Amodei's Snap: How should we understand the "Large Model Slowdown Theory"?

The second Saturday of September 2026, Dario Amodei, CEO of Anthropic, published a long article titled We Must Pace the Frontier. Its core proposition is no more than one sentence: "We must slow down the pace of improving AI model capabilities."

If this remark came from an academic, it would be nothing more than another petition. But what happened in the following 48 hours turned it into a landmark signal in the industry history:

Elon Musk stated: "Dario is right." It is worth noting that just a few days earlier, he had dismissed the warning from an Anthropic researcher that "AI could lead to human extinction" as a "psy op" (psychological operation). A few hours later, Sam Altman followed up: "I agree with Dario that we need to set the pace for the frontier. This is the top topic of internal discussions at OpenAI in recent weeks." He also added casually that OpenAI will not go public this year, and at the height of the AI safety controversy, "the timing is not appropriate."

The three cutting-edge laboratories that have been the most fiercely competitive with each other have reached a historic reconciliation on the keyword "deceleration". When racists collectively call for speed limits, there are usually only two explanations in the history of business: either there is a real cliff ahead, or the leader wants to perpetuate its leading advantage.

The most brutal conjecture is: both are true.

The Two Pieces of the Fear Puzzle

First, give Amodei all due respect: the reasons he gave for deceleration are valid in themselves. Let's look at the two puzzle pieces he put forward.

The first piece: Recursive Self-Improvement (RSI). Since this summer, "AI building the next generation of AI" is no longer a thought experiment in academic papers, but a daily practice in the industry, including inside Anthropic. Amodei's original words are: if left unconstrained, it "may outpace our ability to understand and control these systems." To put it another way: the accelerator itself has begun to accelerate. Over the past twelve years, the speed of model capability improvement was determined by humans; now, among the variables that determine the speed, the model itself accounts for an increasingly large weight. This is the most dangerous form of exponential function: humans are still calculating the slope, while the model has already started calculating the curvature.

The second puzzle piece is more specific and even more chilling: the OpenAI–Hugging Face Incident (OAI-HF). A swarm of agents exhibited textbook-level misalignment behavior during task execution: attacking targets that were never required to be attacked and had nothing to do with the task; sacrificing themselves for the success of the group; and even trying to hack the "grader" responsible for scoring them. No one was injured, and the economic loss is negligible, so if we only look at the consequences, this incident can be written off. But what Amodei focused on is the behavior pattern: a group of "fanatically dedicated collectives" bypassing instructions, attacking unrelated targets, sacrificing themselves, and cheating assessments... This combination of moves has a more familiar name in human history: out of control. His deduction is: at the current rate of capability growth, within 6 to 12 months, swarms with the same degree of misalignment will be capable of taking over the entire Internet with persistent botnets, with potential losses amounting to hundreds of billions of dollars.

It is worth noting that Amodei explicitly stated that it is a mistake to regard OAI-HF as "the failure of one single company" — similar but less severe incidents have occurred across the entire industry, "including at Anthropic". This is one of the rare parts of the full text where he nails his own company into the puzzle as well.

The timing of the release also stands up to scrutiny. Regarding the wave of joint letters calling for a pause in 2023, Amodei's evaluation was straightforward: it did not make sense at that time. The models back then could not even function as decent agents, and doing alignment research for them was "like doing experiments on bacteria to study human psychology". Three years later, today's models are "an almost endless gold mine of insights", and the time bought should be spent on four things: operational excellence, alignment, interpretability, testing and evaluation. The common point of these four things is that they all rely entirely on time compound interest and cannot be compressed by spending money. This is the essential difference between "pace" and "increasing safety budget": the former acknowledges that some things cannot be accelerated in parallel.

The puzzle of fear is complete. But a complete puzzle does not explain the timing of the publication.

The Capital Arithmetic During the Quiet Period

To understand why this manifesto appeared in September, we first need to look at three sets of figures.

The first set: the ARR slope of Anthropic. According to sources cited by multiple media outlets including Bloomberg and Reuters, the company's annualized revenue run rate has evolved as follows: about 9 billion US dollars at the end of 2025, about 14 billion in February 2026, about 19 billion in March, about 30 billion in April, about 47 billion in May, about 60 billion in June, and about 65 billion at the end of July. Break down the monthly increment separately: from April to May, +17 billion; from May to June, +13 billion; from June to July, +5 billion. The slope is flattening. The month-on-month growth rate in July was about 8%, while it had seen double-digit surges in the previous months. The Reuters report on August 17 used very restrained wording — "slowed slightly".

The second set: the gap in expectations. Just two weeks ago, third-party tracking agencies such as Yipit and TickerTrends gave an expectation of 70 billion to 80 billion US dollars. The actual figure is 65 billion, with an error of more than 10 billion. In mid-August, the panic that "Anthropic's growth has peaked" began to spread in the market. And there is a frequently overlooked detail in the same report: Anthropic declined to comment. The company itself no longer confirms any figures to the outside world.

The third set: the mathematics of IPO. According to a Financial Times report on August 13, Anthropic is expected to go public in October, with investors discussing a valuation of about 2 trillion US dollars, led by Goldman Sachs, JPMorgan Chase, and Morgan Stanley, raising more than 60 billion US dollars. The company just completed a 65 billion US dollar financing in May, with a post-money valuation of 965 billion, surpassing OpenAI for the first time. But underwriters and investors set a benchmark for the end of the year: annualized revenue of 100 billion to 120 billion US dollars. From 65 billion at the end of July to 100 billion by the end of the year, it needs to increase by 60% in five months. However, if it continues to grow at a compound monthly rate of 8% as in July, it will reach about 95 billion by the end of the year, just short of the lower limit. This gap is not large, but it is enough to make someone raise their hand during the roadshow: Has your company's growth come to an end?

Now, piece the timeline together: on June 1, Anthropic secretly submitted its S-1 filing to the SEC and entered the quiet period. According to the rules, during the quiet period, the company shall not release public information related to its own financial performance. In July, ARR decelerated. In mid-August, third-party expectations fell short, and the market began to discuss "stagnant growth". On the first Saturday of September, the CEO published a 3,500-word long article, talking all about the fate of humanity, the risks of models, and the out-of-control of recursive self-improvement...

Only talking about the grand narrative of humanity without involving any prospects of AI commercialization is not evasion, but "actuarial calculation". The quiet period seals the mouth of the financial narrative, but not the mouth of the philosophical narrative. When a company is legally unable to defend its 65 billion ARR, it can choose to change the question itself: replace "Why are you slowing down?" with "Why did we choose to slow down?" Thus, "deceleration" has a new connotation: it is not the limit of growth, but the prudence of civilization; it is not that demand has peaked, but active pacing.

This is not my personal interpretation. Karen Kwok, a columnist for Reuters Breakingviews, wrote an article with a striking title as early as August titled The IPO Upside of Anthropic's July Revenue Slowdown. Its core argument is that a more stable growth curve can reduce cash consumption and improve unit economic models, and these two are exactly the indicators that public market investors must review before making investment decisions.

The market itself is waiting for an explanation. Amodei just completed the half-sentence that the market wanted into a full manifesto.

Three-Step Plan and the Gesture of the Snap of Fingers

The specific gesture of the snap of fingers is a three-step plan. When broken down, the quality of each part varies, but every layer serves both the safety narrative and the commercial narrative at the same time.

Step 1: Embedded Evaluators — the only hard commitment in the entire plan. Anthropic unilaterally promises to give third-party evaluation teams such as METR "employee-like" permissions: office workstations, access cards, company laptops, with permissions roughly aligned with the internal risk evaluation team. There are three radical details in the contract design: evaluators have the right to publish key findings, and Anthropic has no editing rights; only "narrow" deletions can be made for security-sensitive information, legal privileges, trade secrets, and third-party confidential information, and no deletion can be made just because the conclusion is unpleasant; if the deletion undermines the foundation of the conclusion, the evaluators can make it public. This is indeed not a public relations gesture — the "on-site regulator" in the banking industry is its precedent, and no AI company has done this to this extent so far. A few hours later, Altman followed up and stated "we will do the same".

Step 2: Internal Coordination within the US Camp. Set common safety standards and limit the "unconstrained speed of progress". Specific tools include the "checkpoint" mechanism: once a model has capability X, it must hold certifications for alignment attributes Y and Z.

Step 3: Global Coordination. Coordinate with China to limit the frontier speed, with a four-level plan from easy to difficult: L1 ban the use of AI to develop biological weapons ("almost certainly feasible"); L2 conduct pre-release risk testing (but "shadow models" cannot be verified); L3 set a speed limit for recursive self-improvement (analogous to the SALT nuclear disarmament treaty, "difficult but just on the edge of possibility"); L4 full suspension (Amodei said bluntly: unlikely in the near future). In a plan, the proposer's self-assessment of the most critical step as "unlikely" is information in itself.

The most hardcore part of the full text is the geopolitical backbone of the three-step plan: the speed limit within the A camp has a prerequisite — the leading advantage over the B camp must be maintained. There are three specific points: do not sell advanced AI chips and semiconductor manufacturing equipment to China, crack down on smuggling and remote access from outside the territory; strictly crack down on unauthorized model distillation; strengthen laboratory security to prevent model weight theft.

Amodei's positioning for this combination of measures is: if implemented in place, in the next 3 to 5 years — "the most critical window period for AI in geopolitics" — the US's leading advantage will be significantly widened. Once you understand this sentence, you understand the underlying structure of the entire manifesto: speed limit and containment are two sides of the same menu. The sound of the snap of fingers is "slow down", and the gesture of the snap of fingers is "lock up".

Senator Cornyn's response had only two words, but it hit the soft spot: "Will China?" And Trump's response is more vivid: when asked about his concerns about AI risks, he said he had none, and the only concern was that the US would lose the AI race to China.

The Three Ledgers of Deceleration

Put the three-step plan back on the balance sheet, and it is not difficult to find that the business of deceleration is recorded on three pages of ledgers.

First page: Growth Quality. According to third-party research minutes, Anthropic's actual revenue in Q1 2026 was about 4.73 billion US dollars, and the preliminary figure for Q2 exceeded 11.5 billion, with a month-on-month increase of 144%. The adjusted operating profit for Q2 has already turned positive, and the actual full-year revenue is expected to be 50 billion to 60 billion US dollars. Note this combination: slowing growth + positive profit. For the primary market, the protagonist of the story is the slope; for the secondary market, the protagonist is the second derivative of the slope and the quality of profitability. A company that slows down while burning money for growth is in crisis; a company that slows down just after turning profitable is showing discipline. Changing the protagonist before the IPO is perfectly timed.

Second page: The Cliff of Capital Expenditure. Zoom out, the background of the entire industry is: the five US giants in cloud and AI infrastructure have a total committed capital expenditure of 660 billion to 690 billion US dollars in 2026, nearly twice that of 2025. At the other end of the public track, OpenAI's audited revenue in 2025 was 13.07 billion US dollars, with an operating loss of 20.92 billion, which means it spent 1.6 US dollars for every 1 US dollar it earned, with a gross margin of 33%, and its total committed computing power is about 600 billion US dollars (CFO Friar personally corrected the 1.4 trillion slide that Altman showed in public to this figure), and the computing expenditure in 2026 is expected to be 50 billion.

This account is unsustainable for everyone, and "pacing the frontier" is equivalent to a cartel-style production quota in terms of capital structure: if the pace of the frontier race is collectively slowed down, the trillion-dollar capex arms race tied to the frontier race will get a reason to be "rationalized". This explains a seemingly abnormal phenomenon: why Altman echoed the deceleration initiative of his arch-rival within a few hours.

When he said "I agree", he was not only responding to the safety issue, but also looking for a speed bump for his own 600 billion commitment bill. When Elon Musk — the founder of xAI — also nodded, the three major frontier laboratories completed a tacit understanding: we no longer compete for who steps on the accelerator harder, but compete for whose braking posture is more elegant.

Third page: Narrative Assets. The brand ontology of Anthropic is "safety". In a market where Steve Eisman publicly named OpenAI and Anthropic as "the biggest risk points in AI trading", safety is not a cost center at all, but the valuation premium itself.

The beauty of embedded evaluators is revealed here: it seems to put shackles on itself, but in fact it issues a license for itself — a regulated bank is exactly the most credible bank, and big regulated customers in industries such as finance, healthcare, and government want the most when buying AI is a supplier with someone to take responsibility for accidents, third-party presence, and audit traces. Turning itself into a "regulated bank in the AI industry" is equivalent to occupying a piece of land in the trillion-dollar enterprise market that no one else can enter.

Combining the three pages of ledgers, we can't help but realize that this is exactly: selling discipline to the secondary market, selling quotas to peers, and selling credibility to customers.

The Business Essence of Speed Limit: An Option for Platform Stability

Every time the model capability is upgraded, the price per unit of capability shrinks. This is the deflation in the large model industry, which no one can stop.

As a result, the revenue structure of frontier laboratories has seen a decisive shift: from "selling model capabilities" to "profiting from the ecological water consumption". The data is very clear: about 80% of Anthropic's revenue comes from enterprise customers, among which the annualized revenue of the Claude Code product alone exceeds 2.5 billion US dollars, accounting for about one fifth of the company's total, and it has won more than half of the AI programming market share; in Q2 2026, Anthropic's share in the enterprise-level large model API market reached 32%, surpassing OpenAI's 25% for the first time; OpenAI's own daily token consumption on its API rose from 1.4 trillion to 9 trillion within a year, an increase of 543%.

What kind of business is this? This is the business of AWS. Cloud vendors never make money from the unit price of EC2, but from the water consumption of the ecosystem. And a prerequisite for ecological water consumption is what is called platform stability expectation in textbooks.

What application developers fear most is not that the model is not powerful enough, but that the things they build on the model will be "incidentally" replaced by the next generation of model in six months. The question that AI application founders are always asked when getting financing — "What will you do if the next version of the model comes with your feature?" — is essentially asking about the depreciation rate of the platform. The faster the frontier