On September 18, 2026, a class-action complaint landed in the Northern District of California. Four plaintiffs — Charles Buist, Nick Spetsas, Cheyenne Hunt, and Christine Bullock — are paid subscribers to frontier AI platforms. Their allegation is narrow and mechanical. Anthropic, OpenAI, SpaceXAI, and Google have entered into an unlawful agreement to restrain trade. Not that the models are unsafe. Not that the training data was misappropriated. That four firms agreed to slow down.
The interval that matters is seventy-two hours. On September 12, Anthropic CEO Dario Amodei published "We Must Pace the Frontier," an essay advocating deliberate deceleration in model deployment. That same day, Elon Musk, Sam Altman, and Demis Hassabis publicly endorsed the pacing strategy. On September 15, OpenAI's Chris Lehane confirmed to TechCrunch and Bloomberg that the major labs had been coordinating on safety protocols for weeks. Three days later, the filing.
Check the logs, not the tweets. That instruction is the first thing I give any analyst who joins a desk. It is also the reason this case is difficult. In every market I have analysed professionally, coordination leaves a residue — signed transactions, timestamps, state diffs, nonce sequences. Here, the coordination is documented in an essay and confirmed in interviews. Everything after that is inference dressed as evidence.
Context
The defendants are the four firms best positioned to define what "frontier" means. Anthropic, OpenAI, SpaceXAI, and Google. SpaceXAI's presence on the list is the detail most commentators skipped, and it matters: the coordination described is not purely a model-lab phenomenon. It spans compute, deployment, and distribution.
Section 1 of the Sherman Act prohibits contracts, combinations, or conspiracies in restraint of trade. Two elements must be proven. There must be an agreement — not a parallel outcome, an agreement. And that agreement must unreasonably restrain trade. Courts split these into per se violations, which require no market analysis because the conduct is inherently destructive, and rule-of-reason cases, where plaintiffs must define a market, prove power, and show anticompetitive effect.
Horizontal output restrictions sit near the per se end of that spectrum. When direct competitors agree to limit production volume, the historical treatment is hostile. The plaintiffs are building their case on exactly that frame. If competitors agree to slow capability growth or release cadence, they are, by definition, restricting output. The labs will argue that safety coordination is a legitimate, precompetitive justification — that the agreement improves a product dimension consumers value, namely reliability and risk containment.
That defense has a problem. It requires the court to accept that the same firms setting the safety threshold are the ones being constrained by it. There is no external validator. There is no oracle. The threshold is a number chosen by the parties, published voluntarily, and revocable at will.
Code is law; hype is just noise. A pacing essay is neither.
The political layer arrived first. Days before the filing, Senators Josh Hawley and Ted Cruz blocked a national-security antitrust exemption for AI companies inside the NDAA manager's package. Hawley's line was blunt: "No antitrust exemptions for AI. Not a chance." The labs had sought legislative cover to formalize their safety pacts. They did not get it. White House AI czar David Sacks complicated the narrative further, dismissing the self-regulatory proposals as potential regulatory capture or a distraction timed to the election cycle.
Third-party observers have already reached for historical precedent. Cohere CEO Aidan Gomez, in a recent essay, called the push for a FINRA-style safety body "a cartel by any other name." He drew a direct line to the SEC's 1975 NRSRO designation — the mechanism that created a small, self-selecting set of credit rating agencies — and to Europe's 1985 Motor Vehicle Block Exemption. Both frameworks were justified as consumer protection. Both functioned as entry barriers.
The case sits in the pleading stage. No merits ruling exists. No response from the defendants had been filed at the time of writing.
Core
I have spent the last decade building measurement systems for markets where the coordination is real but the paper trail is absent. That is the actual technical problem here, and it is not a legal problem. It is a data-architecture problem.
Start with the dependent variable. In crypto, the variable was clear: transfer volume, unique active wallets, gas consumed, liquidity depth, slippage under volatility. In frontier AI, "output" is not one number. It is inference tokens served, API concurrency ceilings, model release intervals, benchmark deltas per generation, and the price of a marginal unit of capability. Each of these is measurable in principle. None of them is disclosed on a schedule anyone can audit.
The first analytical move is to stop treating "pacing" as a philosophy and start treating it as a parameter. A pace is a rate. A rate can be plotted. If four labs simultaneously stretch their release intervals from nine months to fifteen, that is not automatically coordination — but it is a testable series, and the counterfactual is available: what were the intervals before the endorsements, and what did each firm's own roadmap say the intervals would be?
I have run this exact analysis before, on a smaller stage. In 2021, during the NFT cycle, I built a wallet-clustering regression to separate genuine collector demand from manufactured volume. The headline number was that roughly 40% of floor price movement traced to bot activity. The lesson generalises. Aggregate price and aggregate output can both be manufactured, and the manufacturing is invisible in the headline series. It only appears when you decompose by actor.
Which brings the analysis to thresholds.
DeFi taught me that a governance-set constant is still an arbitrary number, no matter how elegantly the curve is parameterised. Aave and Compound do not discover interest rates. They choose them — through utilisation curves whose slopes and inflection points are set by token votes and admin proposals. The market supplies the input. The protocol supplies the shape. Nothing about that shape is emergent from supply and demand; it is a design decision wearing a market's clothing.
Safety thresholds in AI are structurally identical. No natural constant says a model must not exceed capability level X on evaluation suite Y. Someone picks the evaluation suite. Someone picks the threshold. Someone decides which release counts as a violation. If the picking is done collectively by the firms being measured, the curve is not a market signal. It is a governance parameter.
I spent four months in 2017 reverse-engineering Groth16 proof verification for early zero-knowledge protocols. Custom Python, constraint by constraint. I submitted three pull requests that reduced gas costs by twelve percent. That work taught me a durable rule: the value of a system lives in its constraints, and constraints are only meaningful if someone independent can verify them. A constraint nobody can verify is a promise. I have never seen a market price a promise correctly for long.
Here is the part the antitrust framing obscures. In DAO governance, the phrase "code is law" collapsed the moment people looked at the upgrade path. The smart contract does not govern itself. Upgrade rights sit with a multi-sig, and the multi-sig is three to nine humans. The immutability was always a thin surface over an administrative core.
The AI pacing framework has the same architecture and less substance. There is no smart contract. There is no multi-sig. There is no on-chain commitment anyone can inspect to verify that a lab actually decelerated. The commitment exists in prose. If a firm quietly accelerates while publicly pacing, nothing in the framework detects it. The pacing pact is a policy without an enforcement layer, which makes it indistinguishable — from the outside — from a public-relations campaign with a legal shield attached.
The Layer 2 market offers the second structural parallel, and it is the one I find most uncomfortable. For three years I have argued that dozens of rollups competing for the same small population of active users is not scaling. It is fragmentation. Liquidity sliced into thinner fragments, each with its own sequencer, its own bridge risk, its own rent extraction. The aggregate number went up. The per-unit depth went down.
Frontier AI is now doing the same thing with the word "frontier." Four labs, plus a queue of well-funded challengers, each claiming to operate at the frontier, each drawing on the same narrow pool of capital, compute, and specialised researchers. When the incumbents agree to a common pace, they are not regulating a race. They are setting the terms of a shared scarcity.
So what would a Section 1 restraint look like if it were measurable? It would look like a gap between output actually produced and output that would have been produced absent the agreement. That is a counterfactual, and counterfactuals are the weakest evidence in any quantitative discipline. Which is precisely why the plaintiffs will lean on the seventy-two-hour sequence and the Lehane confirmation rather than on a model. They have a timeline. They do not have a ledger.
I built a version of the ledger once, for a different market. In 2024 I worked with a boutique quant fund to design an on-chain surveillance dashboard for institutional clients — anomaly detection across smart-money flows in Layer 2 ecosystems, with a 92% accuracy rate on short-term volatility spikes. The pipeline was straightforward: normalise the flow, establish a baseline, flag the deviation, hand it to a human. The hard part was never the detection. It was the baseline.
The same problem holds here. The roughly 85% probability I placed on the Terra/Luna depeg, flagged two weeks ahead of the collapse, came not from sentiment but from oracle dependency mapping. I could see the structure of the fragility before anyone could see the price move. That kind of forecast is possible only when the underlying system emits data continuously and honestly.
Frontier AI emits almost nothing on a fixed schedule. That is the asymmetry that will decide this case — not the rhetoric.
Even without a ledger, labs emit residue. Check the commit graph, not the press tour. Model card revisions carry timestamps. API version deprecation notices carry sunset dates. Pricing pages carry historical snapshots in archive services. Benchmark harnesses carry commit hashes. None of this proves an agreement. All of it establishes a release cadence that can be plotted, compared across firms, and tested against each firm's own prior statements. A cadence shift that is simultaneous across four independent roadmaps is a data point. A cadence shift that is simultaneous and unexplained is a stronger one.
Antitrust law has handled this shape before, in other industries. Trade associations that share information — capacity plans, output forecasts, pricing intentions — have been litigated for a century, and the doctrine is uncomfortable for everyone involved. Sharing safety research is not the problem. Sharing deployment schedules is closer to the line. The distinguishing variable courts have used is whether the exchanged information concerns a competitive variable and whether it reduces uncertainty about rivals' future conduct. A published pacing framework does both. It announces a competitive variable — deployment speed — and it reduces uncertainty about what every other signatory will do.
The plaintiffs also inherit a market-definition problem with no clean crypto analogue. Subscription access to frontier models is bundled, tiered, and priced below marginal cost at the top end. Revenue comes from enterprise contracts and API metering, not from the twenty dollars a month the named plaintiffs pay. That makes the consumer-harm theory harder to quantify than it looks. In every on-chain market I have modelled, harm showed up as slippage — a measurable transfer from one cohort of users to another. Here, the harm is a capability that was not shipped. You cannot compute slippage on a product that does not exist. You can only argue about whether it would have.
Contrarian
Now the honest part, because the tidy version of this story is wrong.
Public endorsements are cheap. Musk, Altman, and Hassabis agreeing with an essay on the same day is suggestive. It is not a conspiracy. Firms inside a shared technological paradigm converge on shared language constantly, and the convergence is frequently independent. Correlated timing is not agreement, and agreement is not restraint. Three of the four named companies had independent reasons to slow down in September 2026, and none of those reasons required a phone call.
The most plausible alternative hypothesis is not collusion at all. It is constraint. Compute procurement has a lead time measured in quarters. High-quality training data is not infinite. Benchmark deltas have been shrinking across generations — a well-documented pattern that predates any pacing framework. Deceleration, in that reading, is not a choice the labs made together. It is a physical limit they encountered at roughly the same moment and then branded as strategy.
There is also a structural blind spot in the complaint itself. The relief the plaintiffs want is, functionally, acceleration. A court order that firms compete harder on speed. If granted, it does not restore a competitive baseline. It substitutes judicial judgement for corporate judgement about capability thresholds, at the exact moment when the technical community is least confident about where those thresholds belong.
And the remedy may be worse than the coordination. If the pacing framework is invalidated, the firms with the strongest safety apparatus are penalised for having had one. The firms that never agreed to anything — including labs in jurisdictions that will not enforce a ruling from the Northern District of California — inherit the market. Antitrust enforcement in a globally distributed industry has a jurisdictional ceiling, and this case is standing directly under it.
Takeaway
What to watch is not the verdict. It is the docket sequence. A motion to dismiss will force the plaintiffs to specify the agreement rather than the outcome, and that specificity — or its absence — will reveal whether the case has a theory of harm or only a timeline. Watch whether the FINRA-style safety body reappears with legislative language attached. Watch whether any lab begins publishing capability thresholds on a fixed cadence with third-party verification, because that single disclosure would defuse both the antitrust exposure and the safety critique at once.
The deeper question is the one nobody in the filing addresses. In a market where the four largest participants jointly define the metric that determines whether they are competing, what exactly is the regulator measuring? Check the logs. If there are none, someone will eventually be asked to build them — and the party asked to build them should not be the party being measured.