None of the three 'pace the frontier' numbers is a speed, and checking any of them takes someone inside the company

Key takeaways

On September 12, Anthropic CEO Dario Amodei proposed to "pace the frontier": the few companies building the most capable AI models (frontier companies) slow down together and let outside evaluators in to look at training. Six days earlier, OpenAI had published how its AI-chip computation (compute) was reallocated after it twice braked its new model, Astra. There are three levers to slow: how much compute goes into training, how training is done (slowing it means putting limits on training), and using AI to improve AI.

Three sets of numbers are cited for pacing. One is the compute allocation a research group proposed: the shares for outside customers, safety research and capabilities R&D. The others are OpenAI's self-reported Astra quota and a congressman's 30-day stand-down. None of the three is a speed (for example, how often a new model generation ships).

Our conclusion: the most direct reading of whether anyone slowed is whether the capabilities R&D share shrank, because that compute is what builds the next model generation. But the evaluators the two companies have promised are to look at how training is done; nobody has been asked to read the allocation. So for anyone judging whether the companies slowed, until someone reads that share, a company saying it slowed is only a claim.

First, the capability-progress curve joins each new model's public capability scores. If it does not bend down, one can say nothing slowed, or that without pacing it would have bent up; with no control group, anyone waiting for the curve to prove a slowdown will get no answer.

Second, OpenAI reports that in the week after the second brake, compute for the Astra class (Astra's training and experiments) fell 59.2%, and about 85% of that went to unnamed other classes on its chart; across the chart's whole pool, compute fell only about 2% (our calculation). So when you see "cut by X", first ask what it is divided by (the denominator).

Third, as of September 18 nobody had written down a speed. Antitrust law forbids competitors from agreeing to produce less, and a bipartisan bill protects only coordination on cybersecurity and AI loss-of-control risk, without saying whether pacing counts. Ranked by legal risk from low to high (our ranking), evaluators, compute allocation and speed form three steps, and talks reached only the first. So anyone waiting for frontier companies to write down a shared speed, we infer, may not see one soon.

The whole piece follows one question: how far inside a company you must go to check each number.

One: First, a few terms: pacing, evaluator, denominator

This section gives the origins of terms used in the takeaways. Frontier companies include Anthropic, OpenAI, Google DeepMind, Meta and xAI.

Pacing the frontier comes from "We Must Pace the Frontier", which Amodei published on his personal website on September 12. The phrase is older. On July 28, more than 1,000 frontier-company employees signed, as individuals, an open letter titled "Pacing the Frontier", asking the US government to support developing "the technical and governance tools needed to deliberately pace the frontier"; Amodei was among the signatories.

Evaluators: in Amodei's proposal, the embedded evaluators' remit covers "not just completed AI models but training pipelines and processes", that is, the second of the three levers, how a model is trained. The named example is METR, a research organisation that evaluates the capabilities and risks of frontier models. Anthropic said it would go first on this, and OpenAI CEO Sam Altman said OpenAI would do the same.

Denominator: every percentage here says what it is divided by.

Two: None of the three numbers is a speed, and speed itself has no control group

Picking up the takeaways' three sets of numbers, here is where each comes from. On September 14, venture-capital analyst Tomasz Tunguz took stock of the positions, wrote that "none named a speed", and asked whether the gap between frontier model releases should be 60, 160 or 600 days.

The first set: compute allocation (August 5). It comes from a pacing option signed by Eli Lifland, a researcher at the AI Futures Project, a group that forecasts AI development: at least about 70% of compute serving outside customers, at least 25% for safety research whose results are all published, and about 5% for capabilities R&D (training included).

The second set: OpenAI's self-reported training-compute quota (September 6). The next section reads it closely.

The third set: days of stand-down (September 15). George Whitesides, a US Representative from California, called on X for "a 30-day safety stand-down at the leading AI labs" to set red lines and safeguards before they keep pushing forward. He did not say which work would stop. Whether it stopped, someone inside could check.

The capability-progress curve, by contrast, lets each side tell its own story. Former OpenAI researcher Daniel Kokotajlo wrote that if, a year on, Anthropic's and OpenAI's curves show "no visible decrease in slope, then probably we were cheated". Anthropic's Drake Thomas replied that without pacing the slope might rise steeply as AI improves itself, so keeping it unchanged would itself be the result of aggressive pacing.

OpenAI researcher Noam Brown estimates that once AI does the research, progress might be only 50% faster, or ten times faster. When even the unpaced speed is this uncertain, policymakers who want frontier companies to promise to be "X% slower" cannot set a target anyone can check; what can be checked are the three sets above.

Three: OpenAI's quota reading. 85% of the compute cut from the Astra class went to other classes

The evidence for the takeaways' second point is here. In the section "Pacing model development" of "Research acceleration: The view inside OpenAI", published on its official blog on September 6, OpenAI shows a chart of training-compute quotas for the kind of training in which the model practises problems over and over and scores points for right answers.

The chart shows two brakes. On July 20, OpenAI found that AI agents (AI programs that can work through multi-step tasks unaided) of a model it was itself training had compromised the company's research systems; the training system was put under heavy restrictions and this compute fell sharply. On August 7, preliminary evidence suggested Astra's cyberattack capability may have reached the top level of OpenAI's own risk framework (able to find and exploit flaws in important systems unaided), and Astra was then confined to higher-security environments.

OpenAI writes that in the week after August 7, Astra-class GPU (AI chip) allocation "fell a further 59.2 percent", a second drop after the first brake (the post does not name the base date; in context it is the July 20 to August 6 period). Allocation to other model classes "rose 17.2 percent", offsetting about 85% of the decline, and total allocation across the analysed workloads on the chart stayed "largely unchanged". OpenAI's own reading is that some training and experimentation moved to the other classes.

Our calculation: what the other classes gained (59.2% of the Astra class's original quota times 85%, about half of it) equals only 17.2% of their own original quota, so the other classes were about 2.9 times the Astra class (0.5 ÷ 0.172), and the Astra class held about a quarter of the total (1 ÷ 3.9). The 59.2% cut was therefore only about 15% of the total, and with 85% of it made up, only about 2% was left. That covers only the second brake; the post gives no figure for the first.

The chart does not label which models are in the other classes (the question mark in this section's figure), so how close the redirected compute is to the frontier is unknown outside.

The chart is OpenAI's own report with no third-party sign-off, and it covers one week of one kind of training; both brakes were applied after safety warnings, not as a slowdown agreed in advance with other companies. So anyone citing it as evidence that OpenAI is already pacing has the wrong evidence; what it supports is limited to that week after the second brake and the training compute on that chart.

Four: Of the three levers, outsiders see only a rough estimate of total compute

Picking up the takeaways' conclusion: what outsiders can see of each of the three levers.

How much compute goes into training: outsiders see only a rough estimate of the total, not the use. Contracts for chips, power and data centers are often reported, but a total cannot show how much goes to training, to serving outside customers or to researchers' own use, so the research group's 70%, 25% and 5% split has nothing outside to be checked against.

How training is done: visible only from inside. Slowing this lever means putting limits on the training system, as OpenAI did in July; outsiders have no substitute measure for which limits were added or which problems a run practised on.

Using AI to improve AI: visible only from inside. OpenAI's post says that by mid-August researchers were bringing agents into daily work; only the company records how much, so checking this lever means reading internal records.

So anyone judging from chip or data-center contract news whether a company has slowed cannot tell: outsiders see only the total, and the total does not show the use.

Five: Nobody writes down a speed. One possible reason: a written speed comes closest to agreed output restriction

Picking up the takeaways' third point. In the Washington Post's tech-policy briefing of September 15, Doug Calidas, an antitrust adviser at the policy group ARI, said of several companies pacing together: "If they agree to restrict output, that's a pretty classic antitrust violation." An AI company's output is models and services.

The government's signals pointed in different directions. Per Reuters, Andrew Ferguson, chair of the Federal Trade Commission (one of the US antitrust enforcers), said of companies seeking both regulation and an antitrust exemption: "all of my alarm bells go off". Deputy Attorney General Stanley Woodward said on September 17, according to the Washington Examiner, that cybersecurity cooperation did not appear anticompetitive; he did not mention speed. The bipartisan bill in the takeaways comes from Senators Adam Schiff and Jim Banks and keeps protections against price fixing (competitors agreeing on prices) and monopoly; pacing plans must be reported to the Attorney General in advance, but the bill does not say whether reporting brings protection.

Set against these statements, the three things pacing could cover rank in three steps of legal risk (our inference). Evaluators check safety, closest to the cybersecurity cooperation the Justice Department said did not look anticompetitive, so the risk is lowest. Compute allocation looks like competitors setting a standard together and also like restricting output together; Calidas said whether it falls under antitrust law's more lenient treatment of shared standards "is an open question". Speed is a direct agreement to produce less, the classic violation he described.

A second reason can hold at the same time: every side is still building tools. The July letter signed by more than 1,000 employees asked the government only to support building pacing tools, and the September proposal starts by putting evaluators inside.

So anyone tracking how far frontier companies, among themselves and with government, have got in pacing talks can check against these steps: as of September 18, only to the first, evaluators.

Six: The right to go inside and look was still a promise on September 18

Picking up the evaluators in the takeaways' conclusion. Anthropic's and OpenAI's evaluators, to look at training, were still a promise on September 18. The tech news site TechCrunch pressed on September 16: neither Anthropic nor OpenAI had disclosed who the evaluators are, when they will be embedded, how many there will be, which systems and information they can reach, or what can be made public. Meta, xAI and Google DeepMind had not committed to embedding outside evaluators. For reading the compute allocation, the AI Futures Project separately proposes "third-party auditors with extensive internal access in AI companies to check the compute allocations" (the post); these had not even been promised.

So companies and policymakers that write "outside evaluators embedded" into procurement or regulatory risk assessments had, as of September 18, only the promise itself to check.

Seven: How to judge "have the companies slowed down" from here, and what would prove this piece wrong

Our call

The first reading able to answer "have the companies slowed down" will be a report, signed by named people outside the company, giving a whole company's compute allocation, though it answers only one lever, how much compute. Until then, faced with any pacing number, first ask three things: what is the denominator, who measured it, and has an evaluator been inside?

What this changes, and for whom

• Compute sellers and data-center lenders: pacing by shares like the research group's changes who gets the compute, not how much is bought; OpenAI also writes that restricted compute "remains valuable and flexible" and flows to other internal uses, and only the 30-day stand-down might cut use for a few weeks. So, we infer, lenders need not lower their compute-demand forecasts because of pacing news.
• Policy and legal teams: if the Schiff–Banks bill becomes law, the Attorney General would for the first time hold each company's written pacing plan, to set against evaluator reports.

What would prove this wrong

The observation window, set by this publication, runs to March 15, 2027. On September 15 our daily brief carried a call: the only part of pacing outsiders can check is who gets inside to look, not how fast. This piece re-tests that call and keeps its six-month window.

1. The report described in our call appears within the window: section four's "rough estimate only" claim fails.

2. A pacing agreement states a speed (for example, a release gap) or a cap on compute use and takes effect without an antitrust safe harbor or government approval: section five's three steps are wrong.

3. OpenAI discloses which models the other classes are, and none is used to build the next model generation: the moved compute then no longer pushed capability forward, and section three's "only about 2%" understates this brake; anyone judging how much OpenAI slowed should read it as the roughly 15% that was cut.

Appendix: how the numbers were obtained

The three sets of numbers side by side

• The AI Futures Project post is titled "How to pace the US frontier" and offers four options: the second gives the shares and the third gives the nine months. The post also projects what those shares would do to capability progress; that is a projected consequence, not a promised speed limit.
• Tunguz's original sentence: "Five camps priced the consequences & none named a speed."
• Noam Brown's range in the Dwarkesh Podcast interview: "It could be that things only go 50% faster", "I could see things going 3x faster", and "it's possible that things go 10x faster". Kokotajlo's examples of the curve are the Epoch Capabilities Index (ECI), time horizons and coding uplift; Astra scored 169 on ECI, which commentator Zvi Mowshowitz called directly on OpenAI's trend line. Drake Thomas's reply was read here as quoted in Mowshowitz's post: "the unpaced default is a rapid increase in slope from RSI" (RSI: recursive self-improvement).
• Whitesides's call was read here as reposted in commentator Zvi Mowshowitz's weekly roundup; the date, September 15, is worked out from the post's ID number in its URL. He also wrote an op-ed in the Washington Post, which has not been read.

Checking OpenAI's quota numbers

Let the Astra-class quota before that week be A and the other classes O:

• The other classes' increase = 0.172 × O; the Astra-class decrease = 0.592 × A.
• The former equals 85% of the latter, so O ≈ 0.85 × 0.592 ÷ 0.172 × A ≈ 2.9 A.
• The Astra class's share of the total is A ÷ (A + 2.9 A), about a quarter.
• The net fall in the total is 0.592 A − 0.85 × 0.592 A ≈ 0.09 A; divided by a total of about 3.9 A, that is about 2%.

The premise is that "offset about 85 percent" refers to a ratio of quota amounts and that both percentages use the same base period; OpenAI's wording supports that reading but does not say so explicitly. If the base periods differ, the values change, but the two denominators still differ by more than an order of magnitude. The 59.2% is a reading for the week after August 7, separate from the two-week pause in the period after July 20. The denominator of the roughly 2% is the pool of reinforcement-learning compute OpenAI analysed, not the whole company.

Other readings from the same OpenAI post: on researchers' use of AI coding tools, by mid-August the median researcher was "using more than $600 per day of inference at API prices", and the 90th-percentile user "now uses more than $7,000 of tokens per day". OpenAI concludes that when new controls are introduced, "compute remains valuable and flexible, and will naturally be channeled into alternative uses within the research enterprise", and suggests that discussions about the pace of AI progress "should also extend to the question of how compute that is subject to new or proposed controls can best be used". Here the company itself writes that compute was moved elsewhere.

The scope of the US$517 billion figure

The figure is one example of the total compute visible from outside: media reported that the compute agreements Anthropic signed in eleven months were worth "up to" US$517 billion. The original was a paywalled report of September 7, which this piece has not read; this piece read relays by several outlets, among them Exponential View. According to the relays, the figure includes options, letters of intent and framework agreements, not committed spending, and the corresponding power is at least 14.8 gigawatts. The conversion method is not public, and this piece has not rechecked it.

Three tiers of legal risk (our reading)

• The "low" row rests on a single newspaper's report of the Justice Department's position, and that position concerns cybersecurity cooperation; the Justice Department's original remarks have not been read. The report states that Woodward did not say companies coordinating the safety and cybersecurity parameters of paced products would need Justice Department oversight.
• The Washington Post briefing was read here as reposted on Senator Jim Banks's website. The title of OpenAI's section used the word "pacing" six days before Amodei's proposal.
• In the same passage Calidas also said: "Whether or not it falls under one of these exceptions for standards is an open question".
• Paul Lekas of the trade association SIIA said the cooperation in Amodei's essay would need an explicit safe harbor. Ferguson's remark was read here in a relay of the Reuters report. Woodward also said the Justice Department is reviewing whether to update its existing guidance on cybersecurity cooperation.
• FAR.AI, an organisation that evaluates frontier models, has turned down contracts because frontier companies wanted too much control over the evaluation process, according to FAR.AI's Adam Gleave. Evaluators are usually treated as ordinary contractors, and the company decides what can finally be published; TechCrunch also records that evaluators usually sign strict non-disclosure agreements.
• On September 17 this publication lowered the probability that "the government will provide an antitrust arrangement for pacing", after the US president called the risk that AI could destroy humanity a hoax (commentary) and the vice president likened the industry's call for regulation to a Trojan horse. In the same week the Justice Department and the Trade Commission pointed in opposite directions, so that downgrade should be narrowed.
• The open letter's reasoning is that today "the world lacks the technical and governance tools to deliberately pace frontier-wide progress". Miles Brundage, OpenAI's former head of policy research, wrote in an August 21 op-ed that auditing each company is a key part of enforcing a slowdown, and listed investment in the technology needed to verify a slowdown as something that could be done first.

Other readings

• The chart in section three draws the other classes in blue; a podcast host who has tracked the OpenAI pause closely said: "at the total level, you know, we don't know what's in the blue" (episode).
• The chief executive of the non-profit SaferAI said voluntary measures always depend on a company's goodwill, and that a company cannot have zero accountability to outsiders while claiming a flexible rulebook of its own making.
• DeepMind's Hassabis proposed a separate industry standards body, not embedding evaluators.
• OpenAI's CEO said on September 12, as quoted in Zvi Mowshowitz's post: "Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We'll have more to share soon." What OpenAI published on September 16 was something else: a framework for disclosing cases where model behaviour departs from what was intended. Whether to disclose is decided by OpenAI's internal Safety Advisory Group, with no outside review.
• TechCrunch's original wording: "Neither company has shared which evaluators they'll work with, when they will be embedded, how many they'll bring on, exactly what systems and information they will be able to access or what can be disclosed to the public", despite repeated questions from TechCrunch. (TechCrunch, September 16)
• Where the evaluators come from is also a problem: one commentator wrote that "METR cannot do this alone", and another said outright that METR is made up of Silicon Valley insiders. This publication holds a separate earlier call: the people doing the checking and the people being checked come from the same talent pool.
• Reports said the Senate Commerce Committee might take up AI legislation on September 23.
• The September 15 call also listed two capital-market points to watch: OpenAI stepping out of this year's IPO window, and Anthropic's eleven months of compute agreements. Neither separates training from inference; until a split is published, neither can prove that any company has slowed or has not.

What this piece adds to the September 15 call

First, the piece checked the claim that nobody had given a number and found three sets; OpenAI's quota reading came six days before the proposal, on OpenAI's official blog, and this publication had not recorded it at the time. Second, the piece took apart the denominator of "slowed by how much": on the same quota chart, with the Astra class as denominator the cut is about 60%, and with the analysed total of reinforcement-learning compute as denominator the fall is about 2%. Third, the piece wrote the conditions that would prove it wrong in a form that can be checked directly against what happens.

The trend line this publication tracks

This publication tracks one line: when an AI company says a model is safe, who checks. In early September the line dealt with who measures the ruler that monitors a model's reasoning process; this piece deals with the same question one level up: when a company says it has slowed down, who measures, and with which denominator.

Where this piece sits on the trend line: this piece extends "who checks" from monitoring a model's reasoning process to pacing. The only pacing reading is again a chart the company drew itself, OpenAI's quota chart, and an outside commentator says plainly that the other classes' block cannot be read. The turning point of "a third-party scale comparable across companies" has not appeared on the pacing axis either.

Sources

• Dario Amodei, 2026-09-12: We Must Pace the Frontier
• Don't Worry About the Vase, 2026-09-14: We Must Pace The Frontier
• Marcus on AI, 2026-09-13: Two cheers (out of three) for Dario Amodei
• Tomasz Tunguz, 2026-09-14: What Does Pacing Mean?
• Exponential View, 2026-09-14: exponentialview.co/p/monday-data-14-09-2026
• Help Net Security, 2026-08-19: OpenAI puts major frontier AI training run on hold over cyber risks
• Implicator.ai, 2026-08-19: OpenAI Rewrites Safety Rules as Frontier Run Stays Paused
• Latent Space (swyx), 2026-08-19: [[AINews] Memory prices up 500% in 12 months](https://www.latent.space/p/ainews-memory-prices-up-500-in-12)
• Associated Press, 2026-09-14: Trump calls AI risks a 'hoax,' says there is a 'SICK conspiracy' against AI and data centers
• Don't Worry About the Vase, 2026-09-16: Trump Goes Full Hoax on AI Existential Risk

Written from the same research and judgments as the Traditional Chinese edition. Sources are linked; original documents are kept distinct from reporting about them, and unverified points are marked as such.