1; to founders with equity stakes structured to cash out at any narrative-supportable moment; and to a thin layer of well-remunerated research scientists whose salaries are the only labour cost visible to a casual reader. The labour that produced the underlying science was performed by graduate students, postdocs, and publicly salaried professors and will not be compensated if the bezzle collapses. That is the architecture any nationalisation inheriting the wreckage acquires.
In Doctorow’s chokepoint-capitalism framework — and the corpus is now long enough that the assumption seems fair — this is the chokepoint operation in its 2026 form: a two-sided market, lock-in on both sides, monopsonistic purchasing power on the supplier side, Most Favored Nation clauses embedded in enterprise contracts, and a bezzle sustained by an “AGI” narrative that lets the labs justify whatever capital expenditure the moment requires. None of this is hidden. All of it is documented. The argument is about what to do about it, not whether it is happening.
I want to be careful here, because the Schneier–Sanders proposal has real merit and I do not want to be unfair to two writers whose analysis I usually learn from. The distinction they draw between product innovation (which they would put under something like national-lab operation) and compute infrastructure (which they would treat as a regulated public utility) is genuinely useful. The list of precedents they cite is not made up. Switzerland’s national AI capability, Spain’s Barcelona Supercomputing Centre extensions, and Singapore’s AI Singapore exist as public instruments. The CIFAR Pan-Canadian AI Strategy — the network anchored at Mila in Montreal, the Vector Institute in Toronto, and the Alberta Machine Intelligence Institute in Edmonton — is the closest thing in North America to the model they are gesturing at, and it has been running since 2017 on something close to public-good logic, with deliberately non-commercialised research outputs and an explicit openness norm.
The trouble is that those public instruments were built as public instruments. They were not built as failed private firms with failed private equity structures that the public then inherited. CIFAR is the right institutional type, but its roughly one hundred million dollars a year is two orders of magnitude smaller than the capex bill the AI labs have been running, and the political economy of standing up a US equivalent at the relevant scale — sustained appropriations, insulation from procurement-cycle pressure, a research-firing cadence that holds the talent — is the real work, not the structural diagram. The argument for also building public compute capacity in the United States along the lines the Schneier/Sanders piece describes is reasonable. The argument for nationalising the wreckage of the OpenAI–Anthropic equity structure in order to do so is the same logic that produced the wreckage.
There is a quieter trap in the proposal they do not name. The phrase “nationalise if they fail” is structurally indistinguishable from “bail them out if they fail,” because the assets that survive the failure — the model weights, the contracts, the talent, the data-centre leases — are the assets the public would acquire, and any nationalisation in the moment of collapse is necessarily a purchase from the equity holders whose bet did not work. The same equity holders whose preferred-share-and-ratchet structure the federal-review apparatus is still working through, and whose employee-organised leverage over Project Maven-style defence contracting Alphabet quietly eroded after 2018 to prevent another revolt. The federal government is famously good at paying retail for assets at wholesale moments. The revolving door between Alphabet, Meta, Microsoft, OpenAI, Anthropic, and the federal AI policy apparatus would administer any nationalised lab — and the composition of the National AI Advisory Committee, chartered under the 2020 National AI Initiative Act and including industry leadership from across the leading AI firms, is the same architecture at slightly higher resolution. The cost, as a different community once put it about a different industry, would be borne by the people the transition was supposed to help.
The public-ownership argument only holds, however, if it identifies what the public should actually own. A flagship model at the API layer is increasingly substitutable: GPT-5, Claude Opus 4, Gemini 2, and the top-tier open weights do roughly the same things for roughly the same price on the workloads most enterprises actually deploy. That is the commodification claim, and it is real at the inference margin. What it omits is the deployment stack: the agentic orchestration, the tool-use loops, the eval and alignment work, the integration with customer systems, the routing between models by capability and cost, the proprietary data flywheels, the inference-optimisation software one layer below the API the customer sees. That stack is what the hyperscalers and the frontier labs are vertically integrating into. It is also where the rents, when they arrive, will arrive. The hyperscalers’ acquisition pattern — buying the inference-optimisation layer to narrow the frontier-lab moat — is not in the past tense; it is the present tense, and it is the thing the public-ownership question is actually about. The model is the pretext. The stack is the asset. Public ownership has to target the stack: model weights, training pipelines, evaluation harnesses, the inference-optimisation layer, and the compute that runs it.
Which brings the column to the institutional-design work that has to be done now, before the founding, because the founding is the only moment at which it gets done. The Schneier–Sanders proposal gestures at this when it specifies “democratic oversight” and “alignment to democratic values,” but the institutional mechanism they describe — congressional oversight, civil-service compensation, the national-lab template — is closer to the postwar arrangement than to anything more recent, and the postwar arrangement gave us Los Alamos. The public-successor institution needs to be designed, on purpose, with the public-interest mandate baked into its charter in enforceable form, and not left to the discretion of whatever executive branch happens to be in power at the moment the question becomes live. The CRTC on telecom is one model of how a public regulator can hold a piece of public-interest infrastructure accountable over decades — and, more honestly, one model of how a public regulator can hold a piece of public-interest infrastructure mostly accountable, with the failure modes legible.
The Harold Innis frame is the political-economy lens worth bringing to the same question: frontier AI is the 21st-century extractive staple whose rents ought to be captured publicly, the way the 19th-century Canadian staple economy was supposed to capture the rents of the fur trade and the timber trade and the railways, and mostly did not. The institutional failure mode is the same.
Two design requirements need to be baked into that institution at founding, not added later.
The first is that public ownership does not automatically mean democratic governance. The history of public-interest science in the United States runs through multiple institutional forms: Bell Labs operating under the regulated-monopoly regime that made its transistor patents available royalty-free; the ARPA-funded university contracts that produced ARPANET and therefore the internet; the NIH-funded university research that produced the foundational mRNA work; and the postwar national labs, which produced the foundational research for nuclear power and a great deal of physics. All of these were public-interest missions operating under various institutional forms. Running alongside them are the security-state national labs, which produced the bomb, the surveillance state, and a great deal of what the civil-libertarian community has spent fifty years trying to dismantle. Both of those are public. The question of which kind of public is what gets decided at the founding, and the moment of founding is precisely the moment at which the founding documents and the founding charters matter most.
The second is that the comp structure is the founding document. The 2026 frontier-lab pay premium — top-quartile AI engineers at OpenAI pulling total comp above $1 million, with the bulk in equity — has done structural damage to the public-sector science workforce that no single appropriations cycle in the last forty years has matched. A public successor that pays civil-service scale will haemorrhage talent to whatever private labs survive the bubble. A public successor that pays frontier-lab scale will replicate the compensation pathology that produced the equity story in the first place. The institutional-design problem is to write a compensation regime that is high enough to retain the people the institution needs and low enough that the institution is not, within a decade, captured by the same equity expectations it was built to displace. There is no clean precedent for this. It has to be invented, and it has to be invented at the founding, because a compensation regime is the easiest thing in the world to amend and the hardest thing in the world to amend back.
So — what would the public actually do, given the chance? Not inherit the wreckage, and not pretend the architectural diagram Schneier and Sanders sketch is the same as the political economy of building it. Restore the pre-restructuring governance of the AI labs that once had nonprofit, mission-bound forms — OpenAI’s pre-2019 charter, Anthropic’s pre-2021 public-benefit framing — and consolidate their research outputs under a public foundation that cannot issue equity and cannot be acquired, so that the output of the existing labs becomes the input to a new public institution rather than an inheritance the public purchases from the old one. Route the next round of frontier AI research through instruments that are already structured to handle it: the NSF National AI Research Institutes network at twenty-five-plus institutes and roughly twenty million dollars each over five years, the Department of Energy national-lab complex already operating the kind of large-scale scientific computing the AI compute ambitions resemble, more of the CIFAR model at several times its current scale. Build the public compute utility, the genuinely good idea in the Schneier/Sanders proposal, as a civilian supercomputing capability — modelled on NSF ACCESS, the DOE INCITE program, and the National Energy Research Scientific Computing Center — rather than as a nationalisation of assets a private equity structure built up to a fractional share of the global GPU fleet. The compute utility is a real idea and worth building. The compute utility the Schneier/Sanders proposal envisions is a derivative instrument: a public balance sheet underwriting an equity structure that did not work.
There will be an appropriations item in the 2027 budget cycle for whatever the rebranding committee settles on — AI Safety Institute, AI Compute Initiative, National AI Lab, the precise name is less important than the institutional form. That item will be the founding. NSF or Los Alamos is what gets decided there, and it is what gets decided now. Vannevar Bush made the same argument in Science, The Endless Frontier (July 1945) — that when wartime research outgrew its wartime mission, the successor institution had to be designed on purpose, in public, before the moment closed. The deadline, as the regulatory agencies have always known, is the only part of the regulatory process that the agencies actually respect. The 1950 answer was NSF. The 1980s answer was DARPA’s commercialisation conveyor and Bayh–Dole. The 2027 answer could be NSF. It could also be Los Alamos. The difference is what gets decided at the founding, and the founding is now. Gravity finds Wile E. Coyote on a Tuesday afternoon. The work, when he falls, is to be standing under a competent institution — built from the public research the public already paid for, not from the wreckage of an equity bet that did not work.