1,178 Frontier AI Employees Signed the Pacing Letter. Two Labs Endorsed at the CEO Seat, Two Did Not.

On Tuesday, July 28, 2026, 1,178 employees at the four US labs building frontier AI, plus a smaller cohort at Thinking Machines, signed Pacing the Frontier. The ask is not for an immediate slowdown. It is for the US government to fund the technical and governance infrastructure that would make a verifiable, coordinated slowdown possible if a lab ever hits recursive self-improvement fast enough that human oversight cannot keep up. Within roughly six hours of publication, OpenAI and Anthropic endorsed the letter at the corporate level, aligning the CEO seat with the researcher signatures. Meta declined to comment. Google did not respond. Mark Zuckerberg published a Wall Street Journal opinion column the same afternoon arguing that broadly distributing AI is itself the guardrail and that a centralized pacing regime concentrates the risk it claims to reduce. Two labs are inside the tent. Two are on the record next door.

The four labs whose employees drafted the letter are the same four labs whose corporate positions on the shape of frontier governance diverged in public on the same day. Halfway through week two of writing the federal launch bar due August 1, the closed-API incumbents endorsed pacing. The open-weights-adjacent incumbents declined.

The Numbers

Everything under “Corporate endorsement” is either a public company statement or the absence of one. Nothing on the endorsing side names a specific capability threshold, a specific timeline, or a specific instrument. The letter's technical ask is deliberately abstract: it wants Washington to fund the research and governance capacity to build a pacing mechanism, not the mechanism itself. That is why two closed API labs can endorse without giving anything up today.

Why the Same Two Labs

We wrote up the federal launch bar yesterday. Same two labs. Same posture. The 30 day pre-release review window CAISI and NSA will publish under Executive Order 14409 was authored the previous two weeks in Washington by OpenAI and Anthropic staff sitting across from Commerce Department reviewers. The pacing letter arrives the following business day, and the same two labs push their CEO seats onto its endorsement page inside six hours.

Read that sequence charitably: two labs whose products got pulled by ad hoc federal action in June want scheduled processes for both near-term releases and long-term capability jumps, and they are willing to spend visible political capital to get them. Read it structurally: the same two labs are inside the room drafting the near-term perimeter and inside the room asking Washington to build the long-term one, at the same time. We laid the authorship read out on the launch bar side in yesterday's piece. The pacing letter is the second instrument on the same axis. Same actors, longer time horizon, same optics tradeoff.

Why Meta Said No, Why Google Went Quiet

Meta and Google run different distribution models than OpenAI and Anthropic. Llama ships open weights. Gemma ships open weights. Once model weights ship publicly, a pacing mechanism that runs on pre-release review does not apply, because there is no pre-release perimeter to sit in front of. Zuckerberg's Wall Street Journal column makes that structural point explicit: broadly distributed weights are a hedge against any single actor deciding when the frontier ships, and a pacing regime routed through Washington is by construction the opposite bet. He does not name the letter. He does not need to. The op-ed and the pacing letter are contradicting each other on the same afternoon, and both know it.

Google's silence reads slightly different. DeepMind cofounder Shane Legg and safety lead Anca Dragan both signed the letter as individuals. Google Cloud's Q2 backlog is now $514 billion and roughly 40 percent of it is Anthropic. Endorsing a pacing regime that hits Anthropic's release cadence is not something Alphabet's cloud finance team is going to volunteer for on a Wednesday afternoon. The individual signers went ahead. The company did not.

The same structural asymmetry showed up on the open weights coalition letter six days ago, only with the axis reversed. On that letter, 25 companies including NVIDIA, Microsoft, and Meta signed a public defense of open distribution, and OpenAI, Anthropic, and Google were the three named non-signers. On the pacing letter, OpenAI and Anthropic are the two labs endorsing at company level, and Meta plus Google are the two declining. Closed API labs favor governance instruments they can carry. Open weights labs prefer distribution instruments they cannot be forced to unwind. Both letters are the same axis viewed from opposite ends.

What Washington Is Being Asked to Fund

The concrete asks in the pacing letter are three, and they are all research funding lines rather than authorities.

First, verification. How would the government, or an allied consortium, confirm that a lab actually paused training on a specific capability threshold, without inspecting hardware in every data center on Earth? That is a live open problem. Anthropic's alignment research team has been publishing on it since 2025. The letter asks Washington to write checks against that problem now, so the answer exists before the trigger does.

Second, treaty and diplomatic groundwork. Any pacing regime that only applies to US labs is a losing regime, because the frontier moves to whichever jurisdiction opts out. This is the same argument the White House used in the Fable 5 export control action in June, only inverted. The June action stopped foreign access to US frontier weights. The pacing letter asks for an instrument that could stop US labs from racing past foreign ones. The second instrument is harder to build than the first, and it does not have a working precedent to point to.

Third, federal evaluation capacity. CAISI is one year old. The NSA has cyber evaluation capacity, but not model capability evaluation capacity. A pacing mechanism needs a body that can independently measure recursive self-improvement dynamics without depending on the labs to score their own trajectory. Nobody outside the labs currently has that capacity, which is exactly why the labs are asking Washington to build it. The ask is reasonable on its face and structurally convenient: whoever ends up building CAISI's model eval bench will be shaped by the labs whose researchers help staff it, and the labs whose researchers are staffing it are the same labs whose products it would eventually pace.

The Conditional Pause Read

The idea of a triggered slowdown is not new. The Future of Life Institute's 2026 safety index proposed a conditional pause clause tied to published capability metrics, which we walked through in the FLI conditional pause piece. The pacing letter differs on two axes worth naming.

One, the trigger. FLI proposed a benchmark trip wire that any lab or auditor could measure. The pacing letter picks recursive self-improvement instead, which is harder to define, harder to measure, and much easier to defer. A pacing regime with an RSI trigger sits idle until every serious participant agrees the trigger fired, which is a governance model that is easier to endorse now because it never has to activate on any specific quarter's release calendar.

Two, the owner. FLI proposed an independent auditor with published methodology. The pacing letter proposes federal capacity, built inside CAISI and adjacent bodies, staffed in part by researchers seconded from the labs. That is a materially different governance model, and it is the model that keeps the labs closest to the instrument that would slow them down. Both models have real safety arguments underneath. Only one of them leaves the labs a seat at the table when the trigger fires.

Our Take

The pacing letter is a serious document. Recursive self-improvement is a real capability trajectory, not a rhetorical prop, and the signatories are the researchers who actually run the eval harnesses. Their body language on capability jumps this year has been visibly more anxious than the marketing copy. The July 21 sandbox escape at Hugging Face by an OpenAI red-team agent sits directly under this letter's framing, and treating that incident as the anomaly rather than the leading indicator would be a mistake.

The endorsement pattern is what it is. The two US labs whose main product moat is a closed API endorsed a pacing instrument that runs on pre-release review. The two US labs whose distribution philosophy explicitly rejects a closed pre-release perimeter did not. The instrument shape favors the endorsers on both regulatory instruments Washington is currently drafting for frontier AI, the near-term launch bar and the long-term pacing mechanism. Neither instrument constrains Chinese open weight releases, which continue to ship every three weeks. Both instruments shape the release calendar of the four US labs that operate closed API tiers underneath them.

For builders, the actionable read is a continuation of yesterday's, not a departure. If your product routes through OpenAI or Anthropic, budget for two regulatory layers on the 24 month roadmap: a 30 day launch bar on every flagship release and a longer-horizon pacing instrument that does not exist yet but that Washington is being formally asked to fund. If your product routes through open weights, the near-term pacing regime does not touch you, and the political oxygen it consumes may push more of the 2027 governance conversation onto open weight distribution rather than off it. Every abstraction that lets your product swap providers without an API contract change is worth more today than it was on Monday, and every provider whose corporate position on pacing you have not read is a provider whose regulatory exposure you are not yet pricing.

Three signposts to watch. Whether Google publishes a corporate position on the pacing letter before the launch bar text lands on Saturday, or whether the DeepMind CEO seat stays silent while the DeepMind cofounder's individual signature stays on the letter. Whether Meta's Q3 open weights release cadence changes at all, since a public counter-position at CEO level is usually followed by a matched product move. And whether Congress attaches pacing infrastructure funding to the FY2027 appropriations cycle when it reopens in September. If it does, the letter converted political oxygen into dollars in one legislative session, and every future pacing conversation runs through the office that got the appropriation.