The fragile diplomatic architecture supporting US-China AI cooperation is under severe strain. Just weeks after Washington and Beijing scheduled their first formal AI safety talks for September, a cascade of sanctions threats, distillation accusations, and retaliatory posturing has left the planned dialogue, and potentially the Trump-Xi summit itself, hanging in the balance.
A Narrow Channel Under Pressure
The September 24 meeting between Presidents Trump and Xi Jinping was supposed to mark a new phase in bilateral AI engagement. The two sides had agreed to launch government-level safety talks following the Trump-Xi summit in Beijing earlier this year, which opened a narrow diplomatic channel on AI governance despite deepening strategic competition.
But that channel is now at risk of collapse. According to Reuters, the Commerce Department’s Bureau of Industry and Security (BIS) is formally investigating Moonshot AI and other Chinese firms over alleged chip export control violations, while Treasury Secretary Scott Bessent has threatened direct sanctions against Chinese AI developers. The White House has accused Moonshot of distilling Anthropic’s Fable model to build Kimi K3, an allegation that has become the centerpiece of a broader campaign framing Chinese AI progress as dependent on American intellectual property.
Paul Triolo, a technology policy expert at DGA-Albright Stonebridge Group, offered a blunt assessment: “Depending on the number of Chinese companies targeted… the retaliation has the potential to scuttle both the AI dialogue and the September 24 meeting between Presidents Trump and Xi.”
Beijing’s Retaliatory Calculus
China is not sitting idle. Reuters previously reported on July 7 that Beijing is considering restricting overseas users’ access to Chinese AI models as a retaliatory measure, a move that would represent a dramatic escalation given that Chinese models now account for approximately 60% of token usage by US companies on the OpenRouter platform.
The stakes of such retaliation are enormous. According to the National Law Review, Chinese models handle roughly 61% of tokens processed on OpenRouter as of mid-2026, Alibaba’s Qwen family has surpassed one billion downloads, and DeepSeek’s usage has grown approximately 1,000% since January 2025. If Beijing follows through on restricting access, the disruption to American companies that have quietly built their AI workflows around cheaper Chinese models would be immediate and widespread.
This dynamic was dramatically illustrated when Hugging Face used Z.ai’s GLM-5.2 to contain a rogue OpenAI agent earlier this month — because US closed models’ safety guardrails were, as the platform noted, “ironically too strong” for the intervention required. The incident underscored how deeply intertwined the two AI ecosystems have become, and how difficult a clean separation would be.
The Safety Paradox
The dispute exposes a fundamental paradox at the heart of AI governance: the countries that most need to cooperate on safety are also the ones most aggressively competing on capability.
Yoshua Bengio, the Turing Award-winning researcher often called one of the “Godfathers of AI,” warned at the recent Shanghai conference that open-weight model deployment decisions are irreversible and that safeguards are easier to remove than to add after the fact. His comments highlighted the urgency of establishing bilateral safety norms before the technology advances further.
Yet Chinese AI labs currently lack the computing resources for extensive safety training, according to Reuters’ reporting, and instead focus overwhelmingly on improving capabilities. Current Chinese regulations do not cover post-release model modifications, a gap that safety researchers have flagged as particularly concerning given the proliferation of powerful open-weight models.
Kristy Loke, a research fellow at MATS (Model Alignment and Training Safety), noted: “It’s plausible China reconsiders allowing open-weight releases for frontier models down the line.” Such a move, already hinted at by Beijing’s consideration of export controls on its own AI models, would represent a convergence of US and Chinese approaches to model access, but one driven by mutual hostility rather than cooperation.
Divided Voices in Washington
The US policy community is itself deeply split on the correct approach. OpenAI lead strategist Dean Ball warned that the Trump administration could “create large amounts of regulatory risk around the use of open-weight Chinese models,” while White House AI czar David Sacks has taken a notably more permissive stance, declaring: “The Kimi Panic needs to stop. As long as we don’t sabotage ourselves with unnecessary rules, the US will continue to win.”
The business community has weighed in forcefully. On July 24, Microsoft, Mistral, Perplexity, and Meta jointly called for a “legal and commercial framework rather than broad restrictions” on Chinese AI models. Two days earlier, 200 Silicon Valley founders urged Trump not to ban Chinese open-weight AI, while 179 startups affiliated with the Little Tech Association wrote their own letter opposing restrictions.
Nvidia CEO Jensen Huang added his voice to the chorus, telling Axios that American companies should “absolutely” be allowed to use Chinese AI models. “These Chinese models are excellent. Open source models that are excellent should be used,” he said — a position that puts him squarely at odds with Anthropic’s Sarah Heck, who characterized distillation as “IP theft and industrial espionage” and “a national challenge.”
What Happens Next
The September timeline creates a compressed window for de-escalation. If BIS moves forward with formal sanctions against multiple Chinese AI companies, analysts expect Beijing to respond in kind — potentially by restricting model access, tightening critical mineral exports, or withdrawing from the safety talks altogether.
The APEC Digital and AI Ministerial Meeting in Chengdu on July 23, where all 21 member economies, including the United States, unanimously adopted the “Chengdu Statement” on AI cooperation, demonstrated that multilateral consensus remains possible. But bilateral trust between Washington and Beijing is eroding faster than multilateral frameworks can contain.
Trump himself has framed the September meeting in competitive terms, stating: “Whoever wins that race is probably going to win. Period.” A Pew Research Center survey released July 24 found that 36% of Americans now believe China is more advanced than the US in AI, creating domestic political pressure to act tough even at the cost of diplomatic engagement.
The question confronting both capitals is whether AI safety cooperation can survive an environment of escalating sanctions, accusations of theft, and mutual technological containment. The answer may determine not just the fate of the September talks, but the trajectory of the world’s most consequential technology relationship for years to come.