China Rejects AI Slowdown Calls: What It Means for the Global AI Race

A single essay from an AI CEO just shattered any remaining illusion of cooperative global AI development. Beijing’s response made the fault lines permanent.

Anthropic CEO Dario Amodei published “We Must Pace The Frontier” in September 2026, calling for deliberate slowdowns in frontier model development, independent safety evaluators inside AI labs, and tighter U.S. export controls on advanced AI to China. Sam Altman and Elon Musk both signed on. China’s Foreign Ministry called it “fearmongering” and a “Cold War playbook.” Trump said “whoever wins AI wins.”

That’s the full spectrum of the debate — in three sentences. The question worth unpacking is what this breakdown actually means for the global AI race in 2026 and beyond.

In brief: China’s rejection of AI pacing proposals isn’t primarily about safety disagreement — it’s about structural competitive advantage. A voluntary slowdown preserves whoever currently holds the lead, and both sides know it.

Three things make this moment particularly sharp:

1. Chinese companies including DeepSeek and Alibaba now deliver competitive AI performance at significantly lower inference costs than U.S. counterparts, narrowing the capability gap that pacing would have frozen in place.
2. China’s State Security Minister Chen Yixin simultaneously acknowledged real AI security risks while publicly rejecting Western-led governance frameworks — a revealing contradiction that shows Beijing isn’t dismissing safety, just whose safety framework it’ll accept.
3. A Trump-Xi summit scheduled for September 24, 2026 is expected to address AI governance directly, making the next two weeks unusually consequential for how this standoff evolves.

How We Got to This Breaking Point

The AI export control timeline matters here. The U.S. started restricting advanced chip sales to China — specifically NVIDIA A100 and H100 GPUs — back in October 2022. The goal was to slow Chinese frontier model development by cutting off compute access. It didn’t work as planned.

DeepSeek’s R1 model, released in early 2025, demonstrated that Chinese labs could train highly capable models at a fraction of the cost U.S. labs reported spending. The reaction in Silicon Valley was visceral — and it recalibrated assumptions about who actually held the lead.

Fast-forward to September 2026. According to NBC News, U.S. agencies have accused Chinese firms including DeepSeek and Alibaba of systematically copying American AI models through “distillation” — training smaller models using outputs from more powerful ones. Beijing denied the allegations. But the accusation itself signals how rattled U.S. labs are by Chinese progress.

Into this environment, Amodei drops a proposal that includes restricting U.S. tech sales to China. Chinese state media’s Global Times immediately characterized it as a “Cold War playbook.” That framing wasn’t accidental — it’s a deliberate recast of a technical safety debate as economic protectionism.

The timeline compression here is striking. In 18 months, the conversation shifted from “China is years behind” to “China is systematically copying our best models because they’re competitive enough to be worth copying.” That context is essential for reading what comes next.

The Pacing Trap: Who Benefits From Slowing Down?

Amodei’s proposal has genuine safety intent — independent evaluators, coordinated standards, international cooperation on risk. These aren’t bad ideas. The export control component, though, undermines the neutrality claim.

According to the South China Morning Post, Chinese researchers framed the core problem precisely: a development pause preserves existing capability gaps. Whoever holds the lead when the music stops keeps that advantage — potentially permanently, given how compounding AI capabilities work.

This is a structurally sound critique. Pacing agreements in technology markets have historically benefited incumbents. That’s not a conspiracy theory; it’s how market dynamics work. Chinese firms are behind on some metrics — raw compute access, frontier model scale — and ahead on others: inference efficiency, cost-per-query, open-source deployment. A freeze locks in that asymmetric position.

This approach can also fail on its own terms. Voluntary pacing only works if both parties comply. Unilateral restraint by U.S. labs, absent Chinese participation, doesn’t reduce global AI risk — it just redistributes who’s doing the frontier development.

Beijing’s Internal Contradiction

China’s public position has a crack in it worth examining. The Foreign Ministry dismissed Western safety frameworks as fearmongering. Simultaneously, State Security Minister Chen Yixin acknowledged real AI security risks and called for accelerating China’s own risk-control systems, according to NBC News.

That’s not a contradiction born of confusion. It’s a deliberate framing distinction: China accepts that AI risks are real, but rejects the idea that a U.S.-led or Silicon Valley-led framework should govern the response. Xi is positioning China as a global open-source AI leader, offering development assistance to BRICS nations — a direct play for AI governance influence in the Global South.

The practical implication: multilateral AI governance is effectively dead as a near-term prospect. What’s actually on the table is parallel domestic regulation — each bloc developing its own standards, with limited overlap and significant friction at the edges.

The Distillation Accusation and Its Consequences

U.S. agencies accusing Chinese firms of model distillation is significant beyond the legal dimension. Distillation — training on outputs of a more capable model — is a standard ML technique. The fact that U.S. agencies are now treating it as IP theft signals an attempt to regulate training data provenance at the geopolitical level.

If that framing gains legal traction, it could justify further export controls not just on chips, but on model weights, APIs, and benchmark access. That’s a meaningful escalation with direct consequences for developers building on Chinese AI infrastructure or using Chinese-origin models.

Two Governance Models With Zero Overlap

The table shows two governance architectures with zero mechanical overlap, even when both sides nominally agree that AI risks exist. That’s why political scientist Ja Ian Chong’s assessment — that the situation requires “simultaneous neutral self-restraint” from both sides, which he considers unlikely — is the most accurate read available right now.

Three Scenarios Worth Tracking

For developers and engineering teams building on AI infrastructure: The distillation accusation carries real near-term risk. If U.S. regulators move to classify certain model training practices as IP violations when Chinese-connected entities are involved, API access restrictions and compliance requirements could land fast. Teams currently using models from DeepSeek, Alibaba’s Qwen family, or similar Chinese-origin systems should audit their exposure now — not after a policy announcement forces the issue.

For companies choosing between U.S. and Chinese AI ecosystems: The bifurcation is accelerating. According to reporting from Madhyamam Online, China’s response explicitly calls for “open, inclusive international cooperation” — but that’s diplomatic language for building an alternative ecosystem. Expect separate benchmarks, separate safety certifications, and separate deployment standards within 18 months. Picking an ecosystem now carries longer-term lock-in than it did a year ago.

Watch September 24: The Trump-Xi summit is the single highest-signal event in the near term. If AI governance appears on the joint communiqué in any form — even vague language about cooperation — that would be a meaningful de-escalation signal. If it doesn’t appear, or if the summit produces direct accusations around distillation, the bifurcation timeline compresses significantly.

What Comes Next

This was never a technical disagreement. It’s a structural one. The framing of “AI safety” lands differently when the person proposing the safety rules is also the person who benefits most from freezing the current standings.

A few things are now close to certain:

• China’s pacing critique is economically coherent: slowdowns preserve incumbent advantages, and Beijing isn’t wrong about that
• Beijing’s internal contradiction — accepting AI risk while rejecting Western governance — signals a parallel-track future, not a cooperative one
• The distillation accusation may be the precursor to a new wave of AI-specific export controls extending well beyond chips
• The September 24 Trump-Xi summit remains the next major signal to watch

The next 6 to 12 months likely bring diverging safety certification regimes, accelerated Chinese investment in domestic compute alternatives — Huawei’s Ascend chips being the clearest example — and increased pressure on Global South countries to pick a side.

The realistic question isn’t whether global AI governance can be unified. That window closed this month. The question is whether the two blocs can avoid active interference with each other’s development, or whether chip restrictions, distillation accusations, and competing pacing proposals escalate into something structurally harder to reverse.

The September 24 summit is where that answer starts taking shape. What happens there — symbolic statement or substantive framework — changes the calculus on everything else.

References

1. China rejects AI slowdown calls, blasts U.S. tech leaders’ ‘fearmongering’
2. China criticises idea it is in ‘malicious competition’ over AI
3. China rejects AI slowdown calls, blasts U.S. tech leaders’ ‘fearmongering’ - eriinfo

Photo by Igor Omilaev on Unsplash