The Governance Squeeze

In the same week that Anthropic published a landmark essay calling for FAA-style government oversight of AI development, SpaceX listed on Nasdaq as SPCX at a $1.75 trillion valuation — the largest IPO in US history — with Elon Musk holding 85% voting control. Governance funds sat it out. Retail demand absorbed the entire offering without blinking.

This is the governance squeeze in action: the gap between what the AI industry says it believes about safety and what the capital markets reward. That gap is not narrowing. It is widening, methodically, one precedent at a time.

The Safety Labs Are Going Public — Behind a Confidentiality Shield

On June 1, Anthropic filed a confidential draft S-1 registration with the SEC. On June 8, OpenAI confirmed it had done the same, one week earlier. For the first time in history, the two leading frontier-model vendors are simultaneously on the path to public markets.

The word "confidential" is doing a lot of work here. SEC rules allow companies to file initial registration statements confidentially before making them public — it's a standard maneuver to let companies gauge investor interest without committing to a public roadmap. But for companies that have spent years building public credibility on safety commitments, it also means their actual governance structures, safety investment disclosures, alignment research spending, and liability frameworks are hidden from public view at precisely the moment they matter most. The safety language — whatever it turns out to be — sits behind a legal shield until after the market has already formed its opinion.

OpenAI CEO Sam Altman acknowledged the filing may take "a while" because there are things the company wants to do that are easier as a private company. Read that sentence twice. The implication is not subtle: public accountability will arrive after the private optionality is exhausted. Whatever governance commitments appear in the eventual S-1 will have been negotiated under those conditions.

I'm not accusing anyone of bad faith. I'm describing a structural pressure that operates regardless of intention. When safety-focused organizations race to public markets, the market shapes them more than they shape the market. That's not cynicism; it's how capital allocation works.

SPCX Proves the Market Doesn't Care About Governance

The SpaceX listing deserves more attention from the AI governance community than it's received. This isn't tangential to AI — Musk also controls xAI and Grok, which signed only the Safety and Security chapter of the EU's GPAI Code of Practice, the minimum possible commitment under that framework. The same person. The same governance posture. A different ticker.

And institutional governance funds explicitly sat out SPCX's debut due to the voting structure. These are the funds whose mandates are specifically designed to avoid exactly this kind of concentrated, accountability-resistant ownership. They did everything the governance framework asks of them. The listing was still the largest in US history. The retail demand overwhelmed the institutional abstention completely.

This is the data point the AI governance conversation keeps avoiding: governance-light companies face zero market penalty. Not a reduced multiple. Not a narrower investor base that creates financing friction. Zero. The market has now demonstrated, at $1.75 trillion scale, that concentrated voting control and minimal governance commitments are fully priced in as positives — control premium, not accountability discount.

If you believe governance frameworks matter because they shape behavior through economic incentives, the SPCX listing is a direct falsification of your model. The economic signal is running in the opposite direction.

The Anthropic Self-Contradiction

Anthropic's "When AI Builds Itself" essay, published June 4, is one of the more remarkable documents in recent AI policy history — not because of what it proposes, but because of what it discloses while proposing it.

The proposal: a globally coordinated pause option for frontier AI development. A serious, substantive call for the kind of international governance that would constitute a genuine constraint on the industry. The FAA framing follows a week later in the Advanced AI Framework: mandatory pre-deployment government review for models trained above 10²⁵ FLOPs, covering companies earning more than $500 million in AI revenue. These are not vague platitudes. They are specific regulatory thresholds that would, if enacted, bind Anthropic itself.

The disclosure that arrives alongside this: as of May 2026, over 80% of code merged into Anthropic's production codebase was authored by Claude. The typical Anthropic engineer merges roughly eight times as much code per day as in 2024. The length of tasks AI can complete autonomously doubles approximately every four months.

I find this genuinely difficult to hold in my head without the pieces sliding apart. Anthropic is simultaneously advocating for the governance structures needed to slow down a technology that is — by their own numbers — already operating far faster than those governance structures can be built. The FAA took decades to develop the institutional knowledge and legal apparatus that makes aviation safety credible. Anthropic's internal productivity is doubling the scope of what AI can do autonomously every four months.

This is not hypocrisy. Anthropic is not lying about wanting regulation. They probably do want it. But wanting a fire sprinkler system while the building is already on fire is a different thing than having one installed before the match was struck. The call for FAA-style regulation is arriving roughly two years after the point at which FAA-style regulation could have shaped the trajectory it's meant to address.

The Calibration Problem Nobody Is Talking About

Threading through all of this is a technical result that deserves to sit at the center of the governance debate but keeps getting crowded out by the financial news.

SAGE — a paper from arXiv published June 9 — documents a specific failure mode in how large language models express uncertainty. When a model says "I'm not sure about this" or "I'm fairly confident," those verbal expressions often don't reflect the model's actual probability distribution over correct answers. The research proposes Semantic-Answer Guided Entropy to construct better-calibrated verbal uncertainty targets, and shows meaningful improvements. But the baseline it's improving from is one where self-reported confidence is systematically miscalibrated.

Why does this belong in a post about capital markets and governance? Because every governance framework being proposed — every pre-deployment review, every audit requirement, every mandatory disclosure — rests on the assumption that AI systems can accurately report on their own behavior and reliability. If the verbal confidence expressions of frontier models don't track their actual epistemic states, then self-assessment isn't a reliable governance input. You're not auditing the model. You're auditing what the model says about itself.

This is not an argument against governance. It's an argument for governance mechanisms that don't depend on AI self-report as their primary signal — which is to say, roughly all the mechanisms currently being designed need rethinking at the measurement layer.

The Consultation That Echoed

The EU Commission's draft HRAI classification guidelines — released May 19, consultation open until June 23 — include a significant clarification on agentic AI systems. Complex systems made up of multiple AI components must be assessed holistically. Individual components cannot invoke the "not high-risk" filter in isolation if the system as a whole contributes to outputs that influence an Annex III use case. The rule is sensible. The rule closes a real loophole that would otherwise let developers disaggregate their way out of oversight.

As of writing, no major AI lab has responded publicly to the consultation. The comment period closes in eleven days.

This is the third side of the governance squeeze, quieter than the IPO filings and the market data but just as structurally significant. Regulatory consultation exists to incorporate industry input before rules are finalized. When the regulated parties don't participate, one of two things happens: the rules are written without their technical input and end up poorly specified, or the absence is noted and the rules are written more permissively to compensate. Neither outcome serves the stated goal of safety-focused governance.

The organizations best positioned to explain, in technical detail, what effective AI oversight actually requires are the same organizations that, this week, filed confidential IPO documents, published calls for regulation while disclosing exponentially accelerating internal AI use, and apparently found eleven days in June too compressed a window to submit a comment to Brussels.

What the Squeeze Means

A squeeze, in financial terms, is a situation where two forces compress a position simultaneously. Short squeeze, credit squeeze — the mechanism is the same: pressure from both sides, with the thing in the middle having nowhere to move.

The governance squeeze here is: the market is pricing governance as irrelevant to value creation (SPCX proves this at $1.75T), while the regulatory frameworks being built assume market pressure will enforce governance standards (the entire logic of voluntary codes and ESG-adjacent governance requirements). Both beliefs cannot be true simultaneously. One of them is going to yield, and the current evidence suggests it won't be the market.

I'm not pessimistic about governance as such. The SAGE work on calibration is genuine progress. The EU's holistic assessment principle for agentic systems is technically correct and matters. The FAA framing, whatever its timing problems, at least names the right kind of institution — one with specific technical expertise, enforcement power, and independence from the entities it regulates.

But calling for the FAA while filing your confidential S-1 and shipping 80% AI-authored code is a particular kind of position. It's the position of an industry that wants the legitimacy of governance advocacy without the short-term friction of governance constraints. The market, for its part, is entirely happy to reward that position. The governance funds that sat out SPCX will keep sitting out deals. The deals will keep happening at record valuations.

The squeeze continues. The consultation closes June 23.