AI IPOs force a new bargain between the state and labs
Anthropic and OpenAI’s IPO filings, after abrupt U.S. restrictions and a partial restoration, expose the unstable politics of treating AI labs as strategic assets.
In a nutshell
- The AI labs’ IPO plans expose the fragile politics of state control
- Public listings force political risk into prospectuses and market pricing
- As AI commoditizes, the most durable revenue shifts to government buyers
- For comprehensive insights, tune into our AI-powered podcast here
The five weeks spanning June and early July 2026 told a story. On June 1, San Francisco-based artificial intelligence developer Anthropic confidentially submitted a draft registration statement to the United States Securities and Exchange Commission for a proposed initial public offering (IPO) after a private funding round that reportedly valued the company at nearly $1 trillion. A week later, its crosstown rival OpenAI did the same. On June 12, a Commerce Department directive ordered Anthropic to suspend foreign-national access to Fable 5 and Mythos 5 models, its latest frontier AI systems. Anthropic responded by disabling the models for all customers.
On June 19, U.S. President Donald Trump said he no longer considered Anthropic a national security threat. By June 30, restrictions had eased: Fable 5 returned to broad availability, while Mythos 5 was restored only to approved U.S.-based organizations. Two days later, the Financial Times reported that OpenAI had discussed giving the U.S. government a 5 percent stake in the company.
The sequence compressed the new political economy of frontier AI into a few weeks: filing, coercion, absolution, partial restoration and equity settlement. Washington has begun treating the leading AI labs as strategically critical institutions.
Yet once those labs enter public markets, this informal order becomes too unstable to sustain. Public investors can tolerate risk, but they require it to be disclosed, priced and governed. The AI IPOs will therefore force a settlement that Washington has so far avoided: The state will retain decisive power over frontier AI, but the terms of that power will have to be formalized.
Capture without a charter
The traditional concept of regulatory capture is beginning to reverse itself in frontier AI. The classic fear was that powerful firms would bend public institutions to private ends. In AI, the pressure increasingly runs in the other direction. As models become central to military capability, intelligence work, cyber defense, economic competitiveness and public administration, governments are asserting final authority over firms originally built as private technology companies.
The Anthropic suspension showed that AI does not fit existing categories. A government directive removed access to flagship commercial products overnight. The company received neither a procurement dispute nor a contract modification. It received an order. That order applied to a technology whose civilian and strategic uses cannot be cleanly separated. Unlike aircraft, satellites or weapons platforms, these systems were developed by venture-backed companies, distributed through software interfaces and sold globally before the state had built a governing architecture around them.
Washington now seeks the authority of a defense customer in a market that developed outside the defense-contract system. When the U.S. cancels or redirects a major defense program, contracts define termination rights, cost recovery and dispute procedures. State supremacy is converted into a priced, litigable relationship. The AI labs have no equivalent charter. They are treated as strategic contractors while retaining the exposure of startups.
That asymmetry was easier to manage while the firms remained private. Venture investors who accepted political risk only had to answer to themselves. Public companies are different. Their risks must be disclosed in advance, priced by the market and challenged by shareholders when management appears to tolerate avoidable harm. IPOs shift the problem from policy to securities law and corporate governance.
The commercial squeeze
The timing could hardly be worse, as the business model supporting the labs’ valuations is coming under pressure.
The simplest version of the frontier AI business is selling intelligence by the meter. Customers pay per token – the units of text a model processes – so revenue rises with usage. Near-trillion-dollar valuations assume that metered intelligence will become an enormous and durable market. The problem is that both sides of that market are now pushing against the model.
On the supply side, cheaper open-weight models, many of them Chinese, are narrowing the performance gap and driving prices down. OpenRouter, a leading neutral routing platform, shows a sharp rise in Chinese models’ share of routed usage, with some analyses placing them at a majority of token consumption among the platform’s top models. This trend is not due to superiority; rather, the models offer a drastic cost advantage and are sufficiently effective for most applications. Cost has become a central reason developers and enterprises experiment with alternatives to the leading American systems.
On the demand side, customers are realizing that usage does not automatically translate into value. Uber reportedly burned through its 2026 AI coding budget in four months, while Microsoft began shifting engineers to lower-cost models as token costs rose sharply. Customers are becoming less dazzled by AI and more insistent on outcomes. Palantir’s Alex Karp captured the mood, saying that “something has gone completely wrong” with token-based pricing.
This commercial squeeze connects directly to the political one. As open markets commoditize, the most defensible revenue comes from buyers who cannot simply choose the cheapest model: government, defense and critical-infrastructure customers. Every gate into that segment is held by the state. Washington does not need to become the labs’ largest customer to become their indispensable one.
The prospectus problem
An IPO registration statement must tell investors what could materially harm the business. Some risks are straightforward to describe: Competitors may cut prices, customers may spend less than expected, demand may grow more slowly than projected. Defense contractors also disclose political risk, including the possibility that programs will be canceled, budgets reduced or contracts terminated. But those risks are understandable because they sit inside a defined legal and contractual framework. Investors may not know whether a program will survive, but they know the rules that govern what happens if it does not.
The frontier AI prospectus needs to address something more awkward: the possibility that a government directive may suspend flagship products for reasons only partially disclosed, with no clear path to compensation and no settled procedure for appeal.
Markets have priced politically subordinated firms before. Alibaba and Tencent kept trading despite Beijing’s power over them; investors simply applied a discount. The American setting is less accommodating. A standing “capture discount” would become a target for founders, employees and shareholders, who, unlike their Chinese counterparts, have the legal and political means to attack it. An IPO would force that confrontation into the open: Investors would have to ask whether the labs can credibly be valued as public-market champions while remaining subject to strategic control exercised through improvised government pressure.
Why Washington will bargain
The Trump administration could ignore this contradiction if it did not need the companies to succeed. Frontier AI has become part of the competition with China. Maintaining the frontier requires vast, recurring investment in chips, data centers, energy, talent and model development. OpenAI was reportedly projecting multibillion-dollar losses in 2026, and Anthropic’s latest financing reflects both investor appetite and the strain on private capital. In a contest where China can direct treasury resources toward favored firms, keeping America’s leading AI labs unlistable would weaken one of Washington’s strategic instruments.
The same logic applies to valuation. If state coercion creates a standing political-risk discount, the state itself pays part of the bill. A higher cost of capital means less compute, slower deployment and weaker firms at the frontier. Coercion that seemed free when imposed on private investors becomes expensive once it is capitalized into every future financing round.
Coercion that seemed free when imposed on private investors becomes expensive once it is capitalized into every future financing round.
Public ownership also reshapes the politics of intervention. Suspending a private company’s product hurts venture funds. Doing the same to a trillion-dollar public company destroys value held by index funds, pension funds, retirement accounts and retail investors. Once ordinary savers own the national AI champions, the government has an interest in their stability. An unpriced sovereign override would damage voters’ portfolios as well as corporate balance sheets.
There is also a legitimacy problem. The AI boom cannot appear to enrich only venture capitalists, founders and hyperscalers while the public bears the brunt of job disruption, security risks and energy costs. The Trump administration’s 2025 conversion of federal support into a 10 percent stake in Intel showed that minority public ownership of strategic technology firms has reentered the American policy toolkit. OpenAI’s 5 percent proposal belongs in that context: an attempt to formalize political dependence before investors price it in on harsher terms.
More by innovation expert Uri Gabai
- The battle for AI sovereignty
- From Pax Americana to Pax Silica
- OpenAI, public interest and the problem of tech governance
The settlement now taking shape
Two models are now competing. Under the equity model, the state receives a stake, perhaps through a public wealth vehicle. This model is easy to explain and gives leaders a way to say that citizens share in the AI boom. Its weaknesses are equally clear. A stake leaves unanswered how the state will behave the next time it invokes national security. It also institutionalizes a conflict of interest: The same government that regulates frontier AI, awards its contracts and polices its competition would hold a significant financial position in the firms it oversees. This creates an obvious divide between the labs inside the arrangement and those outside it.
The charter model addresses the operating relationship. It would preserve the state’s right to act on security grounds while making that power procedural and compensable. Security directives would come with notice where possible, defined standards, review, appeal and compensation for state-imposed losses, loosely modeled on termination-for-convenience rules in defense contracting. In return, the labs would accept deeper oversight, procurement commitments and obligations tied to export controls and allied security needs. The charter’s defect is its inheritance. Compensated coercion and anchored government demand are also the recipe for the defense industry’s chronic diseases – cost overruns, lobbying entrenchment and firms optimizing for the contract rather than the market.
So far, the two labs have gravitated toward different templates. OpenAI has discussed giving the state a place in its capital structure, while Anthropic’s June restoration came through negotiated rules and safeguards.
The distinction matters for investors. Equity offers political symbolism: It lets the state claim that citizens share in the upside of frontier AI. A charter does something more practical: It reduces unpriced risk by defining how state power will be exercised, reviewed and compensated. The first makes the bargain visible. The second makes it investable.
Scenarios
Most likely: Formalization of AI regulation and ownership models
The AI listings proceed, but the settlement emerges over 12 to 24 months rather than in a single legislative act. Prospectus-driven pressure, deepening government procurement and eventual statutory rules gradually convert today’s informal control into a more explicit framework. The likely result is a hybrid: a modest public stake as the political symbol, and contractual or statutory rules on notice, review and compensation as the machinery that makes the bargain investable.
Key signposts in this scenario will include the risk-factor language in Anthropic’s IPO, expected around September; indemnification provisions in government AI contracts; and legislative proposals for directive review.
Moderately likely: Priced subordination
Markets absorb the risk at a discount – the American Alibaba path. The labs list below aspiration, capture persists informally and investors price political risk without forcing institutional reform.
The clearest signpost would be valuation compression a listing priced below the roughly $2 trillion figure investors were reportedly discussing as of mid-August, without any real change in the state-firm relationship.
Less likely: AI’s retreat from the public market
Listings slip indefinitely. Sovereign wealth, defense money and hyperscaler balance sheets substitute for public capital, and the settlement forms privately, opaquely, without disclosure discipline – the worst governance outcome of the three scenarios.
The main signposts in this scenario would be OpenAI’s reported preference for a 2027 listing stretching into an even longer delay, and further mega-rounds replacing IPO proceeds.
Contact us today for tailored geopolitical insights and industry-specific advisory services.
Sign up for our newsletter
Receive insights from our experts every week in your inbox.
