"A Truce in the AI War?" — The Exhaustion of Companies Visible in the Pacing of the Frontier (updated September 27, 2026)

A strange sign of a "truce" has appeared in the U.S. frontier AI competition.

Dario Amodei of Anthropic has proposed an intentional adjustment of development speed, and Sam Altman of OpenAI, Elon Musk of xAI, and Demis Hassabis of Google DeepMind have successively expressed agreement with this direction. Furthermore, 1,386 people working at major AI companies have signed "Pacing the Frontier." However, the comments from the signatories are not the official views of their respective companies. What can be confirmed at this time is that executives and employees have almost simultaneously begun to view the "too fast competition" as an issue (Pacing the Frontier, CNN).

Is this cooperation for safety? Or is it a way of coordinating terms on the edge of collusion to escape a war of attrition that continues to burn through computational resources?

1. Why "Safety Cooperation" Looks Like Collusion

Amodei's proposal is not a complete halt. He proposes creating "checkpoints" where external evaluators are placed within AI companies to slow down development only when dangerous capabilities reach a certain point. Furthermore, he argues that the U.S. government should mediate to ensure that competitive disadvantages are not concentrated on a single company. Since this is an ongoing conversation regarding safety, he argues that even limited antitrust exemptions may be necessary.

Crucially, Amodei explicitly states that by slowing down, they can gain "1 to 2 years" without losing commercial advantage or the U.S. lead over China (We Must Pace the Frontier). In other words, this is a concept not just for safety, but for leveling the competitive playing field.

On the other hand, according to guidelines from the Federal Trade Commission (FTC), cooperation between competitors does not immediately become illegal. Joint safety standards can serve the public interest. Collusion is strongly suspected to be illegal when agreements are made regarding prices, supply volumes, or market division to avoid competition (FTC "Dealings with Competitors"). There is no evidence in public records that companies have agreed on model prices, supply volumes, or release dates. While saying it "looks like collusion" may be valid, we are not at the stage of concluding that "illegal collusion has been established."

2. The Supply Side is Already Out of Breath

Even so, there are physical reasons why pace adjustment is being welcomed: computational resources.

On September 10, 2026, OpenAI temporarily suspended new subscriptions for the $200/month ChatGPT Pro. TechCrunch reported that demand for the latest model, Astra, was straining the infrastructure, leading the company to stop accepting sign-ups for the most heavy-load $200 plan (TechCrunch).

Anthropic also does not appear to be a supplier with ample capacity. The company announced a contract with SpaceX to utilize the entirety of Colossus 1's computing capacity. This is nominally over 300 megawatts, with more than 220,000 NVIDIA GPUs. It is a larger commercial contract than simply "borrowing Musk's servers," but the fact that they need to secure capacity to this extent speaks to the intensity of the inference supply competition (Anthropic).

However, it cannot be definitively concluded from this that "they proposed slowing down because computational resources ran out." According to OpenAI's public documents, while they paused reinforcement learning for Astra for two weeks due to safety reasons and reduced GPU allocation for it by 59.2%, they increased allocation for other models by 17.2%, offsetting 85% of that. This serves as evidence that total computational resources did not stop, but were redistributed (Summary of public documents by Xcademia). While resource tightness can be confirmed, there are no internal documents showing that this is the direct cause of the pace adjustment.

3. Opus Inside Kimi

Another factor that changed the perception of the competition was Anthropic's threat report in September 2026.

Anthropic does not provide Claude commercially in China, and its terms of service prohibit Chinese companies from using it via overseas subsidiaries (Anthropic's Regional Restrictions). Furthermore, the company reported that Chinese AI companies used aliases, forged/stolen cards, stolen API keys, and proxies to circumvent regional restrictions. The usage patterns indicated by Anthropic's investigation are clearly unauthorized and circumvention-based, rather than standard contractual usage.

Regarding Moonshot AI, Anthropic claims that Kimi forwarded customer requests to Claude without notifying the users, and then displayed Claude's responses as Kimi's responses. The reported scale is as follows (Anthropic Threat Report):

To be precise, it has not been confirmed that "all of Kimi's processing was Opus." What has been confirmed is Anthropic's finding that a significant amount of requests were secretly forwarded to Opus, and the responses and reasoning processes were saved for model training. However, this greatly undermines the premise of viewing the performance of Chinese models as an achievement independent of U.S. models. Alibaba, DeepSeek, and Zhipu were also identified in the same report for large-scale distillation and research and development using U.S. models.

4. What was leaked was not "all of China's secrets"

Even more serious is that user data from the Chinese side flowed into the bypass routes.

According to Anthropic, users suspected of having ties to the People's Liberation Army (PLA) were feeding footage from hundreds of surveillance cameras in Chengdu into Kimi. This included cameras surrounding PLA facilities, research institutes affiliated with China Electronics Technology Group Corporation, and state-owned enterprises. Another engineer from a state-owned enterprise was inputting internal code and valid credentials from multiple Chinese companies. It is also reported that via DeepSeek, internal documents containing specifications, organizational structures, and strategic objectives regarding flagship AI projects of Chinese companies were transferred to Claude.

Therefore, the statement that "government and military secrets were completely leaked to the United States" is too broad in scope. What public records confirm is that multiple significant specific instances—such as surveillance data appearing to be PLA-related, code and credentials from state-owned enterprises, and internal documents from police and corporations—flowed to Anthropic. While the full picture remains unknown, it can be said that the failure in information management was of a national security grade.

The Chinese government rejected the U.S. claims as "groundless," arguing that distillation is a common practice. However, at the time of reporting, no technical explanations contradicting individual cases from Moonshot, DeepSeek, Alibaba, etc., have been presented (AP News). In this article, we will simply record this rebuttal as an event where a rebuttal was made.

5. Not a ceasefire, but a deceleration while maintaining superiority

Taken together, "Pacing the Frontier" reveals a different aspect.

As U.S. companies build smarter models, the demand for inference expands, forcing them to secure massive amounts of GPUs and power. At the same time, Anthropic has presented evidence that major Chinese labs have been illicitly using U.S. models and using them even for distillation or as substitutes for their own services. From a situation where no single company could stop due to the competition with China, a calculation has become easier to justify: "If the opponent's speed also depends on access to U.S. models, we can tighten access while taking time to establish our own safety measures." This is the hypothesis of this article.

There are no minutes or internal testimonies that directly demonstrate this view. It cannot be said that China's entire AI development has reached a standstill, nor that China is no longer a threat. Amodei himself calls for export controls and measures against model theft due to the risk of China catching up. At the core of the proposal is not the abandonment of competition, but buying time for safety preparations within a range that does not erode U.S. superiority.

The fears regarding safety, the strain on server resources, and the exposure of the circumvention by Chinese entities have surfaced simultaneously; "deceleration," which was taboo until yesterday, can now be discussed as a business strategy. To me, this proposal looks less like a ceasefire agreement and more like a blueprint for designing conditions where one can step on the brakes while maintaining superiority.

6. Agreement alone does not become a common brake

The agreement of various companies does not mean they have already slowed the pace of the capability race. Anthropic released Fable 5.1 and Mythos 5.1 on September 1, and OpenAI released GPT-6 Astra on September 3 (Anthropic, TechCrunch). What Amodei immediately promised in his September 12 proposal was the first stage: the permanent acceptance of third-party evaluators within the company. The second stage—aligning the speed of capability improvements—is positioned as future coordination between the industry and the government.

On social media, observations are spreading that next-generation models such as Opus 5.2, Fable 5.2, GPT-6 Sol, and Grok 4.7 are imminent. As of September 14, no announcements or model cards from the respective companies confirming these names and release timings have been released. Nevertheless, unless release delays, reductions in training scale, or the start of accepting third-party evaluators are demonstrated, the discrepancy between the words saying "it's too fast" and the actions of continuing competition will remain.

Government alignment is also not in sync. On September 13, when asked whether AI development should be decelerated or regulated, President Trump stated, "The winner in AI wins," indicating a stance of not wanting to cede U.S. superiority to China. Regarding movements calling for AI safety, he criticized "negative forces" who bring up topics that should not be discussed (AP / South China Morning Post). While Amodei's second stage assumes the support of the U.S. government, the President does not agree with the perception of crisis that serves as its starting point.

What has been agreed upon is only a part of the perception of risk, not who will slow down and when. It is difficult for a company to stop alone as long as others are running, and the government pours cold water on deceleration due to the competition with China. In this combination, a state is established where everyone continues to step on the accelerator while everyone speaks of danger.

Model release intervals, postponement of release due to safety evaluations, the actual dates when external evaluators entered, and adjustments under competition law by the U.S. government. Only when these four move will "pacing" transform from a statement by some executives into a verifiable brake.