The ledger remembers what the headline forgets.
On September 18, 2026, four paid subscribers of frontier AI services filed Buist et al. v. Anthropic PBC et al. in the Northern District of California, docket number 3:26-cv-10693. The claim is blunt: coordinated public endorsements of AI capability slowdowns constitute an output-restricting cartel under Section 1 of the Sherman Act. The plaintiffs—Charles Buist, Nick Spetsas, Cheyenne Hunt, and Christine Bullock—are not academics. They are consumers who paid for access to models. They argue that when competitors agree to limit the speed or capability of their products, they are restricting output, regardless of whether their stated rationale is safety.
The lawsuit landed six days after Anthropic CEO Dario Amodei published “We Must Pace the Frontier” on September 12. Within hours, Sam Altman, Elon Musk, and Demis Hassabis endorsed the essay. By September 15, OpenAI policy chief Chris Lehane confirmed to TechCrunch that major labs had been coordinating on safety protocols for weeks. Same-day endorsements. Confirmed coordination. A federal antitrust filing. That sequence is not a coincidence. It is a footprint.
Every bug is a footprint left in haste.
The week did not stop with the judiciary. On September 19, President Trump announced the creation of an AI Force and a new AI Czar. His post on Truth Social dismissed AI safety as a “hoax,” vowed that his administration would not “hinder or stifle the Growth of this incredible Industry,” and floated renaming AI to “Superior Intelligence” or “Extreme Intelligence.” No appointee was named. No implementation details were provided. The executive branch moved from oversight to mandate-for-growth.
One week. One antitrust suit. One executive dismissal. One legislative blockade. Seven days did more damage to the AI safety coordination thesis than five years of debate. The question is no longer whether frontier labs can coordinate on safety. The question is whether they can do so without breaking the law, losing political cover, and exposing themselves to existential litigation.
I have spent the past decade auditing decentralized systems where trust is supposed to replace authority. Crypto taught me a simple truth: coordination that depends on goodwill is not coordination; it is an assumption wearing a white hat. AI safety coordination has just been force-migrated from goodwill to adversarial legal review. The code did not change. The environment did.
I. The Antitrust Trap: Why “Safety” Is Not a Defense
Section 1 of the Sherman Act does not exempt altruism. It prohibits agreements that unreasonably restrain trade. The legal categories are familiar: price-fixing, market allocation, output restriction. The plaintiffs in Buist argue that frontier lab agreements to pace development are output restrictions. They have a clean theory. If AI capability is the product, and if the labs collectively decide to restrain that capability, they are collectively deciding to supply less than they could. Safety rationale does not immunize that decision. It arguably documents it.
The public record is catastrophic. Amodei’s essay called for pacing. Altman endorsed. Musk endorsed. Hassabis endorsed. Lehane confirmed weeks of coordination. That is not a hypothetical agreement derived from circumstantial evidence. That is a paper trail. In antitrust law, direct evidence of a horizontal agreement among competitors to limit output is the classic forbidding scenario. The word “safety” does not appear in Section 1.
The labs will argue that capability restraint is not output restriction because safety is a product attribute. That argument is not frivolous. A car manufacturer that agrees to limit top speed for safety is not obviously restricting output in a legally cognizable sense. But the analogy breaks on measurement. Cars have speedometers. AI capabilities do not have a shared, auditable metric. There is no hash of intelligence. There is no verifiable ledger of model behavior. There is only the publicly articulated intention to pace.
That absence is the core forensic problem. In blockchain, we say pics are noise; the hash is the identity. Here, there is no hash. There is no neutral oracle to determine whether a lab actually slowed down. There are only press releases, endorsements, and coordination calls later confirmed by a senior policy executive. The plaintiffs do not need to prove that output actually fell. They need to prove an agreement to restrict output. They have public statements. The intent is on the record.
The deeper issue is that the labs built the metadata of a cartel without the discipline of a cartel. Real cartels do not publish manifestos. They leave no signatures. These labs published coordinated manifestos and then, apparently, acted on them for weeks. That is not a bug in the safety thesis. It is a design feature that became a liability. The same transparency that built public trust for safety coordination built the plaintiffs’ complaint.
The legal defense will be complicated by the absence of any formal safe harbor. The labs could argue that their coordination was precompetitive because it improved safety, but that argument requires an antitrust rule-of-reason analysis. That analysis is expensive, discovery-heavy, and politically vulnerable. It requires the court to accept that making AI safer is a legitimate goal. The current administration disagrees. The political environment is hostile. The executive branch has explicitly called safety a hoax.
This is where the antitrust trap becomes existential. The labs cannot simply abandon coordination. Their entire reputational and regulatory positioning is built on the claim that they take safety seriously. If they abandon pacing to avoid litigation, they signal that safety is subordinate to the share price. If they continue pacing, they face discovery into every internal safety meeting. Every email about deployment timing becomes evidence. Every pause becomes a smoking gun. The phrase “safety pause” becomes a legal artifact.
Silence in the code speaks louder than the pitch. But there is no silence in a Slack archive.
II. Executive Attrition: The Presidency Becomes a Counterparty to Risk
The Trump administration’s creation of an AI Force is less a policy program than a declaration of war on caution. The President’s language was dismissive: safety is a “hoax.” The industry should not be hindered. The name itself should be changed. This is not the language of a regulator. It is the language of a promoter.
The executive shift did not come from nowhere. David Sacks resigned as AI czar in March 2026 after reaching the 130-day special government employee limit. That limit is a structural asterisk. The position of AI czar has now been vacant, filled, and recreated in a period of less than eighteen months. There is no stability. There is no institutional memory. There is only the current occupant’s disposition toward growth.
The new AI Force is a mandate, not an oversight mechanism. It tells the industry to go faster. It does not ask what breaks at speed. It therefore removes the political cover that safety coordination relied on. Previously, labs could claim that they were aligning with government expectations. Now they are explicitly diverging from them. The executive branch will treat any slowdown as a failure to comply with the national growth agenda.
For the labs, this is a strange inversion. In the early days of AI governance, the fear was that government would over-regulate. That fear always seemed premature. The more realistic risk was that government would weaponize its own mandate to break coordination. The federal government does not need to pass a law to harm safety. It can simply withdraw legitimacy. It can call the responsible actors enemies of growth. It can create a litigation environment where every safety decision is interpreted as a restraint on trade.
The executive dismissal also complicates the FINRA-style safety body. That body, which Anthropic and Google are reportedly building with possible OpenAI participation, needs regulatory cover to operate. If its purpose is to coordinate safety standards among competitors, it faces the same Section 1 problem as public endorsements. Actually, it is worse. A formal body with shared decision-making creates a visible algorithm for collusion. The minutes of its meetings become a roadmap for plaintiffs. The executive branch will not protect it. The executive branch is cheering for its destruction.
History is not written; it is indexed. The index for AI safety governance now includes a presidential declaration that safety is a hoax.
III. The Legislative Blockade: No Exemption, No Safe Harbor, No Legal Architecture
The legislative path closed on September 15, three days before the lawsuit was filed. Senator Josh Hawley, supported by Senator Ted Cruz, blocked a national-security antitrust exemption in the NDAA manager’s package. Hawley’s statement was categorical: “No antitrust exemptions for AI. Not a chance.”
That exemption was not a minor procedural detail. It was the legal window Amodei had identified as a necessary precondition for industry-wide safety cooperation. Without it, any formal coordination among competitors is exposed to antitrust liability. With it, the labs could coordinate on safety standards without triggering Sherman Act exposure. The exemption was the load-bearing wall of the whole coordination thesis. Hawley knocked it out.
The irony is severe. Hawley is often framed as a critic of Big Tech. By blocking the exemption, he has ensured that Big Tech AI labs either fracture into unilateral action or coordinate privately with no legal cover. Private coordination is exactly the conduct that attracts the most aggressive antitrust enforcement. Public coordination can be scrutinized and, if necessary, defended. Private coordination is treated as conspiracy until proven otherwise. The legislative blockade did not prevent coordination. It pushed coordination into the shadows. That is a deterministic path to the discovery room.
The Cruz-Klobuchar-Thune duty-of-care bill still exists. That bill would impose a legal standard on AI developers. But it does not create a safe harbor. It creates liability. In a hostile political environment, that bill is not a solution. It is an additional tool for attack. If the duty of care is defined too loosely, it gives plaintiffs another theory. If it is defined too narrowly, it gives the executive branch a reason to veto. Legislation cannot be a substitute for a regulatory exemption. The exemption is the only mechanism that converts coordination from crime to compliance.
Congress has now delivered a triple failure. It refused to exempt safety coordination. It has not passed a comprehensive AI governance law. And it has allowed the existing National Security exemption process to be weaponized against the industry. The result is a legislative vacuum where no lawful path to coordination exists.
IV. The Governance Gap: Not Legal, Not Political, but Structural
The week of September 12-19, 2026, shredded the three foundational assumptions of AI safety coordination.
First, the labs assumed that competitors can coordinate on capability restraint without violating antitrust law. The Buist complaint challenges that assumption directly. It argues that coordination is output restriction. The assumption is now a litigation risk.
Second, the labs assumed that the federal government would enable or tolerate such coordination. The Trump AI Force explicitly rejects that assumption. The president has called safety a hoax and promised no hindrance. The government is no longer a neutral platform for safety. It is an acceleration machine.
Third, the labs assumed that Congress would provide the legal framework to formalize coordination. The Hawley-Cruz blockade forecloses that path. The exemption that would have created a formal framework is dead. The duty-of-care bill remains, but it cannot substitute for a safe harbor.
All three assumptions collapsed in one week. What remains is a liability vacuum. No actor—not the labs, not the executive, not Congress—has the authority and the will to set the terms of safe frontier development.
The Amodei pacing framework proposed voluntary deceleration as a governance model. The antitrust suit argues that voluntary deceleration among competitors is illegal. The executive branch says deceleration is unnecessary. Congress has blocked the mechanism that would have made it legal. The four corners of the state have each pointed in a different direction. There is no center.
This is the same structural failure I have watched in decentralized finance. In crypto, every high-yield protocol eventually discovers that the economic model is an assumption, not an invariant. The same is true here. The “safety coordination” thesis was a governance model that assumed goodwill and legislative cover. The cover is gone. The goodwill is now a legal exhibit. The model must be rebuilt from first principles.
V. What Survives: Capital Pressure, Institutional Momentum, and the No-IPO Signal
It would be a mistake to write the obituary too quickly. The institutional momentum around AI safety has not vanished. The FINRA-style safety body is still being built. Anthropic’s reported IPO deliberations continue. The capital pressure to show growth is real. But survival and functionality are different things.
The FINRA-style body is the most important survivor. It is a formal institution with a specific mandate. If it is structured properly, it could provide a legitimate venue for coordination. FINRA itself is a self-regulatory organization empowered by statute. The AI version would need a statutory basis or at least an unambiguous regulatory blessing. Neither exists today. Without that basis, the body is not a safe harbor. It is a honeypot for antitrust discovery.
The capital pressure is real, but it is not uniformly directional. OpenAI’s no-IPO stance is the strongest signal that at least one major lab is willing to absorb the capital cost of maintaining safety-first positioning. In the current market environment, going public releases liquidity pressure. Refusing to go public is a hedge against short-term growth expectations. It allows a lab to coordinate without the disclosure regime of a public company. That is a capacity for safety that no other lab has demonstrated.
Yet here is the contradiction: the no-IPO stance protects OpenAI from shareholder pressure, but it does not protect it from antitrust liability. The Buist plaintiffs are not shareholders. They are subscribers. They paid for access to models. They can sue regardless of the capital structure. The legal exposure is about coordination, not incorporation.
What survives is a zombie framework. The institutions exist, but the legal and political substrate that made them viable is gone. They can function as forums, as research bodies, as standard-setters. They cannot function as enforcement mechanisms. Enforcement requires legal authority. Legal authority requires a legislature that is willing to create safe harbor. Congress is not willing. The executive branch is actively hostile. The courts are now in play.
The safety-thesis contradiction I identified in Anthropic’s IPO analysis—safety positioning versus capital pressure—has expanded into something worse. It is not a contradiction. It is a collapse. Safety positioning now contends with legal exposure, political hostility, and legislative gridlock. All four forces arrived in the same seven days.
The map is not the territory; the chain is both. The map of AI safety governance was drawn with a pen. The territory includes subpoenas, discovery motions, and executive orders. The two no longer overlap.
VI. The Contrarian Case: Coordination by Other Means
Let me steelman the bulls. The collapse of top-down coordination may not be the disaster it appears to be.
First, the antitrust lawsuit is not guaranteed to succeed. The plaintiffs must prove that capability restraint is output restriction. That is a novel legal theory. AI capability is not a discrete product with a stable unit of measurement. Courts may decline to impose Sherman Act liability on public safety commitments without evidence of actual output reduction. The labs could also defend coordination as precompetitive. Consumer preferences include safety. Coordinating on safety might increase overall demand for AI, thereby increasing output in a broader sense. That argument is available, and it is not irrational.
Second, the FINRA-style body may become more effective without political baggage. The administration’s hostility forces the body to become a private standard-setting institution rather than a quasi-public regulator. Private standards have their own legitimacy. They are not subject to the vagaries of presidential succession. They can persist through regime change. The lab-led body could become the real governance mechanism precisely because it no longer depends on false government support.
Third, the no-IPO stance may become contagious. If OpenAI demonstrates that a private company can maintain safety credibility while capital markets demand growth, other labs will follow. The legal exposure remains, but the capital pressure drops. That is not nothing. It is the beginning of a different safety architecture: safety through independence.
Fourth, litigation itself may clarify the rules. A judicial opinion on whether safety coordination is an antitrust violation would resolve the ambiguity that currently paralyzes the industry. The labs are afraid to coordinate because they do not know the law. A court decision would tell them. Even an adverse decision is better than permanent uncertainty. The law is the map. The map is missing.
I do not agree with every point in this counterargument. But it deserves respect. The coordination collapse does not automatically equal a safety collapse. The labs may adapt. The question is whether they adapt into a lawful framework or into a shadow network.
The evidence so far is not encouraging. Lehane’s confirmation of weeks of coordination proves that the labs were comfortable operating without explicit legal cover. That comfort is a liability. The only safe response to uncertainty is unilateral action. Unilateral action is the opposite of coordination. It is the end of the shared pacing framework.
The legal gray zone is not a neutral space. It rewards secrecy. Secrecy destroys trust. Trust is the only real infrastructure of safety coordination. The chain does not forgive.
VII. The Accountability Call
The frontier of AI safety is no longer about model capabilities. It is about legal architecture. The labs cannot solve this with another safety paper. They cannot solve it with a red team report or a public commitment. They need a safe harbor that is enforceable and politically protected. They need Congress to act, or they need a court to bless a particular form of coordination, or they need to restructure themselves so that coordination is not legally visible.
The hardest solution is the only honest one. The industry must move from voluntary coordination to institutionalized, legally binding governance. That requires a new kind of contract: not a compact among competitors, but a regulatory compact between the industry and the state. The state has to grant the exemption. The state has to create the body. The state has to enforce the rules.
That compact does not exist. It was blocked by Hawley and Cruz. It was dismissed by the President. It is being litigated by four paid subscribers. The week of September 12-19, 2026, did not just damage the coordination thesis. It revealed that the thesis was never a thesis. It was an assumption floating on good intentions.
Precision is the only apology the chain accepts. There is no precision in the current AI governance discourse. There is no hash, no ledger, no verifiable commitment. There are only endorsements, subpoenas, and an executive branch that calls safety a hoax.
The challenge for the next year is not to sustain coordination under the old assumptions. It is to invent a new governance primitive. That primitive must assign liability, grant safe harbor, and create an auditable record of restraint. Without it, the industry will oscillate between catastrophic speed and paranoid silence. The market will do what it always does: optimize for what is measurable. Without a measure of safety, the only measurable thing is speed.
The ledger remembers what the headline forgets. The headline is “AI Safety Coordination Collapses.” The ledger is seven days of legal decisions, political appointments, and legislative vetoes. The ledger is unforgiving. The next entry will be written by the courts, or by a Congress too late to matter. The industry should not wait for that entry. It should build the safe harbor before the next week arrives. Weeks like this one do not forgive in proportion to preparation. They forgive not at all.