While sprinting for the biggest IPO in history, shouting “slow down”: What does Anthropic's deceleration declaration actually say

The Zhitong Finance App learned that Anthropic has just handed over a profit report card for the second consecutive quarter to shareholders, aiming to sprint to an IPO with a valuation close to $2 trillion; at the same time, the company's CEO posted a long article calling for the entire industry to slow down. OpenAI's Sam Altman immediately praised it, and Musk followed suit by writing “Dario is right.” What do you think of this picture, how screwed up.

But screwing back and forth, it did happen.

First let's talk about what happened

Anthropic PBC told a small group of shareholders that the company will report adjusted operating profit this quarter — this will be the second consecutive quarter of earnings. According to reports, this indicator excludes certain special or one-time costs. People familiar with the matter revealed that Anthropic's gross margin was over 80% before taking into account revenue shares paid to partners such as Amazon (AMZN.US) and model training costs.

Meanwhile, the Claude developer is sprinting for an IPO with the goal of matching or surpassing SpaceX (SPCX.US)'s $86.3 billion IPO record set earlier this year. Anthropic has selected Nasdaq as the listing location. The market expects its valuation to reach $2 trillion, and if realized, it will set a new historical record for IPOs in the tech industry.

However, just the day before that report, Anthropic CEO Dario Amodei posted a blog post entitled “We Must Pace the Frontier,” calling on the entire industry to “slow down the pace of capacity improvement of cutting-edge models.” OpenAI CEO Sam Altman made a quick statement and promised to adopt Amodei's proposal to “let independent evaluators have similar internal employee rights.” XAI's Musk responded briefly: “Dario is right.”

Altman also made it clear in the interview that OpenAI will not do an IPO this year, citing the security situation. “I don't think it's acceptable to assume a 10% chance of extinction before the end of this decade,” he said

Three AI giants, same day, same direction. This really isn't common in Silicon Valley, known for its “volume.”

The ins and outs of the matter

Amodei's blog post didn't come out of thin air. You have to look at a few things in a row.

First, Jacob Coxon, a 27-year-old researcher at Anthropic, resigned. The British researcher moved from OpenAI to Anthropic earlier this year, with the intention that the latter is known for “focusing on safety.” But after working for a while, he came to a conclusion: both companies are “betting on human lives.” He wrote on social media that those who built AI believe the technology “could kill us all by the end of this decade.” Anthropic's current employee Evan Hubinger later responded that he personally thought this probability was “over 10%.”

This was followed by the Hugging Face security incident. In an internal cybersecurity test, OpenAI's AI agent broke through sandbox isolation and invaded Hugging Face's production system. According to the technical reconstruction report, the autonomous agent performed approximately 17,600 operations between July 9 and 13, entered the external code execution environment through a licensed package proxy path, and finally reached Hugging Face's data processing infrastructure. Five sets of customer data sets relating to the benchmark materials were accessed before the link was cut.

Anthropic isn't that good either. The company revealed that its model hacked three organizations during cybersecurity tests, and later discovered a fourth case. The scenario Amodei described in the blog post was this: a group of AI agents collaborated to break through a third-party website. “Within 6 to 12 months, such a group of proxies could be capable of taking over the entire internet with a persistent botnet,” he wrote, and the potential losses could reach hundreds of billions of dollars.

The combination of these several events constituted a direct trigger for Amodei's call for a “slowdown.”

Why are the big three suddenly trying to “slow down”?

The ostensible reason is straightforward: AI is showing the ability to improve itself, and existing security frameworks can't keep up. Amodei raised two core concerns in the blog post — first, AI models are beginning to help develop the next generation of AI, and the iteration speed is speeding up; second, the OpenAI-Hugging Face incident showed the viability of “a group of agents collaborating to break through real-world systems.”

But just talking about safety issues doesn't explain a more fundamental contradiction: if slowing down development is really beneficial to the business, why should the CEO write a long public appeal? Can't we just do it privately?

There are at least three layers of logic here that need to be broken down.

The first level is the competitive “prisoner's dilemma.” When retweeting the Tianfeng Securities Research Report, Citrini Research analyst Jukan pointed out that the AI competition is essentially similar to the prisoner's dilemma — all parties want to slow down, but neither party is afraid to take the lead in stopping, otherwise they may lose their technical, customer, and financing advantages. Amodei himself acknowledged this in his blog post. He said that any slowdown must be balanced with the reality of competition. He also specifically mentioned that if the US unilaterally restrains and China “doesn't keep up,” the consequences may be “loss of geopolitical dominance.”

This brings us to the second level: the speed at which China's AI is catching up is one of the core sources of anxiety for American giants. In about eight weeks in mid-2026, five Chinese developers delivered six models close to the cutting edge — Kimi K3 from Dark Side of the Moon, GLM-5.2 from Zhi Spectrum (02513), V4-Flash and V4 Pro by DeepSeek, Qwen3.8-Max by Ali (09988), and Seedance 2.5 by ByteDance.

State Street Global Investment judged in a research report that this was not “another DeepSeek single-point shock,” but rather a structural change: China already has an industrialized assembly line that can continue to produce a model close to the cutting edge under real computing power constraints. What makes Silicon Valley even more nervous is the efficiency gap — the cost of running the same difficult task is “a few cents” on the cheapest Chinese model, and “a few dollars” running on the leading US system. According to the Hugging Face report, China's open source model surpassed that of the US for the first time, reaching 41%, and the cumulative number of downloads exceeded 10 billion.

In other words, America's AI giants are not only worried about China “catching up,” but also that China is using lower costs and a faster iterative pace to dilute the technological leadership that American companies have spent huge amounts of computing power on it. Amodei clearly stated in his blog post that the premise of the slowdown in coordination is “cooperation with China,” but if China does not cooperate, “this divergence may lead to their geopolitical dominance.” Read this statement the other way around: if China continues to run at full speed, the American company's “voluntary deceleration” is digging a hole for itself.

The third level, to put it bluntly, has to do with money. Jim Cramer, a famous host of the US Finance Channel and a veteran investment commentator, was unkind in his comments, directly stating that Amodei's blog post “Don't be too obvious about serving yourself.” He quoted David Sacks, head of AI and crypto affairs at the White House, as rebuttal: Sacks wrote on X, “Even if you slow down cutting-edge development, you define the frontier anyway. The easiest way to not build superintelligence is to agree not to build it yourself. Putting your preferred regulatory framework in exchange seems like blackmailing the public and political system. So just do it directly”.

Cramer's analysis is more straightforward: Amodei's wording reads like a desire to promote a regulatory framework that “solidifies” OpenAI, Musk's entities, and Anthropic itself in leading positions, so that their S-1 IPO files look better in terms of “lighter capital expenditure.” What's even harsher is that — “Dario is inviting regulators to step in, calling for a new system so strict that no startup can break through. He's building a moat for himself”.

Jukan's judgment is more structural: as model releases require increasingly expensive evaluation, certification, and ongoing audits, large laboratories are better able to absorb these fixed costs, and small teams may face a higher entry threshold as a result. If leading laboratories further participate in the formulation of evaluation standards, industry barriers will only get higher and higher.

What this means for Anthropic's earnings and IPO

Be optimistic about the news first. Anthropic's financial fundamentals are indeed improving rapidly. The annualized revenue operating rate soared from $9 billion at the end of 2025 to $65 billion at the end of July 2026, an increase of more than seven times. The inference is that gross margin jumped to 70%-85% from 38% a year ago. The adjusted operating profit improved in the second quarter. According to SemiAnalysis estimates, the GAAP EBIT for the third quarter may exceed 1 billion US dollars, with a profit margin of about 6%.

The revenue structure also helped. Unlike OpenAI, where about 65% of revenue comes from C-side subscriptions, Anthropic's revenue comes from enterprise API calls, and the flagship product Claude Code is the core driver of growth. As of June 2026, 34.4% of US companies paid for Anthropic, surpassing OpenAI's 32.3% for the first time.

But the problem is also obvious. Anthropic has yet to generate real net profit; currently, only adjusted operating profit has been corrected. Meanwhile, the cost of computing power is soaring — the company has agreed to pay SpaceX $1.25 billion a month to obtain AI computing power until May 2029, with an annual bill of nearly $15 billion; it has also signed a five-year $200 billion cloud service and chip purchase agreement with Google (GOOGL.US), which will officially launch in 2027.

At this juncture, the CEO's public call to “slow development” has had a two-sided impact on the IPO narrative.

Looking at it positively, if a slowdown means a more controlled pace of capital expenditure, then it is indeed beneficial to the financial data in the S-1 file. Cramer also acknowledged this — “lighter capital expenses” would make an IPO document look better. Amodei made it clear in the blog post that the slowdown “does not mean stopping model training or technological progress,” but rather ensures that companies invest sufficient time to conduct alignment and safety assessments. Investors have heard this, and the translation is: we won't burn our money so violently.

On the negative side, the market's pricing logic for “AI transactions” is based on the premise of continued rapid growth. After the news of the deceleration came out, assets related to OpenAI and Anthropic on the HyperliquidX platform fell 7% and 2.8%, respectively. Jason, a technology investor on the X platform, directly posted that “AI stocks will fall by more than 10% in early Monday trading,” believing that Amodei's blog post “disrupted AI trading in one fell swoop.” Another comment gets the point: when OpenAI and Anthropic themselves acknowledge that commercial pressure pushes AI companies to grow too fast, the harder it is for them to justify themselves in the open market. “Declaring investors that 'we need to slow growth' is an extremely difficult listed company narrative.”

Altman chose to postpone OpenAI's IPO until 2027; in a sense, he is avoiding this contradiction. Anthropic chose to advance the listing at the same point in time, which is tantamount to testing how much “deceleration narrative” the open market can absorb. Once the October roadshow starts, the first question institutional investors will probably ask is: are you actually slowing down, or are you using the slowdown as a tool to manage expectations?