When ChatGPT stunned the world in late 2022, US artificial intelligence labs enjoyed what looked like an unassailable lead, measured not in months, but in years. By early 2025, DeepSeek’s R1 had compressed that advantage to a matter of months. In July 2026, with Moonshot’s Kimi K3 and Alibaba’s Qwen3.8 landing within a fortnight of the latest US frontier releases, the gap has effectively collapsed to weeks. According to a detailed analysis published this week by the South China Morning Post, that compression is forcing an uncomfortable strategic reckoning in Washington, and there is no consensus on what to do next.
From Years to Weeks
The collapse of the American lead is not a matter of interpretation. It is being measured in benchmark scores, developer downloads, and enterprise procurement decisions. Kimi K3, whose open weights are scheduled for release on July 27, is already forcing Moonshot to pause new user subscriptions because global demand has outstripped the company’s compute capacity. Alibaba’s 2.4-trillion-parameter Qwen3.8, previewed earlier this month, trails only OpenAI’s Fable 5 on internal comparisons, according to the company. And as Nature reported earlier this week, a growing chorus of scientists now argues the frontier gap has closed outright.
“I don’t think anybody truly knows what’s going on,” Ryan Fedasiuk of the American Enterprise Institute told the SCMP, a candid admission that the analytical frameworks built during the era of clear American dominance no longer map cleanly onto reality.
Three Policy Questions With No Clean Answers
The SCMP piece, authored by Washington-based correspondent Vincent Chow, distills the emerging debate into three interlocking questions.
First: how tightly should Washington police its own frontier? The Trump administration last month forced Anthropic to suspend its Fable 5 and Mythos 5 models over national security concerns tied to their dual-use capabilities. But that instinct sits uneasily beside the fact that Chinese labs are shipping comparably powerful systems into the open. David Sacks, the former White House AI tsar, has been particularly vocal, arguing that “Kimi K3 just fixed 15 critical security bugs that Codex and Fable refused because of ‘cyber guardrails.’
There’s no reason to limit American models on tasks that Chinese models handle without issue.” The argument cuts to the core of America’s self-imposed handicap: if adversaries face no equivalent restraint, guardrails may be exporting capability rather than containing it.
Second: should the United States restrict domestic adoption of Chinese open-weight models? This question has moved from theoretical to urgent. Andreessen Horowitz estimates that roughly 80% of US AI startups now incorporate Chinese base models into some part of their stack. Qwen alone surpassed one billion cumulative downloads on Hugging Face by March. Axios has reported that Kimi K3’s release has emboldened parts of the Trump administration to push for broad bans on Chinese open-weight models, while “pro-competition” officials, worried about crippling the domestic startup ecosystem, have pushed back.
Ben Hayum of the Center for a New American Security captured the establishment’s surprise: analysts had assumed Beijing would abandon open-source once its models crossed critical national security thresholds. It hasn’t. “Some US restrictions on adoption of Chinese models could make sense given the risk of ideological bias and sleeper agents,” Hayum said.
Treasury Secretary Bessent has tried to thread this needle, insisting the administration “supports open source models” while threatening sanctions on adversarial distillation, a pointed reference to the White House’s accusation that Moonshot distilled Anthropic’s Fable to build K3. Bessent has also hinted that US firms may eventually be required to disclose their use of Chinese models, a disclosure regime that would function as a soft ban by raising compliance and reputational costs.
Third: how does Washington drive global adoption of American AI? This is the question that increasingly animates a rising school of thought inside the policy community. Fedasiuk argues the US should abandon a rear-guard focus on model capability parity and pivot to a “forward-looking” strategy of accelerating AI infrastructure buildout globally, particularly across allied Gulf states such as the UAE and Saudi Arabia. In his framing, the contest is now “all-out infrastructure competition”, compute, data centers, energy, deployment — rather than a benchmark race the US can no longer expect to dominate cleanly.
That framing gained urgency after Xi Jinping’s landmark address at the World Artificial Intelligence Conference in Shanghai, where 28 countries, largely from the Global South, signed onto China’s “inclusive” AI-for-all vision. Beijing is not merely competing on model quality; it is offering a diplomatic package around open weights, cheap inference, and sovereignty-friendly deployment terms that resonates well beyond the developed West.
The Open-Weight Paradox
Perhaps the sharpest irony in Washington’s predicament is that China’s continued willingness to release frontier models openly — Kimi K3 on July 27, Qwen3.8 shortly after, has flipped the traditional narrative about which side benefits from restrictions. For a decade, US officials assumed American openness accelerated global research while Chinese controls held Beijing back. In 2026, the polarity has partially reversed. American labs are being pulled behind closed doors by national security review, while Chinese labs continue to ship weights that global developers can inspect, fine-tune, and deploy.
Making matters more complex, Beijing itself is now weighing export controls on its own AI model weights — a mirror-image policy that could close the open-weight window from the other direction. Should the Ministry of Commerce move forward with the controls under consultation with Alibaba, ByteDance, and Z.ai, the current moment of Chinese openness may prove a temporary window rather than a durable posture.
The Ideological Pivot
Wang Yaqiu of the University of Pennsylvania’s Project on the Future of US-China Relations predicts Washington will increasingly shift its rhetoric to highlight ideological differences between American and Chinese models as the capability gap becomes indefensible as a talking point. There are early signs.
The Center for AI Standards and Innovation, housed under NIST, was originally tasked with evaluating Chinese models for “alignment with Communist Party talking points and censorship,” but conspicuously stopped publishing such findings in its most recent reports. CAISI’s director, Chris Fall, left the role after just three months, a departure that observers read as symptomatic of the disarray in Washington’s AI benchmarking apparatus.
President Trump has confirmed AI will feature prominently on the agenda when Xi Jinping visits on September 24. “We’re leading China. We want to keep it that way,” he said this week. “Whoever wins that race is probably going to win. Period.” Yet a Pew Research Center survey released Thursday found that 36% of Americans now believe China is more advanced in AI than the United States, versus only 12% who say the US leads. The domestic political predicate for Trump’s confidence is eroding as quickly as the technical one.
Washington has weeks, not years, to decide whether restriction, adoption, or infrastructure competition will define the next phase.