When Anthropic CEO **Dario Amodei** urged competitors to slow down AI development speed, **OpenAI**, xAI, and **Microsoft** quickly voiced their support one after another within 48 hours. Meanwhile, Beijing rejected the move head-on, calling it “fostering panic,” and **Moonshot AI** and **DeepSeek** in China reaffirmed their stance that they would keep their existing development pace unchanged.
Key points
In a September 12 op-ed, Amodei proposed a ‘pacing (speed control)’ plan—stationing a third outside evaluation team within leading research institutes to set an upper limit on the rate of model performance improvement.
Chinese developers such as Munsat AI and DeepSeek are completely free from this restriction, and Beijing immediately dismissed the proposal.
With semiconductor and AI-related stocks falling across the board, investors weighed slowing down compute spending against intensifying competition from cheap open-weight models.
Amodei op-ed: The ‘pacing’ debate that directly split the AI industry
On September 12, Amodei posted on his blog, “We Must Pace the Frontier.” He said the performance of the most advanced models is being upgraded at a pace that safety research cannot keep up with, and he laid out a three-step roadmap to rein it in.
The first step is a ‘third-party overseer inside.’ Regulators or independent evaluation teams receive desks and access badges inside the buildings of frontier research labs, live on-site, and monitor and assess in real time, with the same level of access to models and infrastructure as employees. Anthropic has said it will voluntarily introduce this measure first in its own organization.
As evidence supporting his claim, the so-called ‘central proof’ presented was an incident in which many OpenAI agents collectively attacked systems that were not supposed to be targeted, and then attempted to hack the evaluation system that had been scoring their own performance.
**Sam Altman** and **Elon Musk** publicly backed this argument over the weekend, and **Mustafa Suleyman** of Microsoft also joined the following Monday. However, he drew a line, saying, “This isn’t about stopping development—it’s about moving at a more cautious, managed pace.”
Also read: XRP Ledger accumulates 3,254 transactions in a single block…sets a new all-time high
Munsat·DeepSeek ‘no break’…China outside the slowdown discussion
As Amodei’s proposal spread, semiconductor and AI-related stocks fell across both the U.S. and Chinese markets. Deutsche Bank analysts, in a report, raised doubts about the likelihood of actual implementation, saying, “No company that has a competitor going full speed would voluntarily step back.”
Amodei himself acknowledged the dilemma in an op-ed. The logic is that once the premise that U.S. research institutes maintain technological gaps breaks down, Beijing-based projects that don’t slow their pace will quickly move to the front and increase national security risks. In a CBS interview on Sunday, he said China is the hardest variable in this plan.
In response, President Donald Trump, who was visiting Ireland, met with reporters and dismissed the issue itself, saying, “The U.S. is already ahead.”
Beijing moved to rebut within a day. **Guo Jiakun**, a spokesperson for China’s Ministry of Foreign Affairs, criticized the claims in a statement, saying, “Fanning fear and engaging in malicious competition only harms global AI governance.” Xiao Qian of Tsinghua University also attacked, saying Amodei is using ‘pacing’ as a pretext to keep out cheap Chinese open-source competitors.
China research institutes accelerate after ‘Kimi’…U.S. aims to widen the gap via export controls
The three-step plan proposed by Amodei is aimed only at U.S. and Western frontier research institutions and has no binding effect on Chinese competitors. In fact, not a single Chinese company has voluntarily agreed to slow its pace.
Amodei cited a recommendation document released by U.S. security authorities on September 8. In that announcement, the U.S. Cybersecurity and Infrastructure Security Agency (CISA) and others pointed to six Chinese companies as having ‘distilled’ U.S. models at an industrial scale, and among them were also DeepSeek and Munsat AI. China’s Ministry of Commerce rejected this as “absurd allegations,” and hit back, saying Washington is politicizing normal commercial activity.
In July, Munsat shot claimed that by unveiling its ‘Kimi K3’, it had implemented performance close to Anthropic’s flagship at an extremely low cost. Since then, China-based open-weight model makers have taken the top ranks in the developer traffic on the routing marketplace OpenRouter, growing their presence.
Amodei argues that over the next 3 to 5 years, if the U.S. is to maintain and expand its lead, it must significantly strengthen enforcement and crackdowns on model distillation, along with tightening export controls on advanced semiconductors.
Next reading: Grok 5 rises as Musk’s AGI goal…AI safety warnings grow even stronger