The AI Pace Debate and the New Cold War Frame: Geopolitical Implications of Anthropic's Safety Discourse

Executive Summary

Anthropic CEO Dario Amodei's call to slow the pace of AI development has transformed into a new front in the U.S.-China tech rivalry, as the safety discourse became intertwined with demands for stronger chip export controls against China. The Chinese Ministry of Foreign Affairs and state-run media immediately countered, framing the move as a Cold War tactic driven by commercial interests, while the Trump administration rejected domestic calls for tighter regulation, citing the need to maintain a competitive edge over China. Ahead of the September 24 U.S.-China summit, it is clear that the language of 'safety' is being used by both countries not as a genuine agenda for cooperation, but as a bargaining chip and a tool to secure a competitive edge. Future developments could follow one of three paths: the institutionalization of safety dialogue, a continued gap between discourse and implementation, or the acceleration of governance fragmentation and technological bloc formation. The most likely path is the continued gap between discourse and implementation. South Korea will need a dual-track approach: avoiding the wholesale adoption of either the U.S. or Chinese narrative, managing security-linked and commercial sectors separately, and simultaneously building its own risk management capabilities.

I. Analysis of the Current Situation

The AI Pace Debate: Anthropic CEO's Call for a Slowdown and the U.S.-China New Cold War Frame

1. Background and Timeline

The flashpoint was a blog post by Anthropic CEO Dario Amodei on Saturday, September 12 (local time) [3][9]. The title was "We Need to Moderate the Pace of Frontier Development" [3]. In the post, Amodei argued for slowing the rate of improvement in AI models [10]. He warned that swarms of autonomous AI agents could take over the entire internet via botnets within 6 to 12 months, causing hundreds of billions of dollars in damages [4][11].

This warning came just days after researcher Jacob Coxson, who had left OpenAI to join Anthropic, departed the industry, criticizing U.S. companies for "gambling with our lives" [15]. The combination of incidents involving loss of control over autonomous AI and the departure of a key researcher lent empirical weight to the safety discourse [3]. Shortly after Amodei's post, Sam Altman (OpenAI), Demis Hassabis (Google DeepMind), and Elon Musk (xAI) successively voiced their agreement [2][16]. The unified stance from the heads of competing frontier model developers was seen as an unusual development by the U.S. media [10][16].

The problem was that Amodei's proposal did not stop at the safety discourse. In the same post, he also called for strengthening AI chip export restrictions targeting China [7][17]. Safety regulation and the containment of China were presented as a single package. At this point, the debate immediately shifted to the framework of U.S.-China technological rivalry.

2. Current Situation

China's Ministry of Foreign Affairs immediately issued a rebuttal. Spokesperson Guo Jiacun, taking over the briefing from Wang Wenbin, stated, "Fear-mongering, confrontation, and malicious competition will only hinder progress in global AI governance and serve no one's interests" [14]. He added that AI development is a matter concerning the common welfare of humanity [14]. The state-run Global Times went a step further, characterizing Amodei's remarks as a product of "hostile and baseless" commercial interests [17]. An editorial from a People's Daily affiliate, cited by the Spanish-language outlet 14ymedio, described it as certain elements in the U.S. tech sector pulling out the "Cold War playbook" to target Chinese AI [18].

Reactions within the United States are mixed. President Trump has repeatedly drawn a line against strengthening AI regulations, citing competition with China [1]. South Korean newspaper The Hankyoreh noted that the call for a slowdown is unwelcome on both sides: China sees it as "kicking away the ladder" to raise barriers to entry, while the U.S. fears it could allow China to catch up [1]. Al Jazeera dubbed the situation a "silent cold war," analyzing that the demand to slow down AI has, paradoxically, drawn a new technological battle line between the U.S. and China [7].

The timing of this debate, erupting just before the scheduled September 24 U.S.-China summit, gives it added weight. The two countries had already agreed to launch an intergovernmental dialogue on AI at their May summit in Beijing, and the upcoming Washington meeting is intended as a follow-up [6]. However, the Trump administration is treating safety discussions solely within the framework of maintaining a competitive edge over China and has squarely rejected the industry's calls for a slowdown. Meanwhile, China is attempting to shift the discussion toward the right to set global rules by insisting that the very definition of safety must first be agreed upon [6]. With the White House's cooperation track and Congress and regulatory agencies' pressure track already operating separately, it is highly likely that individual regulations like semiconductor and AI export controls will continue independently, even if a symbolic agreement is reached at the summit [6].

3. Key Actors and Positions

Anthropic (Dario Amodei)is the epicenter of the debate, having simultaneously raised safety concerns and arguments for containing China. Amodei later sought to manage the fallout from the escalating controversy by clarifying that his company had never advocated for a complete ban on open-weight models [5]. In a CBS interview, he also appeared to moderate his tone, stating, "There's no need to panic today" [13].

OpenAI, Google DeepMind, and xAIhave echoed the call for a slowdown, but their policy interests are not perfectly aligned. OpenAI has conveyed to the U.S. government that "premature regulation" of open-weight models should be avoided, based on the logic that regulation could push innovation abroad, ultimately benefiting China [5]. This shows that while they join the safety discourse, their interests diverge when it comes to the methods and intensity of regulation.

The Trump Administrationis effectively rejecting calls to strengthen AI regulation, countering with concerns about undermining competitiveness against China [1]. Even ahead of the summit, it maintains a stance of treating safety dialogue as a subordinate agenda to the primary goal of maintaining a competitive edge [6].

The Chinese Governmentis responding through a combination of official criticism from the Ministry of Foreign Affairs and editorial critiques in state-run media. Spokesperson Guo Jiacun's briefing framed the U.S.-led safety discourse as an act of obstruction to governance [14], while editorials in the Global Times and People's Daily affiliates framed it as a Cold War-style containment attempt based on commercial interests [17][18]. From China's perspective, the real threat is not the safety discussion itself, but the underlying demand for semiconductor export controls.

4. Core Issues

The core of this issue is that the safety discourse is inseparable from the logic of industrial competition. Amodei's proposal was presented in the universal language of existential risk to humanity, but it simultaneously contained a measure aimed at containing a specific country: restricting chip exports to China [7][17]. Citing this, China is attempting to frame the safety discourse itself as a cover for a containment strategy, thereby shifting the debate into a contest over the legitimacy of international AI governance discussions [14][18].

Even within the United States, the political support for a slowdown is shallow. Under the Trump administration's deregulatory, competition-first stance, the likelihood of substantive federal regulation is low, and actual measures have been limited to self-regulation, such as the introduction of third-party evaluation organizations [3]. This is why the gap between rhetoric and implementation is expected to persist for a considerable time [3].

Another point to consider is the possibility that safety dialogue and individual export control measures will be managed on separate tracks, as this debate coincides with the summit. Because the two countries attach different political objectives to the term 'safety,' it is highly probable that even if a symbolic agreement is reached at the summit, substantive regulations like semiconductor and AI export controls will continue on a separate track [6]. With the Taiwan issue already solidified as a structural red line, the AI safety discussion is likely to be treated as a secondary issue layered on top of, rather than alleviating, the fundamental conflict structure of U.S.-China relations [6].

II. In-Depth Analysis

The AI Pace Debate: Anthropic CEO's Call for a Slowdown and the U.S.-China New Cold War Frame

1. Analysis of the Root Causes

The ostensible cause of this debate is safety concerns. However, the true nature of the issue is revealed only when we examine why this safety discourse emerged at this particular moment and why it was presented in tandem with calls for chip export controls against China. Amodei's September 12 post cited empirical risk signals, such as the loss of control over autonomous AI agents and the departure of a key researcher [3][15]. Yet, in the same text, he also urged for strengthening AI chip export restrictions targeting China [7][17]. In effect, safety and geopolitics were presented as an inseparable policy package.

Behind this combination lies Anthropic's market position. Anthropic is a latecomer, trailing OpenAI and Google DeepMind in commercial scale. For a company that has built its identity around safety to lead the charge for a 'slowdown' while simultaneously calling for export controls that would exclude its Chinese competitors can be seen, as the Global Times pointed out, as a structure where commercial interests and the safety discourse are intertwined [17]. The Chinese state media's characterization of this as a "hostile and baseless" remark was a direct response targeting this combined structure [17].

President Trump's immediate rejection of calls for stronger safety regulations can be read in the same context [1]. From the White House's perspective, there is a significant concern that strengthening AI regulation could result in slowing down in the competition with China [1]. Ultimately, the root cause is that the language of 'safety' is being used as a tool to secure a competitive edge, both within the U.S. and in U.S.-China relations.

2. Structural Context

Political Structure: The Dual Use of the Safety Discourse

The reason this debate did not end as a simple safety dispute between companies is that it coincided with a political event: the U.S.-China summit. The two countries had already agreed to launch an intergovernmental dialogue on AI at their May summit in Beijing, and the September 24 Washington meeting was scheduled as a follow-up [6]. Amodei's remarks at this juncture created a dynamic where the safety discourse was mobilized as a bargaining chip for both the U.S. and China ahead of the summit.

The problem is that the two countries attach different political objectives to the word 'safety.' The Trump administration has treated "safety discussions solely within the framework of maintaining a competitive edge over China" and squarely rejected the industry's calls for a slowdown [6]. China, by insisting that "the very definition of safety must first be agreed upon," is attempting to "shift the discussion of technological risk into a matter of the right to set global rules" [6]. Spokesperson Guo Jiacun's emphasis that "AI development is a matter concerning the common welfare of humanity" [14] can also be seen as an attempt to redefine safety as a universal governance issue, not a tool for one country's advantage. In this structure, with the White House's cooperation track and Congress and regulatory agencies' pressure track already operating separately [6], it is highly likely that individual regulations like semiconductor and AI export controls will continue independently, even if a symbolic agreement is reached at the summit [6].

Economic Structure: Export Controls and Supply Chain Realignment

Amodei's call for chip export restrictions is an extension of the U.S. system of semiconductor export controls against China that has been in place since 2022. It is unusual for a U.S. AI company to voluntarily call for stronger government controls. Typically, the industry has opposed measures that reduce market access, but this time, frontier model developers have been more vocal than the government in demanding tighter regulations. This is not unrelated to the fact that for companies that have already secured a supply chain advantage, export controls have the effect of excluding competitors. On the other hand, there are also voices within the U.S. industry arguing that "premature regulation" of open-weight models should be avoided. The logic is that if regulations are tightened, innovation could move offshore, and China could fill the void [5]. Anthropic has deliberately distanced itself from this debate, but Amodei himself had to issue a separate post clarifying that he "never argued for a ban on open-weight models" [5]. This indicates that interests regarding export controls and openness are divided even within the U.S. AI industry.

Security Structure: The Securitization of Private-Sector Discourse

A prominent structural feature of this incident is how a blog post by a private company's CEO was instantly transformed into an international security discourse. Al Jazeera's labeling of this as a "silent cold war" [7] encapsulates the fact that a private actor's statement, not an official government announcement or treaty, became the trigger for drawing new lines of confrontation between nations. This illustrates the dual-use nature of the AI industry and the current power structure where a few Big Tech companies are directly involved in shaping national security discourse. The fact that misuse of AI by state-backed actors has already been empirically reported [3] is another factor accelerating this securitization.

3. Historical Precedents and Comparative Cases

This incident has structural parallels with the nuclear technology control debates of the Cold War era. Then, too, the technologically superior bloc sought to block the opposing bloc's access under the pretext of non-proliferation, while the other side countered by framing it as a hegemonic strategy to stifle technological progress. A key difference, however, is that unlike nuclear technology, AI lacks a clear physical proliferation pathway, making it difficult to specifically target open-source models and algorithms for control. Consequently, the effectiveness of export controls is inherently more limited than that of the nuclear non-proliferation regime.

A more recent precedent is the escalating semiconductor export control situation since 2018. A series of measures, including sanctions on Huawei and restrictions on advanced foundry equipment exports, were introduced under the justification of "security concerns," and China has consistently rebutted them as measures to "hinder technological progress" [1]. The language used by China in the current AI pace debate—such as "fear-mongering" and "Cold War playbook" [14][18]—follows the same pattern as its recurring responses during past semiconductor control episodes. The difference this time is that the flashpoint was a private company CEO, not the government, and the moral justification of 'safety' was brought to the forefront.

Another point of comparison is the AI Safety Summit discussions that have been ongoing since 2023. Although various countries have pledged cooperation on AI safety since the Bletchley Park summit in the UK, implementation has tended to be subordinated to the logic of protecting domestic industries and securing a competitive advantage. In the current case as well, actual measures are "limited to self-regulation, such as the introduction of third-party evaluation organizations," and pressure for federal-level regulation in the U.S. remains low [3]. This follows a recurring pattern of a gap between declarative agreements and substantive implementation, similar to past attempts at forming international technology norms.

4. Key Variables Shaping Developments

The first variable is the outcome of the September 24 U.S.-China summit. Even if AI safety dialogue is discussed as a summit agenda item, it is "likely to be treated as a secondary outcome of the summit" [6]. With the Taiwan issue already solidified as a "structural red line" [6], it is difficult for the AI safety discussion to gain independent significance. It is judged that even if a symbolic joint statement is issued at the summit, individual regulatory measures such as semiconductor and AI export controls will continue on a separate track [6].

The second variable is the internal division within the U.S. industry. The U.S. foreign policy stance could change depending on how the demands for a slowdown and stronger export controls from the Anthropic camp are reconciled with the opposition to "premature regulation" from the open-weight model camp [5]. As long as the Trump administration maintains its deregulatory stance [3], it is unlikely that industry-led calls for regulation will translate into federal policy. However, this balance could be shaken if further real-world incidents involving autonomous AI agents occur.

The third variable is China's response. So far, China has limited its response to verbal rebuttals through Ministry of Foreign Affairs briefings and state-media editorials [14][17][18]. However, if U.S. export controls are actually strengthened, China may accelerate the development of its own AI ecosystem or take retaliatory measures. This could serve as a catalyst for the global AI supply chain and standards system to solidify into a bifurcated U.S.-China structure.

The fourth variable is whether the risks of autonomous AI agents actually materialize. If the scenario Amodei warned of—an "internet takeover via botnets within 6 to 12 months" [4][11]—is realized, or if a comparable incident occurs, it could trigger a shift from the current self-regulatory safety discussions to a mandatory regulatory regime. Conversely, if no such incident occurs, this debate may be recorded as a case where commercial interests were cloaked in the guise of safety discourse and eventually fizzle out.