Over the weekend, Anthropic CEO Dario Amodei called for slowing—not stopping—frontier AI so safety can catch up. Elon Musk, Sam Altman, and Demis Hassabis publicly agreed. That settles one question from Friday: the leaders agree AI should be paced. Their support for pacing varied, but the decision now moves from whether rivals can agree to what they will actually do—and whose terms could govern a global plan that depends on China.

Amodei said Anthropic would give independent evaluators access comparable to employees who assess risk, and Altman said OpenAI would do the same. The shared commitment could give the public and policymakers a view into safety decisions while systems are still being built, instead of leaving companies to describe their own work afterward. Neither company has named the evaluators or said when access will begin, so the public cannot yet judge how the promise will work in practice.

Amodei’s framework asks China to join a global pacing effort while also calling for measures designed to protect and widen the United States’ lead in advanced AI. Chinese officials called for international AI cooperation, while state media described his framework as containment. China is not rejecting cooperation; it is rejecting terms that ask it to help pace the race while accepting second place.

AI-related stocks fell around the world following the warnings, amid other financial pressures. The selloff did not prove that a slowdown is coming, but the effects of slower AI spending could spread through the chip, data-center, and energy industries that have expanded around the AI boom and into the wider economy. For businesses and investors counting on rapid AI gains, that makes the pace of AI development an economic issue before any company has actually slowed down.

Now that American AI leaders agree that frontier development should be paced, whose terms will govern any global agreement with China? Right now, there are no shared terms. The executives have not announced a joint slowdown, shared safety limits, or an international framework China has accepted. How much their agreement can change depends on what each leader actually supported and what the companies have promised to do.

Frontier AI pacing means deliberately slowing the development of increasingly capable AI systems when necessary to give safety, alignment, monitoring, and oversight more time to keep up.

Amodei says pacing means letting frontier AI capabilities advance more slowly so alignment and safeguards have time to keep up, while AI’s benefits continue. It does not mean stopping training or technical progress. The extra time only helps if companies and governments use it to strengthen safety and make development more visible to outsiders.

Amodei’s three steps move from oversight inside one company to agreements across countries. First, independent evaluators would receive employee-like access inside individual AI companies. That would allow them to examine safety work while models are still being developed. Second, companies and governments in democratic countries would agree on common standards and limits, with regulation covering companies that do not participate voluntarily. Third, governments would pursue global coordination, especially with China, because companies in one group of countries cannot slow far beyond competitors that continue moving at full speed. The three steps are meant to make company promises visible, bring reluctant competitors under the same rules, and keep international competition from undoing the entire effort.

Within roughly nine hours of Amodei’s post, Musk, Altman, and Hassabis all backed Amodei’s call to pace frontier AI. That was a rare point of agreement among four competing labs, but their responses did not carry the same weight. Musk’s entire public response was “Dario is right,” which supports Amodei’s direction but says nothing about what xAI would do. Hassabis said Amodei was pointing toward the right path, but added that the details needed work and connected his support to Google DeepMind’s July proposal for an industry standards body. Altman said he agreed that frontier AI should be paced, explained that OpenAI had been discussing the issue in recent weeks, and promised that the company would also give independent evaluators employee-like access.

Anthropic says its evaluators would be able to check whether the company follows its safety commitments, report incidents, and examine training processes and development tools instead of seeing only a finished model. They would also be allowed to publish key findings without Anthropic editing them, apart from narrow redactions for sensitive information. If a redaction affected their conclusions, they would have to disclose that. OpenAI said more details would follow, so it has not yet explained whether its evaluators would receive all of the same access and publishing rights.

Amodei and Altman both say government rules will be needed to make pacing apply across the industry. Amodei argues that regulation is the most effective way to cover American companies that will not participate voluntarily. Altman similarly welcomed a federal framework with consistent requirements for frontier AI safety. At the same time, both said companies should begin the work they can do themselves rather than wait for legislation.

Altman said OpenAI now prepares a safety case—an explanation of how it plans to manage the risks—before beginning a frontier reinforcement-learning run expected to make a model significantly more capable. OpenAI is therefore examining risk before capability-increasing training begins, rather than focusing only on a completed model before release. Those earlier checks and continued monitoring take time and money, which is why Altman says pacing should make progress slower than it otherwise would be.

The weekend agreement followed a two-month shift from employee advocacy to company positions and then executive support. In July, 1,386 employees at frontier AI companies personally called on the U.S. government to support an international effort to develop tools for deliberately pacing automated AI development. Because the signatures were personal, the statement showed support inside the labs without committing their employers. Anthropic took the next step in August by saying industry-wide pacing would require companies and governments to create a lawful, verifiable, and effective system. OpenAI chief scientist Jakub Pachocki added a similar argument from inside OpenAI on September 6. He said no lab had solved alignment and monitoring well enough to continue scaling at maximum speed much longer. He therefore supported voluntary slowdowns while assigning international coordination to governments. The weekend endorsements moved the idea another step forward by bringing the leaders of competing companies into public agreement.

The four leaders have not agreed on rules that would tell a lab when to slow, what it must restrict, or who would enforce the decision. That means no company is bound to slow a specific project, and every lab still faces pressure to keep up with its rivals. The Wall Street Journal captured that tension by noting that the same companies calling for restraint continue to spend enormous sums on building more capable systems. The weekend changed what their leaders say in public, but it has not yet changed how the competition works. Amodei’s proposal goes further than the other responses, and it makes preserving America’s technological advantage part of his route to global pacing.

Amodei argues that democratic countries can slow only as much as their lead over China allows. If American companies move more slowly while Chinese projects continue at full speed, he believes China could catch up or pull ahead, creating what he views as a national-security risk. His answer is to protect and widen the American lead so U.S. companies have more room to pace development without losing first place.

Amodei calls for limiting the computing power China can obtain by blocking sales of powerful AI chips and semiconductor-manufacturing equipment. He also wants stronger enforcement against chip smuggling and against Chinese companies remotely using data centers outside China. His remaining proposals target access to American AI itself: action against unauthorized distillation by companies in authoritarian countries and stronger security to prevent the theft of model weights. Together, the measures are meant to limit both the computing resources China can use and the American model capabilities it can acquire.

Amodei argues that these restrictions could significantly widen America’s lead over the next three to five years, giving the United States more leverage to negotiate an eventual pacing agreement with China. His strategy is to make it harder for China to catch up first, then negotiate from a stronger American position. Whether China would accept that logic is another matter.

Amodei says global pacing requires China, but he offers several possible levels of cooperation depending on what the two countries could realistically agree to. The narrowest option would prohibit specific dangerous uses, such as using AI to help create biological weapons. A broader agreement would require both sides to test models for serious cyber, biological, and alignment risks, possibly through an international standards body. A stronger agreement would limit how quickly AI can be used to improve AI because that process could make capabilities advance faster than people can understand or manage them. The most ambitious version would place a broader speed limit on frontier development or pause it altogether, giving safety work more time to catch up if capabilities begin advancing too quickly.

Amodei considers a broad limit or pause unlikely in the near term because either country could secretly continue advancing while the other obeyed the agreement. The country that continued could then gain a decisive lead, creating the same security risk the agreement was supposed to reduce. That makes the strongest form of pacing the hardest to verify and the most dangerous for either side to enter without trusting the other.

Altman, Hassabis, and Musk did not publicly endorse Amodei’s China restrictions or his complete international framework. Altman separately warned that too much AI power concentrated in one country, laboratory, company, or individual could produce a dystopian outcome. That supports the concern that no single actor should control advanced AI, but it does not establish that Altman accepts Amodei’s strategy for keeping China behind. Chinese officials answered that proposed bargaining position by challenging its terms rather than dismissing the risks of advanced AI.

China’s Foreign Ministry answered Amodei by calling his warnings about China counterproductive. Spokesperson Guo Jiakun said fearmongering, confrontation, and vicious competition would disrupt global AI governance and benefit no one. He also called for countries to work together on AI. Beijing’s official response therefore supports international cooperation while rejecting the way Amodei proposes to pursue it.

The state-run Global Times went further by describing Amodei’s proposal as a Cold War-style plan to contain China’s technological development. That response followed China’s rejection of American accusations that Chinese companies had improperly copied capabilities from U.S. models. China instead accused the United States of seeking an AI monopoly. Within that wider dispute, Beijing sees restrictions on chips, computing power, and American model capabilities as an attempt to control who can compete, even when those restrictions are presented as part of an AI-safety plan.

Chen Yixin, the head of China’s Ministry of State Security, has also warned that advanced AI in the hands of hostile foreign actors could threaten China’s political, institutional, and ideological security. He pointed to American AI systems that can find software vulnerabilities and perform complex hacking tasks, arguing that they could be used against critical infrastructure. He also warned that foreign intelligence services could use AI to obtain sensitive data, trade secrets, and personal information. Those are Chen’s official warnings rather than evidence of specific attacks, but they show that China also views advanced AI as a national-security threat.

Amodei makes a similar security argument from the American side. He says a Chinese lead in advanced AI would create a grave danger for the United States and other democracies, even if Chinese systems avoided the alignment failures he fears. Asia Society Policy Institute fellow Lizzi C. Lee said both countries want to stop the other from gaining a technological chokepoint. That helps explain the contradiction Beijing sees in being asked to cooperate on frontier safety while the United States restricts China’s access to frontier computing. Both sides describe AI as a security problem, but each also sees the other side’s technological power as part of that problem.

President Donald Trump responded to the executives’ warnings by pushing back against calls for his administration to slow AI development. With public concern growing over AI misuse and data-center construction, he described the opposition surrounding AI and data centers as a “sick conspiracy” that could allow China to overtake the United States. He also emphasized that the United States currently leads China and said whoever wins AI wins. That logic fits his America First message: continued speed protects the American lead, while slowing gives China an opening. The leaders building frontier AI are reaching the opposite conclusion, warning that moving too fast could allow capabilities to outrun safety. That leaves the Trump administration at odds with the people leading the technology’s development.

AI leaders are now calling for restraint inside an international competition that both governments view as a security contest. That tension is reaching financial markets as investors consider whether new safety commitments could slow the spending and growth built around continued AI progress.

AI-related stocks fell across the United States, Europe, and Asia following the executives’ warnings, even though no shared slowdown had begun. The Philadelphia semiconductor index fell 5.2%, while Nvidia lost 3%, putting some of the steepest pressure on companies that supply the computing power behind AI development. Europe’s technology sector fell 2.2%, and SoftBank dropped more than 10% in Japan. The warnings were one of several pressures on the market, but the global reaction showed that the possibility of slower AI progress can affect stock prices long before it changes a training run or product release.

Stock prices include what investors believe a company will earn in the future. AI-related companies have gained value partly because investors expect laboratories and cloud providers to keep buying chips, building data centers, and expanding the energy infrastructure needed to run them. Investors are therefore paying today for profits they expect that spending to produce later. If pacing delays those purchases or the products they support, expected sales and profits can fall, causing stocks to lose value before a company cancels a single project.

Some businesses are also relying more heavily on debt and financing arrangements between companies that buy from one another to support expensive AI projects while borrowing costs remain high. That debt increases the risk because companies must continue paying interest even if development takes longer and the expected revenue arrives later. Investors could then begin questioning whether AI projects will earn enough money to justify what companies spent to build them.

If laboratories respond to safety concerns by developing models more slowly, they may need fewer chips or wait longer to buy them. Delayed data centers would reduce demand for construction, computing equipment, and energy infrastructure, while later product releases would make businesses wait longer for revenue from new AI services. Interactive Brokers analyst Steve Sosnick said a slowdown and reconsideration of AI spending could therefore affect important market sectors and the economy. Companies have not announced those cuts, but his warning explains how slower development could spread beyond the AI laboratories themselves.

Investor Michael Burry, best known for betting against the U.S. housing market before the 2008 financial crisis, offered a more skeptical explanation for the executives’ warnings. He called them “hype and puffery” and said they could provide “cover for real uncontrollable slowing growth.” In other words, he believes the safety argument could give the industry a convenient explanation if the AI boom is already losing momentum for business reasons. Deutsche Bank took a different view of the spending outlook, saying intense competition among companies and countries made a near-term slowdown unlikely. Their comments offer two interpretations of the market reaction: Burry sees signs of underlying weakness, while Deutsche Bank expects the competitive race to keep companies spending.

The selloff alone does not show that an AI bubble is bursting. It shows that AI stock prices are sensitive to anything that threatens the rapid spending and growth investors already expect. Evidence of a broader collapse would become stronger if companies began cutting projects or failing to earn the returns used to justify their valuations. Neither has been established by this market reaction.

Altman indicated that OpenAI would not proceed with an initial public offering this year amid safety concerns. That matters because an IPO would ask public investors to decide how much OpenAI is worth while the company is warning that safety work may make progress slower and more expensive. Anthropic was taking a different path: Reuters reported that the company was continuing toward a public offering expected the following month and had discussed Nvidia as a possible anchor investor, according to sources. The plans could still change, but the contrast shows that laboratories supporting pacing are making different financial decisions while continuing to raise the enormous amounts of money needed to compete.

For people outside the market, the consequences depend on what happens to actual AI spending. If development continues at full speed, spending on chips, data centers, energy projects, and related jobs could keep growing, but company leaders warn that safety work could fall further behind. Communities already opposing data-center construction would also face continued expansion. If development slows, researchers and outside evaluators could have more time to identify and reduce risks, and some pressure from data-center construction could ease. However, projects, revenue, and jobs tied to the AI expansion could also be delayed. The public is therefore being asked to weigh the risks of moving too quickly against the economic costs of slowing down. Doing that requires more than promises: people need to know which safety decisions would actually change and whether anyone outside the laboratories can verify them.

For the first time, leaders of four competing frontier AI companies—Anthropic, OpenAI, Google DeepMind, and xAI—are publicly saying that winning the AI race should not require developing the technology as fast as possible. That matters because the people building some of the world’s most advanced AI systems are acknowledging that there may be times when protecting people requires giving up some speed. The question now is whether that principle will actually change what their companies do.

People would have to live with the consequences of continuing frontier AI development at full speed or slowing it so safety work can catch up, even though companies and governments will make that choice for them. Moving at full speed could bring new AI products, investment, and economic benefits sooner while giving safety teams less time to understand new capabilities before they spread. Slower development could give researchers more time to identify and address dangerous behavior, but it could also delay some of those economic benefits. None of this proves that AI will cause human extinction, that catastrophe is imminent, or that today’s systems have reached the outcomes researchers fear. It means the public is being asked to accept development risk in exchange for faster progress while companies still decide how much risk is too much.

Independent evaluators could make those decisions more visible by examining safety work during development and publishing findings that would otherwise remain inside the companies. Under the commitments announced so far, however, evaluators have not been given authority to stop training or block a release. Future agreements or government rules could give outside evaluators or regulators additional authority. For now, the commitments are primarily about giving outsiders greater visibility into how companies make safety decisions.

The agreement will become more consequential if an evaluator’s findings cause a laboratory to change or delay development, government rules make safety limits mandatory, or the United States and China accept terms both sides will follow. China’s rejection of Anthropic CEO Dario Amodei’s current terms makes international coordination especially difficult because American laboratories will remain under competitive pressure if Chinese developers continue advancing at full speed.

The industry has moved the debate forward without resolving its central conflict. American AI leaders are beginning to agree that frontier development may need to slow, but global pacing ultimately requires cooperation with China while Anthropic CEO Dario Amodei’s proposal—the most developed framework so far—also seeks to preserve and widen the U.S. technological lead.

Public agreement is only the beginning. The real test comes when slowing down means risking the very thing every company and country in the AI race is trying to avoid: falling behind.

Q: What does “pacing” frontier AI mean?
A: Pacing frontier AI means deliberately slowing the development of increasingly capable AI systems when safety, alignment, monitoring, or oversight need more time to keep up. Dario Amodei’s proposal does not call for stopping AI development altogether. It would allow capabilities to continue advancing while creating additional time to understand and reduce potential risks.

Q: Are AI companies actually slowing down?
A: Anthropic, OpenAI, Google DeepMind, and xAI have not announced a coordinated slowdown in AI development. Their leaders have publicly supported the idea that frontier AI may need to be paced, but they have not agreed on shared limits, specific conditions that would trigger a slowdown, or a system for enforcing those decisions.

Q: What could actually cause an AI company to slow development?
A: A frontier AI company could choose to slow development if safety testing or monitoring showed that new capabilities were advancing faster than the company could understand or manage their risks, but the four companies have not agreed on common triggers requiring them to do so. OpenAI says it now prepares a safety case before beginning certain frontier reinforcement-learning runs expected to significantly increase model capabilities, while Amodei argues that companies need enough time for safeguards and alignment work to keep pace with those advances.

Q: What would independent AI evaluators actually do?
A: Independent AI evaluators could examine safety work while frontier models are still being developed rather than seeing only a finished system. Anthropic says its evaluators would be able to examine training processes, development tools, incidents, and compliance with safety commitments and publish key findings. OpenAI has also promised employee-like evaluator access but has not yet provided the same level of detail. Under the commitments announced so far, evaluators have not been given authority to stop training or block a model release.

Q: Would the government have to regulate an AI slowdown?
A: Government regulation could be needed to make frontier AI pacing apply across the industry because voluntary agreements would not bind companies that choose not to participate. Amodei argues that regulation is the most effective way to cover U.S. companies that reject voluntary limits, while Altman has called for a consistent federal framework for frontier AI safety. Both also say companies should begin taking safety measures without waiting for legislation.

Q: Why does slowing AI require China to cooperate?
A: Global AI pacing would require cooperation with China because U.S. companies could lose their technological lead if they slowed significantly while Chinese developers continued advancing at full speed. Amodei proposes protecting and widening the U.S. lead before negotiating broader pacing agreements with China. Beijing supports international AI cooperation but rejects an approach that combines calls for cooperation with restrictions designed to limit China's access to advanced AI technology.

Q: Why would a global AI slowdown be so hard to enforce?
A: A global AI slowdown would be difficult to enforce because countries would need confidence that competitors were obeying the same limits instead of secretly continuing to develop more powerful systems. Amodei considers broad limits or pauses especially difficult because a country that secretly continued advancing could gain a major strategic advantage while its competitor complied. That makes verification and trust central problems for any international pacing agreement.

Q: Could slowing AI development affect the economy?
A: Slower frontier AI development could affect demand for chips, data centers, energy infrastructure, investment, and jobs tied to the AI expansion. AI-related stocks fell across the United States, Europe, and Asia after executives raised the possibility of slower development, although no coordinated slowdown has begun and the warnings were only one of several pressures affecting markets. If pacing eventually delays AI spending or new products, the economic effects could extend well beyond the AI laboratories themselves.

Editor’s Note: This article was created by Alicia Shapiro, CMO of AiNews.com, with writing support, AEO/GEO/SEO optimization, image concept development, and editorial structuring support from ChatGPT, an AI assistant. All final editorial decisions, perspectives, and publishing choices were made by Alicia Shapiro.