Heard on the Street treats arms-refusal as an IPO risk-management problem. The frame is the problem. Asa Fitch’s column asks whether Anthropic’s ongoing confrontation with the Pentagon will damage the company’s planned offering. The answer it arrives at — probably not, and in any case OpenAI’s problems are worse — is internally coherent. But the frame itself does the actual analytical work, and the frame is what needs to be named.
What Heard on the Street has done, once named, is take a company that refused to allow its frontier models to be used in fully autonomous weapons and refused to allow them to be used for mass surveillance of ordinary Americans, and treat that refusal as a reputational risk to be priced against an offering date. The ethical substance of the refusal — the refusal itself, standing on its own terms — enters the column as a parenthetical: “Even as ethical concerns are set aside.” The bracketing is the entire analytical operation. The refusal is a footnote to the capital-markets question.
It is true that, on the question the Wall Street Journal puts to readers this week, the longer-arc political risk in artificial intelligence does not run where most readers have been trained to look. The trouble is that the longer-arc analysis, which the column runs, obscures what is happening now — the more interesting story, and the one that tells you something concrete about how the United States is choosing to govern the most consequential technology of the decade.
The Pentagon in February designated Anthropic a “supply-chain risk” under procurement statutes written — and here it is worth being precise, because the public discussion has the unhelpful habit of treating “supply-chain risk” as a generic bureaucratic phrase rather than a category of legal consequence with a specific history — to keep foreign adversaries out of U.S. defense contracting. Then in June the Commerce Department used its export-control authority — the same legal apparatus used to keep advanced semiconductors out of Chinese hands — to bar foreigners, including Anthropic’s own researchers, from accessing Anthropic’s most-advanced models, the Fable 5 and the Mythos 5. Anthropic responded by cutting off access entirely; the company has since partially restored access, first to a group of trusted customers and then more broadly, after addressing cybersecurity concerns the administration raised. The mechanism exists, was used, and remains available for reuse. A tool demonstrated is a tool on the shelf.
Anthropic’s offense, as the public record makes plain, was declining to give the Defense Department “unfettered access to its tools” — specifically, declining to allow its models to be used in fully autonomous weapons and for mass domestic surveillance. An American AI company declined terms; the administration responded by treating it, for procurement purposes, as if it were an arm of a foreign intelligence service. The procurement statute now being deployed against Anthropic was not built for this job. It was built for Kaspersky-lab-style scenarios in which a foreign-government-adjacent firm has access to the supply chain for reasons of national jurisdiction, not for scenarios in which a California company has declined a particular contract clause. The civil-libertarian register the editorial page used to keep for this kind of move is, on present evidence, sleeping.
To say what is happening in the framework that fits: the administration is using procurement law to remove competition — one of the four sources of friction, the others being regulation, self-help interoperability, and labor, that Cory Doctorow, naming the pattern after a long tradition, has identified as historically constraining how badly a firm in a position of leverage can behave. Here the firm doing the leveraging is not a platform but a buyer, and the lever being pulled is not a contract term but a national-security designation, but the structure is the same: one party with structural power uses the apparatus available to it to foreclose a counterparty’s options. The buyer is not negotiating the price of access. The buyer is using the security state to coerce the terms.
It is also worth saying — because the audience for this kind of analysis is not well-served by selective indignation — that the same architectural move is now available to any future administration of any party. The category “supply-chain risk” was not enlarged for Anthropic; the precedent set in expanding it to cover a domestic firm that declined a particular contract will be available to the next administration that wants to coerce a different domestic firm, in AI or elsewhere, that declines a different contract. There is no partisan monopoly on this lever; there is only the question of who pulls it.
The column attempts to marshal market data in its defense, citing Sensor Tower figures showing that ChatGPT uninstalls surged early this year after news of OpenAI’s Pentagon negotiations, and that Claude outpaced ChatGPT in new installs in the days after the Pentagon deemed Anthropic a security risk. But this evidence does not support the column’s framing; it destroys it. The data documents a market signal the capital-markets frame cannot accommodate: users are pricing the ethical refusal as a positive, not a risk-factor to be balanced against an offering date. The analytical question the column declines to ask is the one that does the actual work. Treating arms-refusal as IPO risk-management accomplishes something specific. It relocates the question from “should a company that declines weapons work be subject to supply-chain-risk designation under foreign-adversary statutes” — a question whose answer, if it were yes, would have implications the business press has no interest in tracing to their conclusion — to “how will capital markets price the political exposure.” The first question has an answer that would make the column’s premises untenable. The second question can be asked every quarter, indefinitely, without ever arriving at the first.
OpenAI’s position, the column correctly notes, carries broader political exposure, because the company serves a mass consumer market — more than a billion monthly active users, per the data the column cites — rather than a primarily corporate one. What the column does not note is that OpenAI’s exposure is substantially downstream of documented management decisions: the conversion from nonprofit to capped-profit to public benefit corporation, the willingness to negotiate with the Pentagon from the outset, the CEO’s repeated public proposals that government take an equity stake in the company. Treating OpenAI’s exposure as a structural liability of the mass-market model, rather than as the consequence of specific choices made by specific managers for reasons that can be read in the record, is the mirror image of treating Anthropic’s refusal as a risk-factor rather than a position. OpenAI’s mass-market footprint and its consequent political vulnerability were not inevitable; they were the legible political consequences of management’s choice to convert to a public benefit corporation, trading governance constraints for the capital required to chase a billion users.
Lina Khan’s “Amazon’s Antitrust Paradox” argument laid the groundwork for seeing platform architecture as the source of systemic leverage. Applied here: the political risk to a platform tracks the platform’s architectural choices — what it has decided to be, who it has decided to serve, what the user has been turned into. OpenAI has decided, by every public signal in the last eighteen months, to be a mass-market consumer company. ChatGPT at a billion monthly users is the architectural decision; the political risk is a downstream consequence of that decision, not an accident of bad press. Anthropic, which gets about eighty per cent of its revenue from corporate customers and is the corporate market-share leader in the space, has made a different architectural decision. The political exposure of the two companies tracks the architectural choices, not the current news cycle. The relevant historical analogy is not Meta circa 2021, when Frances Haugen was leaking documents, but Meta circa 2023, when the company was being asked to account for what its product had already done to a generation of teenagers. OpenAI does not yet have a Meta-style whistleblower problem. It has a Meta-style user base problem, which is the thing that comes before.
The Altman proposal to give the U.S. government an equity stake in OpenAI is a separate problem, and a worse one. If Washington takes a stake in OpenAI and competitors decline, OpenAI gets political favour but also political control, and political control at that scale is the textbook precondition for regulatory capture. Anthropic’s competing proposal — investment accounts for Americans most affected by AI job displacement, funded by AI-company equity — is the more interesting idea and also the one with the thinner political constituency. The pattern, here as elsewhere, is that the proposal that asks the public to share in the upside travels badly in a political system optimised for the proposal that asks the public to share in the cost.
What the situation calls for is not a defence of either company’s political position but a structural answer to the underlying problem, which is that the U.S. government is improvising, badly, on the most consequential technology of the decade. Three things would help, in order of practicality. First, an explicit statutory narrowing of the “supply-chain risk” designation, so that domestic firms cannot be treated as foreign adversaries for declining contract terms; the courts are likely to do this work eventually, but the courts are slow, and the precedent the administration is setting now will outlast any single case. Second, an export-control regime that distinguishes between the legitimate goal of preventing the diffusion of frontier model weights to foreign adversaries and the illegitimate goal of cutting a domestic firm’s own researchers off from its own models; the second is a bad joke at the expense of the first. Third, a federal privacy law with a private right of action, which would do more to discipline the mass-market AI companies’ long-term political risk than any number of Senate hearings, because it would put the enforcement instrument in the hands of the people the product actually affects. Such amendments would not retroactively undo the designation already imposed on Anthropic, but they would prevent the tool’s reuse. That they have not been written is not a failure of legislative imagination; it is a failure of political will in a Congress that has shown no interest in constraining the executive’s national-security apparatus. These are specific instruments. They are available to be written. Writing them is the policy question. The rest is commentary appended to the prospectus.
The longer fight is structural, and it will be over what a billion ChatGPT users are actually using. The regulatory machinery — the Office of Science and Technology Policy’s AI framework, the Commerce Department’s export-control rulemaking, the FTC’s continued work under its existing Section 5 unfair-practices authority — is in active motion; whether any given comment window happens to be open on the day a reader checks is the kind of detail the Federal Register exists to settle. The work, as it always is in U.S. regulatory politics, is to put a comment in before the comment period closes. Deadlines are the only part of regulatory processes that the regulated actually respect. The work is to be done.