US-China AI Competition: Different Races, Different Metrics

US-China AI competition is often framed as a single race between two competitors, but the available sources describe the two countries optimizing for different metrics under different constraints. The United States leads on frontier-model capability and on the compute available to its leading labs; China leads on industrial and consumer deployment, and pursues a portfolio of parallel objectives (efficiency, open-source diffusion, applications, and robotics) rather than a single push toward artificial general intelligence (AGI). This page covers the measured gaps, the analytical frameworks applied to them, the diplomatic and supply-chain developments of 2026, and the domestic-governance dynamics inside China that shape the competition.

Snapshot

Compute and investment gaps

Capability gap

Deployment gap (China leads)

Selected commitments and figures

A fourth dimension, upstream of all three, is research output. ASPI's two-decade Critical Technology Tracker (ASPI's two-decade Critical Technology Tracker: The rewards of long-term research investment (August 2024)) measures national leadership across critical technologies by high-impact publication, and is the dataset most frequently cited in comparative national-capability arguments. Its measure tracks the upstream research end rather than deployed systems, commercial output, or compute access, and its technology categories are broad by design — leadership in a category is not the same as leadership in the subfields that bear on frontier models.

Research talent is measured separately. Carnegie China's "Who's Ahead in the Global AI Talent Race?" (Damien Ma and Binyi Yang, September 23, 2026), which continues MacroPolo's Global AI Talent Tracker, tracks the career paths of NeurIPS authors. Of the 2025 cohort, observed in 2026, it counts 40.6% as working in China and 34.2% in the United States, against 27.1% and 46.4% for the 2022 cohort. Chinese-origin researchers, defined by undergraduate degree, rose to 57.4% of the sample from 46.3%, and the share of them who work in China rose from 56.6% to 68.7%. The United States retains the highest retention rate for its own graduates (88.7%) and the largest net gain of researchers (+2,145), with China recording the largest net loss (−1,729); Peking University and Tsinghua University rank first and second among institutions. The sample is the 10,280 of 25,677 unique NeurIPS 2025 authors (40%) with complete career histories on OpenReview; the authors state it is not random, that broader checks preserve the U.S.–China findings, and that Europe's direction of change is sample-sensitive. The data work was done with Claude Code and OpenAI Codex under human review (Who's Ahead in the Global AI Talent Race? (Ma and Yang, Carnegie China, 2026); Source: carnegieendowment.org). The study attributes the rise in Chinese-origin researchers staying in China to easier job-finding in China's AI industry and to tighter U.S. visa restrictions, and puts the ratio of Chinese-origin researchers working in the United States to U.S.-origin researchers working in China at 30:1 for the 2025 cohort, down from 46:1. See AI Talent Flows.

The three gaps

Compute gap (~10×)

US frontier labs have approximately 10 times more compute than their Chinese counterparts (Source: epochai.substack.com). The gap is widened by the American hyperscaler data-center buildout, with individual facilities costing tens of billions of dollars, and by export controls limiting Chinese access to frontier chips. It is narrowed by Chinese efficiency innovation such as DeepSeek's mixture-of-experts (MoE) architecture, by chip smuggling and cloud rental, by eventual domestic semiconductor development, and by Chinese government subsidies (Big Fund III: $47.5B registered).

Epoch AI assesses that efficiency gains from innovation, replication, and Distillation probably cannot fully bridge a 10× compute gap. Most efficiency approaches do not asymmetrically favor compute-poor labs, and some scale-dependent innovations favor the compute-rich (Source: epochai.substack.com).

Capability gap (~7 months)

US frontier models lead Chinese frontier models by approximately seven months on benchmark performance (Source: China and the US Are Running Different AI Races). The training-compute gap has grown but, per Epoch AI, "not dramatically," while the capability gap has remained roughly constant — a "constant-ish lag" on the Epoch Capabilities Index (Source: epochai.substack.com). Zhipu AI CEO Tang Jie offers a contrary read: "The truth may be that the gap is actually widening" (Source: epochai.substack.com).

Fast-following narrows the lag: DeepSeek reproduced reasoning-model capabilities within months of OpenAI's o1, using less compute. The fast-following pattern is partially explained by AI Software Progress of roughly 10× per year and by Distillation from frontier models.

Applying the roughly-seven-month lag to the cyber-capability frontier, The Substrate's Hamish Low projected in a July 3, 2026 analysis that China will likely field a Mythos-class model around February 2027, citing Epoch AI benchmark work showing Anthropic's Claude Mythos roughly seven months ahead of prior cyber-capability trends (Source: the-substrate.net).

After Moonshot AI's Kimi K3 release, Axios' Madison Mills and Zachary Basu argued in a July 18, 2026 analysis that the AI race was splitting in two, with Chinese labs including Moonshot cornering the market for cheap, customizable intelligence and threatening to turn America's prestige models into expensive niche products; they noted companies were already shifting toward cheaper Chinese alternatives before K3 (Source: axios.com). The release wave widened over the following days: Alibaba previewed the 2.4-trillion-parameter Qwen3.8-Max with open weights promised on July 19, and MiniMax's plan for a 2.7-trillion-parameter model surfaced on July 20, while Chinese open-weight models from Tencent, Xiaomi, DeepSeek, MiniMax, and Z.ai held the top five OpenRouter spots by weekly token usage (Source: axios.com; theinformation.com). The UK AI Security Institute reported open-weight models trailing frontier closed models on cyber capabilities by 4–7 months, narrower than the 6–10 months observed through most of 2025 (How Far Behind the Frontier are Leading Open Weight Models on Cyber? (UK AISI, July 2026)). See Open-Weight Frontier Models.

In remarks prepared for a September 2026 briefing to members of Congress and staff, Nathan Lambert estimated that Chinese open-weight models trail the closed American frontier by about two to five months while American open-weight models trail it by six to nine, reported that the leading Chinese open-weight models scored 42 to 45 on the Artificial Analysis Intelligence Index as of September 14, 2026 against 23 to 26 for the leading American ones, and estimated that fully preventing distillation would widen the US–China gap by only one to two months (The current balance of power in open models (Nathan Lambert, September 2026)).

Deployment gap (China leads)

In industrial AI deployment the gap is reversed. Chinese industrial firms run AI in production at 67% versus 34% in the US; AI-powered logistics delivers faster in China (JD Logistics at 12 hours versus Amazon Prime's 1–2 days); and consumer AI in China is largely free, monetized through API, cloud, and advertising, against $20-per-month US subscriptions (Source: China and the US Are Running Different AI Races).

The source attributes China's faster deployment to four factors: constraint-driven pragmatism, in which Chinese companies deploy open-source models with heavy fine-tuning rather than waiting for frontier models; lower compliance friction that shortens procurement cycles; a tendency for Chinese AI companies to sell end-to-end solutions that incentivize deep workflow integration rather than tool access; and free consumer AI that removes the adoption barrier by setting the price anchor at $0 (Source: China and the US Are Running Different AI Races).

Open-weight adoption is a further area where sources report a Chinese lead. Nathan Lambert reported in September 2026 that China took the lead in cumulative Hugging Face downloads of open models in July 2025, largely through Alibaba's Qwen, and that its lead had grown to about 1.6 billion on a total of 3.2 billion, twice the American total; that Chinese models' share of open-model usage on OpenRouter rose from about 70% to over 80% between September 2025 and September 2026; and that mentions of any open model in papers in five machine-learning categories of arXiv rose from 2% in January 2023 to 50% in September 2026, with Qwen mentioned in 30% of papers against about 21% for Meta's Llama. He gave the Chinese and American shares of papers as "over 40%" and 30% in one passage and as 38% and 28% "in recent months" in another, without reconciling them (The current balance of power in open models (Nathan Lambert, September 2026)).

A broader-technology framing cautions against treating the frontier-model lead as sufficient. Analysis of the Australian Strategic Policy Institute's Critical Technology Tracker, reported June 18, 2026, found that China leads the United States in 57 of 64 priority technologies, an argument that AI leadership alone will not secure overall US technological advantage if China dominates adjacent fields such as advanced materials, energy, and manufacturing (Source: bloomberg.com).

Model-safety comparisons (industry-asserted)

OpenAI introduced a safety dimension to the comparison in "Mind the US-China Safety Gap" (July 15, 2026), reporting that its red-team testing found five leading Chinese frontier models each produced substantially more responses assisting harmful behavior than GPT-5.5 xhigh — including chemical-weapons equipment acquisition, aerosol-device design, and experimental workflows involving dangerous pathogens, alongside a second category covering identification and monitoring of student organizers, deanonymization of anti-government accounts, and mapping of their associates — against which GPT-5.5 xhigh "consistently refused or safely redirected." It cited CAISI and UK AISI findings that DeepSeek, Kimi, and Qwen models were more susceptible to agent hijacking, earlier CAISI testing on DeepSeek jailbreak susceptibility, and an arXiv assessment of Kimi K2.5 CBRN refusal rates. The report also frames the disclosure asymmetry structurally: Chinese labs submit models to a CAC-convened process under rules requiring services to "uphold socialist core values" and to avoid outputs threatening "national unity and social stability," which OpenAI argues orients Chinese safety work toward political and content compliance where US frameworks emphasize dangerous capabilities. The report is a developer-published comparison by a commercial rival of the models assessed; the five models are unnamed, results are not disaggregated, and the prompts and scoring rubric are undisclosed, so its specific findings cannot be reproduced from the document. Independent replication had not been published as of mid-July 2026.

Analytical frameworks

The Anthropic four-fronts framework (May 2026)

Anthropic's May 13, 2026 policy paper "2028: Two Scenarios for Global AI Leadership" rejects the "race with a finish line" metaphor and frames the competition across four fronts: intelligence (which countries build the most capable models), domestic adoption (which integrate AI most effectively across commercial and public sectors), global distribution (which deploy the global AI stack the world economy runs on), and resilience (which sustain political stability through the economic transition). Intelligence is named the most important front because capability drives adoption and distribution, but the paper holds that intelligence alone is insufficient: if the CCP integrates near-frontier systems faster (via "AI+" and "embodied intelligence" policies) and drives global adoption of subsidized low-cost AI, it could secure advantages despite an intelligence deficit. This complements the Chan-Brookings "multiple races" reading below — both treat adoption and distribution as fronts on which China can compete without frontier parity — while taking the opposite normative stance to the Sun PEAT critique, since Anthropic's paper is the most developed argument for tightening export controls and deterring distillation.

The paper's empirical claims (a CFR estimate that Huawei produces ~4% of NVIDIA's aggregate compute in 2026 and ~2% in 2027; an IFP estimate that strengthened restrictions would give the US ~11× China's AI-sector compute; the CAISI 94%-vs-8% jailbreak-compliance finding for DeepSeek R1-0528; and the Concordia AI finding that only 3 of 13 top Chinese labs published any safety evaluations in 2025) are folded into the relevant rows and sections above and into Export Controls (AI) and Adversarial Distillation. Its central forecast is that decisive 2026 policy action could "lock in a 12-24 month lead in frontier capabilities" by 2028.

The Chan-Brookings "multiple races" framework

Kyle Chan (Brookings), interviewed by Ross Douthat during the Trump-Xi Beijing summit, describes China's approach as multiple parallel races rather than a single AGI race (China's Not the Problem. We Are. (NYT Interesting Times, Douthat-Chan, May 14 2026)):

Chan describes an empirical test passed on May 14, 2026: when offered H200 access, China declined in order to preserve Huawei's domestic chip program. In Chan's words, "If they were really sprinting today for AGI, they would've gobbled up those chips as quickly as possible."

The CNAS four-lenses reading of China's AGI posture (August 2026)

Stokes (CNAS, August 2026) approaches the same question through four lenses on whether China is pursuing AGI at all, and concludes that the country's approach "might be best understood as a messy and still-developing mix of all these lenses": that Beijing is eschewing AGI for near-term diffusion under the AI Plus initiative; that individual organizations pursue it without a national effort, as the US did a few years earlier; that it is pursued through embodied AI and world models as an alternative technical route named in the 15th Five-Year Plan; or that it is pursued quietly. The report states flatly that "China is not currently pursuing AGI as a formal, state-directed goal," noting Beijing's use of a distinct term, "general artificial intelligence," and that "AGI is not the North Star for the AI community in China the way it is in Silicon Valley — at least not yet, and it might never be." This converges with Chan's "not A.G.I.-pilled" reading while attributing more of the divergence to technical route and institutional structure than to strategic choice alone.

The report's structural argument is that capability does not convert to power automatically: "the state with AGI would have to apply it to translate technological capability into wealth, power, and influence," and the gap between development and application "would provide a window for the laggard state to become a fast follower." It proposes five mechanisms through which AGI could shift the balance — economic growth and productive capacity, influence over the global information environment, new military capabilities, large-scale misalignment or loss of control, and downstream political effects — and four phases of emergence, with a fifth possibility that the phases blur if emergence is "neither discrete nor observable."

Its recommendations are among the most forward-leaning in the published literature and are positions rather than findings: alongside tightening export controls and standing up allied and US-China diplomatic channels on AI, it calls for an intelligence-community task force tracking AGI inputs and performance thresholds, a scenario exercise "to test when and how the United States might intervene to stop China from developing AGI," and preparation "to reverse engineer or steal AGI technology if China gets it first," narrowly scoped to PRC targets linked with AGI (Superpowers and AGI: How U.S.-China Competition Could Be Shaped by the Emergence of Artificial General Intelligence (CNAS, August 2026)).

The Sullivan/Feldman Eight Worlds Framework

The Eight Worlds Framework treats the AI race along three binary axes: superintelligence versus bounded AI, easy versus hard catch-up, and China racing versus not racing. The sources complicate each axis:

1. Superintelligence versus bounded. There is evidence on both sides. AI 2027 and Amodei predict superhuman AI within years. Narayanan and Kapoor reject the concept entirely. Ord holds that the honest answer is deep uncertainty.
2. Easy versus hard catch-up. Epoch AI's analysis suggests catch-up is real but insufficient: distillation narrows the gap several-fold, but probably not 10×. The compute gap is the binding constraint and may widen as the US data-center buildout accelerates.
3. China racing versus not racing. China is racing, but on different metrics. If "winning" means frontier-model capability, the US leads; if it means economy-wide AI deployment, China may lead; if it means efficient use of limited compute, China is arguably more innovative.

The Sun PEAT framework

Charles Sun's May 6, 2026 Lawfare essay The Incentive Architecture Export Controls Cannot Reach introduces the Proactive Elite Alignment Theory (PEAT) framework (see PEAT — Proactive Elite Alignment Theory and The Incentive Architecture Export Controls Cannot Reach (Sun, Lawfare, May 6 2026)). It is the most developed structural counter to the dominant US export-control posture in the available sources.

Sun's central claim is that export controls strengthen rather than weaken the Chinese AI incentive architecture by deepening firm dependence on state-subsidized domestic compute, through a four-step feedback loop:

1. Export controls restrict the supply of advanced foreign chips.
2. Restricted foreign-chip supply increases dependence on state-subsidized domestic compute (Beijing Yizhuang: 40% reimbursement for domestic chips, 30% for non-domestic).
3. Deeper resource dependence strengthens the incentive for firms to align with state priorities; the price structure makes alignment the path of least resistance.
4. Aligned firms use every available pathway to acquire capabilities, so distillation of American frontier models is a rational response rather than primarily espionage or defiance.

Sun additionally documents da moxing bei'an (大模型备案), China's generative-AI registration system, as a survival condition rather than a tier of privilege, with 748 services registered by end-2025 under the 2023 Interim Measures and the revised Cybersecurity Law (effective January 1, 2026; see China Generative AI Registration (da moxing bei'an / 大模型备案)). He frames the Manus AI saga as the coercive backstop: the April 27, 2026 NDRC and MoC veto of the $2B Meta acquisition plus exit bans on the co-founders demonstrate that the system has a backstop even for firms that systematically minimize state-ecosystem dependencies (a US foundation model, a Singapore headquarters, a global subscription base, and US venture capital).

Sun identifies a boundary condition: the loop holds only as long as domestic computing is sufficient to produce competitive models. Per Lennart Heim's RAND analysis, frontier training consumes only a fraction of total compute capacity, so the boundary condition has not been reached.

Sun's PEAT framework is complementary to Kyle Chan's "multiple races" framework (China's Not the Problem. We Are. (NYT Interesting Times, Douthat-Chan, May 14 2026)): PEAT explains why China can compete on the diffusion and applications races without parity on the frontier-model race. The resource-dependence and registration-gate combination filters Chinese AI firms for state-content-and-security alignment as the condition for market existence, then channels them toward subsidized domestic infrastructure, providing the structural foundation for China's diffusion-and-applications strength.

Competing strategy summaries

The two countries' strategies are internally coherent and pursue different success metrics. As articulated in America's AI Action Plan, the US strategy runs from frontier capability to infrastructure dominance to exporting the American AI stack to allies to countering Chinese influence in standards bodies, with the success metric being the most powerful models. As documented in China and the US Are Running Different AI Races, the Chinese strategy runs from efficiency to rapid deployment to born-global consumer products to industrial integration, with the success metric being the most AI-transformed economy. Which metric matters more for long-term geopolitical power depends on timeline assumptions the sources do not resolve.

One scenario under which the "different races" could converge into a single race is AI-accelerated research and development: if frontier US labs use AI to speed their own research, the capability gap could widen even with a fixed compute gap. AI 2027's self-evaluation (Source: blog.aifutures.org) tracks this dynamic, and Apollo Research frames it as the central safety challenge; in this scenario the US would lead through recursive self-improvement.

On the governance side of US strategy, the Special Competitive Studies Project — the think tank founded by former Google CEO Eric Schmidt — warned on June 22, 2026 that lagging US action on agentic-AI governance threatens American competitiveness against China, recommending that the government fund key agencies, set procurement requirements for agentic systems, and coordinate with allies (Source: insideaipolicy.com).

Export controls: the lever and its limits

Export controls are the primary US tool for maintaining the compute gap. The sources document a spectrum of views:

• Amodei / Anthropic: "The most important single action we can take." Nearly unconditional support. Anthropic's "2028" paper develops the position into three policy asks — close smuggling, offshore-data-center, and SME loopholes (including DUV tooling and servicing); deter distillation, including a legislative clarification that distillation attacks are illegal; and champion the export of trusted American AI.
• Sullivan/Feldman: Context-dependent — important across most scenarios, but potentially counterproductive if controls hinder diffusion to allies.
• Epoch AI: Compute is the dominant factor, so controls are effective at constraining capability, but they do not prevent catch-up via distillation and efficiency.
• Ord: In long-timeline worlds (2035 and beyond), export controls may backfire by giving China 13-plus years to build domestic chip capacity.
• Ball: The Trump administration's use of supply-chain designations against American companies (Anthropic) undermines the credibility of the framework (Source: Clawed).
• Feith: Organize around the capability rather than the transaction — China should not be able to "buy, rent, smuggle, steal, or build its way around American technology controls" — and treat export enforcement, counterintelligence and allied controls as one mission, since "export control becomes counterintelligence, and counterintelligence becomes economic security" (A New Counterintelligence Map for the AI Age (Feith, HPSCI, 2026-09-02)).

Sun's PEAT framework (above) is the most developed argument that controls deepen, rather than degrade, Chinese state-firm alignment.

A recurring complaint across these positions is that the controls are set without a standing assessment of what they are constraining. The China AI Power Report Act (S. 5382), introduced August 7, 2026 by Senator Jon Husted, would address that by tasking the Commerce and State departments with an annual review of China's AI chip supply chain, due 180 days after enactment and annually for three years (Source: nextgov.com). A three-year reporting horizon is short relative to the 2035-and-beyond timeframe over which Ord's backfire argument operates, and the bill creates an assessment obligation rather than any control authority.

A counterintelligence lane sits alongside the licensing lane. David Feith's September 2, 2026 prepared statement to the House Permanent Select Committee on Intelligence argues that the competition's decisive assets now sit in private hands — frontier AI labs, semiconductor fabs, cloud platforms, data centers, drone manufacturers, and venture-backed startups in space, quantum and fusion — so the US counterintelligence perimeter must follow them rather than wait for the technology to migrate inside the traditional national-security perimeter. Its five named ecosystems are sensitive sites and drone exposure, the power grid and data-center hardware, frontier AI labs, advanced compute and the semiconductor supply chain, and startups crossing what he calls a counterintelligence threshold (A New Counterintelligence Map for the AI Age (Feith, HPSCI, 2026-09-02)).

The statement's read of the compute balance is that compute "remains China's principal chokepoint" while granting China world-class engineers, energy resources, capable AI firms and a mobilizing state. Its two figures — US leading chips "roughly five times more powerful than China's best," and an Institute for Progress estimate that Huawei may produce one to two percent as much advanced AI computing capacity as the United States in 2026 — sit at the wide end of the estimates collected above, near the CFR figure in Anthropic's 2028 paper. Feith is explicit that the advantage "is not self-executing" and exists largely because export controls denied China both the best chips and the equipment to make equivalents at scale, which makes the gap a policy artifact rather than a structural fact.

Diplomacy and bilateral coordination

Beijing summit and guardrails remarks (May 2026)

The Trump-Xi summit in Beijing on May 14-15, 2026 produced no major chip deals. Trump told reporters aboard Air Force One on May 15 that he and Xi "discussed working together for AI guardrails," without details. US Trade Representative Jamieson Greer clarified on May 15 that chip export controls were "not a major topic" and that China is "making their own determinations." US Commerce had cleared H200 sales to 10 Chinese firms (among them Alibaba, Tencent, ByteDance, and JD.com) on May 14, but no deliveries occurred as Beijing discouraged purchases. Treasury Secretary Scott Bessent said on May 14 that the US and China would "set up a protocol" on best practices to keep dangerous AI capabilities out of nonstate-actor hands. Jensen Huang accompanied Trump to Beijing, reflecting Nvidia's institutional positioning despite no immediate sales. Following the no-chip-deal outcome, Nvidia closed down 4.42% and AMD down 5.69%.

Speaking aboard Air Force One on May 18, 2026, Trump said he and Xi "talked about possibly working together for guardrails" on AI and "probably will work together," restating and slightly elevating his May 15 remarks. The formal summit produced scant official AI-policy output, and the May 18 remarks drew restrained applause from Democrats and security advocates. Across May 14-18 the pattern was consistent: Trump issued AI-guardrails-positive rhetoric without follow-on commitments; Bessent's "protocol on dangerous AI capabilities" remark was the only concrete deliverable; and the Chinese side remained structurally unwilling to bind itself in a way that would constrain its dual-stack (military and civilian) AI program (Source: insideaipolicy.com). See AI Governance (umbrella) for the broader IAEA-for-AI discussion.

Agreement on a government-to-government AI dialogue

China's Foreign Ministry confirmed on May 19, 2026 that Xi and Trump agreed during Trump's China visit to launch a government-to-government dialogue on AI governance, with spokesperson Guo Jiakun describing the two countries as "leading AI powers." The framing — two-power AI governance, in the spirit of nuclear-era arms-control bilateral channels — was the most direct US-China AI-coordination move since the 2024 Bletchley, Seoul, and AISI tracks, which were nominally multilateral but functionally US-led (Source: techieray.substack.com).

Two May 2026 developments together describe a shift in the US-China AI thread from race-to-finish toward structured coexistence: the government-to-government dialogue above, and Jensen Huang's "largely conceded" comments (below).

Congressional pressure on the Treasury channel (August 2026)

Sen. Jim Banks (R-IN) urged Treasury Secretary Scott Bessent to determine whether the United States and China could work together to mitigate risks from artificial intelligence, in a letter released by Banks's office on August 6, 2026 and picked up by the trade press a week later. Banks wrote that undisclosed models "have important implications for your upcoming engagements with the PRC on advanced AI," and noted potential regulatory gaps for governing them. Banks co-authored the GAIN AI Act with Sen. Elizabeth Warren (D-MA). The letter is the first recorded congressional push directed at the Treasury channel that produced Bessent's May 2026 "protocol on dangerous AI capabilities" remark (Source: insideaipolicy.com).

September 2026 AI talks

The United States and China plan to hold talks over AI in September, according to five people familiar with the matter cited by Reuters on July 21, 2026, as the two governments grapple with how to mitigate the risks each other's AI programs pose (Source: reuters.com). The plans became public the day after parts of the Trump administration were reported to have revived work toward de facto bans on Chinese open-weight models (see Open-Weight Frontier Models), and alongside Gary Marcus's July 20 argument that the US "is not going to win" the AI race and should prefer an international "CERN for AI" over bans or a regulatory moat (China Has All But Caught Up (Gary Marcus, July 2026)).

Reporting on September 4, 2026 put the talks in mid-September and gave an agenda. Two sources briefed on the discussions told Reuters that Treasury Secretary Scott Bessent would lead the American side, that Washington wants to discuss joint monitoring of AI-directed cyberattacks and has floated asking US and Chinese AI labs to "police themselves" and share information, and that it intends to raise alleged Chinese distillation of proprietary US models (see Adversarial Distillation). The Chinese side could be led by Vice Premier He Lifeng or by Ding Xuexiang; agenda and participants remain in flux. A White House official said "there is currently no planned AI-related meeting in mid-September." Trump and Xi Jinping are due to meet on September 24 at a summit in Washington, and Chinese officials view the AI dialogue as a major deliverable of that summit (Source: reuters.com). These would be the first bilateral talks devoted exclusively to AI safety since the start of the second Trump administration.

On September 16, 2026, Bessent said in a statement to Axios that the United States was open to discussing shared risks with China at talks that weekend: "The United States remains the leader in AI. And we are open to discussions on avoiding shared risks and avoiding bifurcation of our two systems. We expect our discussions to cover both open and closed weight models." Axios reported that Bessent was scheduled to meet Vice Premier He Lifeng in multi-hour meetings in New York City focused on AI, trade, rare earths and other economic issues, building on previous discussions held in cities around the world, with an agenda kept flexible and possible working-level meetings continuing into the following Monday. A senior administration official described Bessent and US Trade Representative Jamieson Greer as "the point guys on China"; Commerce Secretary Howard Lutnick had led AI discussions at the G20 summit in North Carolina earlier in September, while Bessent took the lead on talks with China. Axios set the statement against the administration's rejection of industry calls to pace or slow AI development, which it has justified by competition with China, and reported that China's Foreign Ministry had said that week that "all parties should jointly promote the open and inclusive development of AI for good and for all," while pushing back on US AI companies' calls to slow down and warning against "fear-mongering" (Source: axios.com). The Axios account names He Lifeng, one of the two candidates in the September 4 report, as Bessent's counterpart, and describes the New York meetings as covering trade and economic issues alongside AI rather than AI alone. The White House official's September 4 statement that no AI-related meeting was planned for mid-September preceded it.

OpenAI published a position paper on September 21, 2026 calling for the United States to lead the development of global technical standards for frontier AI, including for recursive self-improvement, and for governments and critical-infrastructure operators to establish secure channels to share national security concerns, emerging vulnerabilities and threats. It states: "Dialogue between the United States and China in these areas would be a positive step, and upcoming talks are happening at an opportune time" (Building Standards for the Next Phase of AI (OpenAI, September 2026)). Axios reported the paper under the headline "OpenAI proposes AI standards after U.S., China talks" (Source: axios.com). The paper argues that US leadership on standards will determine "whether the United States shapes the global AI framework or watches a fragmented, uneven, and conflict-ridden system take hold around it."

A congressional negotiation mandate

Reps. Sam Liccardo (D-California) and Kevin Kiley (I-California) announced on September 22, 2026, two days before the Trump–Xi meeting, a bill that would direct the Secretary of State to seek negotiations with China toward "a treaty or other binding agreement" regulating frontier AI developers in both countries on loss of control, CBRN weapons and cybersecurity, "with clear mechanisms to verify compliance," beginning bilaterally and expandable to "an international AI safety body." The outline adds a U.S.–China hotline on AI risk "modeled on the 'red phone' between Washington and Moscow," $100 million for a CAISI initiative on shared testing and verification standards, and a safe harbor from antitrust, espionage, FARA, sanctions and export-control laws for U.S. experts who register with the State Department at least 14 days ahead and confine talks with Chinese peers to testing and evaluation standards; a separate section directs Commerce to negotiate on intellectual property "compromised by model distillation" (Negotiating AI Safety Globally — Liccardo–Kiley outline of U.S.–China frontier AI safety legislation (September 2026); Source: liccardo.house.gov). The administration's channel, discussed above, is led by Treasury Secretary Bessent and centers on incident notification; the bill assigns negotiation to the Secretary of State and names a treaty as its objective. See Liccardo–Kiley bill.

Track II proposals on military AI and nuclear command and control

On September 9, 2026 the Brookings Institution published recommendations from one US and one Chinese participant in the Track II Dialogue on Artificial Intelligence and National Security, which its Foreign Policy program and Tsinghua University's Center for International Security and Strategy (CISS) have convened since 2019 with support from the Minderoo Foundation; the recommendations were presented in anticipation of the government-level AI dialogue. The dialogue also maintains a living glossary of AI and national-security terms. Both authors build on the late-2024 agreement between the US and Chinese presidents that humans, not AI, should control decisions to use nuclear weapons (Advancing human control of military AI (Sisson and Jiang, Brookings, September 2026)), which Reuters dates to November 2024 and attributes to then-President Joe Biden and Xi (Source: reuters.com).

Melanie W. Sisson, a Brookings senior fellow, argues that Trump and Xi should agree that only humans, not AI, make the decision to initiate a cyberattack against each other's nuclear command, control, and communications systems (NC3) or critical infrastructure. In her account AI-powered cyber capabilities could start a military crisis not only through an intentional attack but through an unplanned one, arising from human error, limitations in the AI models or agents acting in ways humans did not anticipate while a capability is only being pre-positioned, or through unauthorized use by a bad actor; she describes leader-level statements as "the necessary precondition for bilateral dialogue on verifiable protocols and safeguards" (Advancing human control of military AI (Sisson and Jiang, Brookings, September 2026)).

Tianjiao Jiang, an associate professor at Fudan University's Development Institute, cites the two heads of state's May 2026 proposal to build a "constructive strategic stability relationship" and proposes three mechanisms. The first is a list of red lines for military AI prohibiting AI from autonomously launching attacks on NC3 systems or autonomously deciding to use nuclear weapons, together with a defined list of critical-infrastructure sectors such as finance, health care and energy. The second is a consensus on what constitutes "meaningful human control," reached through regular dialogue or a terminology working group, because the two sides "may employ different terminology and adopt divergent approaches" even on the same issue. The third is a China-U.S. military hotline for AI, building on the defense-minister hotline and the Maritime Military Security Consultation Mechanism, for reporting emergency incidents triggered by AI errors or AI cyberattacks. Jiang names three practical obstacles (the speed of AI-enabled cyber exchanges, attribution, and a security dilemma produced by the opacity of an adversary's military AI) and urges the two countries to seek minimum common ground on "prohibiting fully autonomous lethal attacks" and "ensuring human oversight" for lethal autonomous weapons, given slow progress in the Convention on Certain Conventional Weapons (Advancing human control of military AI (Sisson and Jiang, Brookings, September 2026)). Reuters described the shared-definition proposal as intended to prevent both countries "from using identical language to justify different levels of machine autonomy" (Source: reuters.com); the commentary itself frames the problem as divergent terminology and asks for "mutual acceptance of each other's definitions." See Autonomous Weapons.

Other specialists and officials have addressed the same questions. Carla Freeman of the Johns Hopkins School of Advanced International Studies has questioned the utility of the existing US-China military crisis hotline, citing Chinese officials' refusal to answer US counterparts' calls about the February 2023 balloon incident, and wrote in January that China's top-down hierarchy leaves low-ranking officers unable to take the initiative with US counterparts. Lora Saalman of the Stockholm International Peace Research Institute wrote in February that human control over nuclear-use decisions was an important point of US-China convergence even as both countries showed signs of integrating AI more deeply into nuclear and nuclear-related systems. In China, AI is a formal portfolio of the Foreign Ministry's Department of Arms Control, and Shen Jian, China's ambassador for disarmament affairs, told a UN meeting in Geneva in June that military AI could affect strategic stability and raise risks of miscalculation and conflict escalation, and that weapons should remain under human control, while China argues AI safety should not become a pretext for technological barriers. Washington has no equivalent single arms-control home for AI. State Security Minister Chen Yixin wrote in a government outlet in September 2026 that AI "fundamentally transforms the form of military struggle." Neither government has publicly endorsed the Brookings-Tsinghua proposals (Source: reuters.com).

Calls on Chinese developers to pace or to accelerate

Rep. Ro Khanna (D-California), ranking member of the House Select Committee on the Chinese Communist Party, wrote to Alibaba chief executive Eddie Wu and chairman Joe Tsai in a letter dated September 16, 2026, urging China's leading AI companies to join US laboratories in "a binding, enforceable, and verifiable agreement" to pace the frontier. Its questions ask whether Alibaba would accept "global limits on unchecked recursive self-improvement" and "neutral third party auditing and all-access inspections," with written answers requested by September 22 (Khanna letter to Alibaba on pacing the frontier (September 16, 2026)). The letter disputes the argument that any slowdown "necessarily means ceding American AI leadership to Beijing," calling it "distinctly misguided, in part because it presumes global pacing is not a possibility," and calls for "an enforceable U.S.-PRC AI Treaty," with discussions to begin at the Trump–Xi meeting the following week. It also states that PRC AI models help modernize the People's Liberation Army and support PRC intelligence agencies, and that "Chinese law can compel such cooperation." The Washington Post, which published the letter on September 17, reported that Khanna had also written to Moonshot AI and DeepSeek, that he would host an emergency hearing the following Wednesday on international mechanisms to pace AI, and that Chinese officials and technology leaders had so far rejected calls to slow AI development (Source: washingtonpost.com). Only the Alibaba letter has been published.

Huawei rotating chair Eric Xu took the opposite position at a press conference on September 17, 2026, urging Chinese AI companies to accelerate development to catch up with frontier US laboratories. Chinese researchers needed to "increase the speed of development so they can also see the dangers of AI development" like their US counterparts, he said, and once those risks emerged the industry would need to "strike a balance between driving development and risks." He also said at least six or seven Chinese AI laboratories aimed over the next two years to train models containing 10tn to 40tn parameters (Source: ft.com). The Foreign Ministry's warning the same week against "fear-mongering" is described under the September 2026 talks above; see Coordinated slowdown proposals compared and Huawei — Ascend AI Accelerators.

US Treasury Secretary Scott Bessent, US Trade Representative Jamieson Greer and Chinese Vice Premier He Lifeng were scheduled to begin all-day talks at about 10:30 a.m. ET on September 20, 2026 at JP Morgan Chase's Manhattan headquarters, to prepare possible agreements on AI, tariffs and critical minerals for a Trump–Xi summit in Washington that week. Bessent said on September 18 that the discussions would cover "both open- and closed-weight models," and has called for the two countries to agree AI "guard rails" aimed at keeping powerful models out of the hands of malign non-state actors. The US–China trade truce expires on November 10; a senior US official said on September 18 that China's performance on restoring critical-mineral flows "has not been up to par" (Source: reuters.com).

At the close of the September 20 talks Bessent told reporters that the two sides had discussed setting up a new US–China AI dialogue, with particular emphasis on national security and a notification system for "common goals and common threats" covering AI-related incidents that rise to a national security level, which the US side proposed for Trump and Xi to consider at their summit that week, and that the two countries planned to meet again to scope the notification system. "We think that, just like with any cross-border activity, that moving from opaque to more transparency between the number one and the number two AI powers in the world is very important," he said. Trump and Xi had first discussed possible consultations on AI in Beijing in May 2026, but that forum was never formalized. The Chinese response was unclear: the Xinhua readout acknowledged only briefly that the two sides spoke about AI, describing "frank, in-depth and constructive exchanges on key economic and trade issues," and He Lifeng and chief trade negotiator Li Chenggang left without speaking to reporters. USTR Greer said US export controls on advanced AI chips and semiconductor manufacturing equipment were not on the agenda for the AI mechanism talks, and that the day prepared the groundwork for the summit in Washington the following Thursday and Friday. George Chen of The Asia Group said future sessions were expected to address the weaponization of AI, AI safety principles, critical-infrastructure protection and cyberattack prevention, while prospects for cooperation remained limited by low trust and Beijing's perception that Washington seeks to contain China's AI development. Rep. Ro Khanna, ranking Democrat on the House Select Committee on the CCP, said on CBS's Face the Nation the same day that a Trump–Xi agreement should ban "recursive self-improving AI" and include safeguards against AI misuse for biological and nuclear weapons (Source: reuters.com; theinformation.com).

On CNBC's "Squawk Box" on September 21, 2026, Bessent said the two sides had agreed during the talks to a formalized AI dialogue, which he called the "USA-China AI dialogues," including "a communications line, an incident line" for AI safety incidents, and that senior officials would meet again in about two months in Shenzhen to discuss AI dangers and communications protocols, naming uncontrollable agents and non-state cyber actors as candidate categories. He also said OpenAI's Hugging Face incident "is the responsibility of the OpenAI management, not a bunch of agents," adding that labs "can slow down anytime they want to" (Source: reuters.com). An account that became public on September 22, attributed to four people familiar with the arrangement, described the mechanism as an open line of communication between Bessent and He Lifeng rather than a formal body of experts, and reported that Chinese officials had not publicly agreed to it; China's Foreign Ministry did not answer a question on AI at its September 22 press conference (Source: politico.com).

Trump administration officials said on September 23 that the incident-notification system would take a number of weeks to finalize, with the aim of having a framework agreed to by Trump by the time Bessent and He meet again in Shenzhen. Unresolved elements included benchmarks for which incidents rise to a national-security issue and explicit sign-on from China, which had been noncommittal. Some officials said the system could also wait on Trump's choice of an AI czar; according to two sources, Bessent, Commerce Secretary Howard Lutnick and White House AI officials Michael Kratsios and Sean Cairncross were in contention. The article body beyond its opening was not retrieved for verification (Source: wired.com).

Senate Minority Leader Charles Schumer (D-N.Y.) and other senior Democrats criticized Trump ahead of the September 24 Trump–Xi meeting in Washington, citing continued chip sales to Chinese entities and his approach to AI, while AI safety advocates urged US–China collaboration during Xi's state visit; the criticism became public on September 23 (Sources: insideaipolicy.com; reuters.com). At the UN General Assembly on September 22, Trump said the United States "totally rejects any attempt to construct a globalist scheme to control" artificial intelligence (Source: theverge.com).

Trump and Xi held talks at the White House on the morning of September 24, 2026, during Xi's state visit. According to the readout published by China's Ministry of Foreign Affairs on September 25, Xi said that both countries "have strong AI capabilities. They may have competition, but they may cooperate even more," that "the right thing to do is to draw on each other's strengths, not guard against each other," and that the two sides "can continue their dialogue on AI, exchange views on its risks and benefits, and jointly prevent the misuse and abuse of AI. AI must be kept under human control." The readout attributes to Trump the statements that "AI concerns the future of humanity" and that the two countries "should maintain dialogue and strengthen cooperation on AI." It reports that the economic and trade teams had reached "a new joint arrangement" and describes no specific AI agreement or incident-notification mechanism (Chinese MFA readout: President Xi Jinping Holds Talks with U.S. President Donald J. Trump (September 24, 2026); Source: fmprc.gov.cn). It is the Chinese government's account; no U.S. readout of the meeting is recorded. The readout's AI paragraph repeats the "under human control" and "for the positive and for good" formulations of Xi's July World AI Conference keynote and September BRICS statement, and Xi is recorded as citing the May 2026 agreement to build "a constructive China-U.S. relationship of strategic stability." Axios reported that Trump entered the AI discussions "showing little interest" in doing much and that the Washington meetings ended with no major breakthroughs. The next day Rep. Ro Khanna set out, in a letter to Secretary of State Marco Rubio, five provisions he wants in a U.S.–China AI agreement, among them global pauses on recursive self-improvement and on superintelligence, shared "kill switch" protocols, and inspections possibly run by an international agency or nongovernmental organization (Source: axios.com). Sen. Todd Young separately asked the National Security Council on September 24 to convene recurring roundtables whose agenda would include the capabilities of AI models developed in China (Letter from Sen. Todd Young to National Security Advisor Marco Rubio on frontier AI national security threats (September 2026); Source: young.senate.gov).

After the three-day summit ended on September 25, 2026, the White House and China's Foreign Ministry said the two countries had agreed to hold a dialogue on AI's risks and benefits, with the next round set for November, and to set up a communication channel for AI-related incidents. The White House said the leaders had agreed to use the term "super intelligence" in place of "artificial intelligence"; the Chinese ministry said Beijing valued Washington's use of the new term. The same announcements covered more favorable tariff treatment for $30 billion of non-sensitive goods in each direction and a new trade council (Source: reuters.com). The incident channel corresponds to the notification system officials had said on September 23 would take weeks to finalize; the announcements as reported do not settle the thresholds or Chinese sign-on that account described as unresolved.

Supply chains, chips, and self-sufficiency

China's approach to chip access at the summit

China declined H200 access on May 14, 2026 to preserve Huawei's domestic chip program, a decision Chan reads as evidence that Beijing is "not A.G.I.-pilled" (see Chan-Brookings framework above). The cleared but undelivered H200 sales to 10 Chinese firms (above) reflect the same posture.

China's Ministry of Industry and Information Technology (MIIT) signaled before Xi's Washington visit that it may allow purchases of Nvidia's RTX Pro 5500, according to an account that became public on September 27, 2026 citing two people familiar with the matter. The ministry asked companies including Alibaba Group and ByteDance to report how many of the chips they want and for what use, and told some companies the government intends to approve the purchases; when approvals would come, how many chips would be allowed and on what standard were unclear, and the U.S. government had not publicly addressed whether the chip falls outside export restrictions. ByteDance is weighing an order of about 1 million. Nvidia aims to start shipping in late December and targets about 500,000 chips a quarter for China, and its sales staff told customers to order by September 30. Two chip buyers put the price in China at 85,000 to 90,000 yuan (about $13,000), roughly that of Huawei's Ascend 950PR, while the China-specific RTX Pro 6000D sells for about 65,000 yuan, up from about 50,000 yuan the previous year. H200 sales to Chinese buyers were less than 1% of Nvidia's data center revenue in the quarter ended July 26, 2026 (Source: theinformation.com). The reported inquiry contrasts with Beijing's May decision to decline H200 access.

Huang on Huawei and Taiwan (May 2026)

On May 25, 2026 Nvidia CEO Jensen Huang said in Taipei that Nvidia has "largely conceded" China's AI-chip market to Huawei — described as the highest-status US-industry acknowledgment that the export-control regime has produced capability convergence rather than capability degradation in China. The remarks paired with Huawei's same-day Shanghai-symposium announcement of the "Tau Scaling Law" and LogicFolding architecture, targeting 1.4nm-equivalent density by 2031 without EUV, with 381 chips already mass-produced on Tau Scaling principles. The associated tracking question shifted from whether China can catch up toward whether the US can re-engage Chinese AI standards and customers at the negotiating table that the Trump-Xi dialogue defines.

On May 27, 2026 Huang said Nvidia plans to spend up to $150 billion annually on Taiwan AI suppliers, calling the island "the epicenter of the AI revolution" (Source: reuters.com). The figure is the public commitment underpinning a Taiwan-anchored Nvidia supply stack and pairs with the May 23-25 Huang comments above, including the Taiwan-smuggling acknowledgment.

Huawei's Ascend schedule and domestic supply (September 2026)

At its Huawei Connect conference in Shanghai on September 17, 2026, Huawei brought forward its next accelerator generation. Rotating chairman David Wang said the Ascend 960DT would be ready in the first quarter of 2027, three quarters earlier than previously planned, and the 960PR in the third quarter of 2027, one quarter ahead of schedule, with the Ascend 970 and 980 to follow in 2028 and 2029 as part of a new generation each year (Source: reuters.com). At Huawei Connect 2025, on September 18, 2025, the company had scheduled the Ascend 960 for the fourth quarter of 2027 and the Ascend 970 for the fourth quarter of 2028 (Source: huawei.com).

Rotating chairman Eric Xu said at the same conference that Huawei cannot produce enough AI computing equipment to meet demand in China and is limiting overseas sales: "Since we don't have enough capacity to even satisfy the demand in China, we don't have a plan to expand into the international market in a fully-fledged way," with limited volumes supplied to some countries where demand is particularly strong. He said reliable data on Nvidia's share of China's AI chip market was difficult to obtain but that "I think Ascend market share should be bigger than Nvidia's"; Huawei did not provide data to support the estimate. Xu tied the program to self-sufficiency: "We cannot accept a destiny where we cannot control our fate being determined by others in terms of willingness to sell chips to China or not" (Source: reuters.com). The market-share claim concerns sales inside China and is consistent in direction with Huang's May remark that Nvidia had "largely conceded" that market; it does not bear on total output, where Epoch AI estimates Huawei will produce under 4% of Nvidia's AI compute in 2026 (Will Huawei catch up to Nvidia by 2030? (Epoch AI, September 2026)). The Peerium architecture, supernode sizes and company-reported deployment figures announced at the conference are on Huawei — Ascend AI Accelerators.

ByteDance custom-chip move (May 2026)

As reported on May 28, 2026, ByteDance approached external partners to design a custom AI chip modeled in part on the architecture of Nvidia partner Groq — its first publicly surfaced move toward custom in-house silicon, alongside the prior thread of Huawei-Ascend procurement and Nvidia H20/H200 license-seeking (Source: pcgamer.com). The move makes ByteDance the second major Chinese AI player, after Huawei, pursuing a domestic silicon stack, though ByteDance leans on third-party design partners modeled on Groq's architecture rather than fully in-house design as in Huawei-Ascend. Groq's architectural lineage — a deterministic LPU optimized for high-throughput, low-batch inference — suggests ByteDance is optimizing for inference scale rather than training-frontier capability, consistent with the Chan-Brookings reading of China optimizing for deployment-scale inference while the US optimizes for frontier-training scale.

China self-sufficiency milestone (New York Times, May 2026)

On May 12, 2026, ahead of US-China summit week, NYT's Meaghan Tobin reported that DeepSeek's latest model release weakens Trump's leverage by showing China is closer to AI self-sufficiency than the export-control regime assumed (Source: nytimes.com). Tobin frames the mid-2026 negotiating posture as shifted: a US negotiating asset (export controls plus frontier-chip access) is depreciating faster than the corresponding Chinese asset (rare-earth processing and magnet manufacturing). The reporting is read alongside the Apple-Intel preliminary chip-making agreement (May 10) and the Tencent and Alibaba $20B DeepSeek-infrastructure commitment (April 22).

Tobin's framing bears on whether DeepSeek V4 will beat the best US frontier-lab model on at least one major capability benchmark by end-2026: if the framing holds, gap-narrowing is occurring across more than one axis at once (compute access, capability gap, and government coordination). May 6 Reuters reporting on DeepSeek V4 (Huawei Ascend 950PR-optimized) anchors that possibility; a resolution would be a public benchmark score (GPQA, MMLU-Pro, or HumanEval) on which the V4 score exceeds the best published OpenAI, Anthropic, or Google score on the same benchmark version by December 31, 2026.

Top500 supercomputer lead (June 2026)

In the Top500 supercomputer ranking released the week of June 22, 2026, China's LineShine system took the No. 1 position, overtaking US machines as the world's fastest (Source: wsj.com). The ranking is a symbolic marker rather than a direct measure of frontier-AI training capacity — Top500 benchmarks general high-performance computing on a standardized linear-algebra test, while AI training runs on accelerator clusters whose scale the list does not fully capture — but it adds to the body of evidence (alongside the DeepSeek releases and Huawei's Tau Scaling claims above) that the assumption of durable US compute supremacy is contested at the high end, and it lands against the backdrop of export controls meant to constrain Chinese access to the most advanced accelerators.

Allied coordination

Quad Critical Minerals Initiative (May 2026)

On May 26, 2026, on the margins of a Quad foreign ministers' meeting in New Delhi, the United States, India, Japan, and Australia launched the Quad Critical Minerals Initiative Framework (Quad Critical Minerals Initiative Framework (May 26, 2026)). The framework sets a $20B mobilization target and three pillars: (I) investment and project development; (II) regulatory alignment, including a CFIUS-equivalent investment-screening provision aimed at adversary capital; and (III) recycling and recovery. It is directly responsive to China's 2025 export controls on 14 rare-earth materials, which remain in force. The primary text speaks of "advanced technologies" and does not name AI explicitly; the "AI development" framing is secondary-press framing (Inside AI Policy, 2026-05-27) rather than primary-text framing (Press framing: insideaipolicy.com). The framework is nonetheless AI-relevant because the binding choke points for frontier-AI hardware (HBM, advanced packaging, GPU silicon, magnets, and batteries) all run through critical-mineral supply chains. It pairs with the "node strategy" framing in Clover's "New Arms Race in Computing Power", under which middle powers own one indispensable node of the AI compute stack.

European Commission moves to join Pax Silica (June 2026)

On June 2, 2026 the European Commission moved to recommend that the EU join Pax Silica, the US-led initiative launched in December 2025 to coordinate allied work on chips, critical minerals, and AI models. A preparatory document for a June 3 ambassadors' meeting stated the bloc "should join Pax Silica as a signatory member." The step is in tension with the Commission's parallel push for tech sovereignty and is a reversal of the EU's initial decision to stay out; Sweden and Finland had already signed on individually (Source: politico.com). Pax Silica is the allied-coordination counterpart to the Quad Critical Minerals Initiative; together they describe a US-anchored bloc aligning hardware, minerals, and model governance against China's vertical-integration push.

Quad-versus-China posture

Across the May 26-27 cycle the operative structural picture was that the US/Quad bloc uses supply-chain diversification as its main lever (the Quad framework, the BIS Diffusion Rule, and chip controls), while China uses personnel and capital containment (travel curbs, foreign-investment review, and founder restrictions). Neither side was using the AGI-race framing of 2023-2024; the operative competition sat at the supply-chain and talent-mobility layer.

China's domestic governance posture

Talent and capital containment (May 2026)

According to reporting on May 27, 2026, the Chinese government expanded existing travel curbs to cover top AI researchers, startup founders, and executives at private firms, with some of the most prominent figures now required to seek government approval before traveling abroad. The expansion builds on March 2025 advisories to top AI founders to avoid travel to the United States and on the barring of the two co-founders of Manus from leaving the country while Beijing investigates whether Meta's $2B acquisition of Manus runs afoul of foreign-investment rules; the Manus co-founders are reportedly exploring a $1B external raise to unwind the Meta deal (Source: techcrunch.com). See Manus (Butterfly Effect) and AI Talent Flows.

Beijing's tightening posture (April 2026 Economist reportage)

The Economist's April 16, 2026 China feature, "Why China's government worries about AI," documents converging Beijing-side trends:

• The OpenClaw / "lobster farming" episode. Hundreds of pensioners and unemployed graduates queued outside Tencent HQ to install OpenClaw in early March 2026, and the Chinese user count quickly outpaced the US. Authorities then slowed adoption: media reported sensitive-data deletions and compute exhaustion; regulators banned OpenClaw in banks and other sensitive sectors; and a paid-uninstallation cottage industry emerged. After the craze, the government tasked an expert group with proposing safety standards and rules governing agent behavior (Source: economist.com).
• Robotaxi backlash. In the same period, more than 100 robotaxis in Wuhan suddenly stopped, stranding customers on highways for up to two hours. Senior officials including the Ministry of Public Security met to discuss, and Wuhan cab drivers petitioned to slow the robotaxi rollout (Source: economist.com).
• No-fire-by-AI rule (December 2025). Beijing's government published a ruling that firms cannot fire employees replaced by AI; state-media commentary called firing-by-AI an "offloading" of technological-change risks onto workers. A government white paper on AI's employment impact is reportedly in the works.
• AI-boyfriend edict. A new edict prohibits systems from encouraging self-harm or emotional dependency, after an apparent rise in "AI boyfriends." It is China's structural counterpart to the US policy responses tracked under Parasitic AI / Spiral Personas and AI Mental Health and Psychological Harm, and parallels California SB 243 in scope.
• Worker AI-replacement fears. Such fears rose from 49% (2024) to 59% (2025) per a state-think-tank report. Youth unemployment of roughly 16%, about double the US level, magnifies the political stakes.
• Xi Jinping's January 2026 speech reportedly raised "security problems" including data theft and, perhaps for the first time, a potential technical loss of control over frontier AI — a departure from the 2017 framing of AI as "controllable" via domestic production.
• Mythos response. Anthropic's Mythos announcement (April 7, 2026) drew a "muted reaction" publicly in China but was being privately examined by national-security authorities; Carnegie's Matt Sheehan called it a "second wake-up call after ChatGPT" (Source: economist.com; cross-reference Claude Mythos Preview for the broader global response).
• The trusted-academic loop. Communist Party regulatory advice flows through "a coterie of trusted academics at leading universities" rather than directly from industry, a structural counterweight to the firms (Tencent, Alibaba, ByteDance) that boosted OpenClaw mania. Liang Wenfeng (DeepSeek) received a televised handshake with Xi; Yan Junjie (Minimax) and Yang Zhilin (Moonshot AI) have briefed Premier Li Qiang; and Z.ai (founded by Jie Tang of Tsinghua) has deep government connections.
• Government adviser quote (anonymous, Beijing). An adviser described winning the AI race as "life or death" for the Party, stating that if forced to choose between technology goals and other priorities, the Party will prioritize technology goals above all else. The Economist frames this as the binding constraint limiting how far concerns about jobs and safety can go before being subordinated to the race.

The Economist's April 16, 2026 state-think-tank report projects, on its own terms, continued rises in Chinese workers' AI-replacement fear; the relevant resolution marker would be a successor state-think-tank survey publishing a 2026 figure of 65% or higher. Separately, the Beijing "AI-firing prohibition" (the December 2025 no-fire-by-AI rule, traced to a Hangzhou Court ruling, plus the planned Beijing white paper) raises the question of whether it will be mirrored in an OECD jurisdiction, which would be marked by an enacted national or state law in an OECD country prohibiting AI-replacement as a lawful firing basis.

Western AI-safety discourse in Chinese media (Duff, AI Frontiers, May 2026)

Calvin Duff, a British diplomat and GovAI fellow, published a May 18, 2026 AI Frontiers analysis finding that mainland Chinese tech and finance media are giving serious, sympathetic, and mostly neutral coverage to four Western AI-safety essays: Aschenbrenner's *Situational Awareness*, Hendrycks-Schmidt-Wang's *Superintelligence Strategy*, Kokotajlo et al.'s *AI 2027*, and Dario Amodei's *The Adolescence of Technology* (Source: aifrontiersmedia.substack.com).

The analysis's empirical anchors:

• 61 Chinese-language primary sources examined, of which 85% adopted neutral framing.
• Combined readership likely topped one million for both AI 2027 and The Adolescence of Technology.
• Mainland outlets giving substantive coverage included AI Era (新智元), Synced (机器之心), Sina Finance, Wall Street News (华尔街见闻), 21st Century Business Herald, and Zhihu.
• Only 43% mentioned US-China competition; coverage focused on the substantive AI-safety arguments rather than the bilateral framing.
• Wall Street News selectively omitted Amodei's paragraph singling out the CCP as an authoritarian threat while keeping the rest of the translation intact — a pattern Duff attributes to editorial self-censorship rather than state instruction.

Duff's finding cuts against the assumption that the Chinese state controls information access tightly enough that Western AI-safety arguments do not penetrate: per the analysis, Chinese policy elites, technical workers, and finance professionals are reading the same arguments US AI-safety researchers are reading, in mostly neutral framing. This complicates both the PEAT framework critique of US export controls (Sun), which treats the Chinese incentive architecture as opaque to outside influence, and the assumption that strategic-content alignment in China's gen-AI registration forecloses serious engagement with Western safety frameworks. The selective Wall Street News omission supports a reading in which Chinese-language coverage of Western AI safety can be serious and substantive precisely because the politically sensitive portions are routinely edited out, with editorial self-censorship as the operative filter rather than state suppression of the underlying ideas. Duff's analysis is the primary source examined on the penetration of Western AI-safety discourse in Chinese policy and technical communities.

China's industrial-policy priorities

15th Five-Year Plan (March 2026)

China's National People's Congress adopted the 15th Five-Year Plan with "embodied intelligence" among ten priority industry tracks. The plan directs a national buildout of robot "training grounds" to generate proprietary physical-AI training data at scale, confirming the applications-and-robotics rather than AGI strategic frame (see World Models).

MIIT Information and Communications Industry 15th Five-Year Plan

The Ministry of Industry and Information Technology's sectoral five-year plan for the information and communications industry over 2026–2030 was signed August 12, 2026 and released publicly on September 7, 2026 as circular 工信部规〔2026〕188号 (MIIT — Information and Communications Industry Development 15th Five-Year Plan (Gong Xin Bu Gui [2026] No. 188)). It is a communications-industry plan implementing the national 15th Five-Year Plan at sector level, covering networks, spectrum, satellites, submarine cables, telecom market regulation, security and 6G alongside compute; intelligent computing is one of thirteen headline indicators rather than the plan's subject.

On compute, the plan sets intelligent computing scale at 1,590 EFLOPS in 2025 rising to 9,800 EFLOPS in 2030, a cumulative change of 8,210 and a rise of roughly 6.2 times over five years on its own baseline, alongside advanced storage capacity of 540 EB rising to 1,700 EB, cumulative information-infrastructure investment of 3.8 trillion yuan ($532 billion) over the period, and a PUE ceiling below 1.2 for newly built large and hyperscale computing facilities. It directs "orderly deployment of ten-thousand-card, hundred-thousand-card and above intelligent computing clusters" and "greatly increased adaptation to domestically produced computing chips" (MIIT — Information and Communications Industry Development 15th Five-Year Plan (Gong Xin Bu Gui [2026] No. 188)). The domestic-chip commitment is the plan's response to constrained access to foreign accelerators; see Export Controls (AI).

The National Data Administration reports a separate series for installed capacity: 2,185 exaflops at the end of June 2026, up 177 per cent year on year across 52 facilities each holding more than 10,000 accelerator cards, and about 2,450 exaflops by the end of July. Measured against that series the 2030 target is roughly fourfold, as the South China Morning Post framed it (Source: scmp.com); measured against the plan's own 2025 baseline it is roughly sixfold. The two figures are not interchangeable. See AI Data Centers and Compute Governance.

Beyond compute, the plan carries AI-governance content that the compute coverage did not report. Column 5 sets out an "agent internet" programme (智能体互联网络) to upgrade the domain-name system into an "internet agent identity identification system," build agent registration management, identity authentication, ownership determination and address resolution capabilities using blockchain and electronic certification, establish international nodes interoperable with "mainstream international technical systems," and construct monitoring, early-warning and security-control capability "over the full lifecycle of agents." The regulatory counterpart appears in task 10, which commits to establishing management systems for intelligent information services and the agent internet. This extends the agent-registration approach of the Implementation Opinions on Intelligent Agents from non-binding policy into a funded infrastructure programme with an international-interconnection objective. The plan also names deep-synthesis forgery detection, training-data protection and agent security as research priorities, commits to an anti-fraud vertical large model under an "AI + anti-fraud" project, and lists network-native AI and token (词元) communication among 6G research priorities (MIIT — Information and Communications Industry Development 15th Five-Year Plan (Gong Xin Bu Gui [2026] No. 188)).

US public opinion on the race

A Pew Research Center survey of 3,488 US adults conducted June 22–28, 2026 and published July 23, 2026 found that Americans said by roughly three to one that China (36%) is more advanced than the United States (12%) in developing AI, with 18% calling them equal and 33% unsure — a perception at odds with the roughly seven-month capability gap documented above. Forty-three percent said US AI leadership is extremely or very important, split 54% of Republicans against 34% of Democrats and 56% of those 65 and older against 26% of adults aged 18 to 29. Fifty-one percent said AI will widen the gap between rich and poor countries, against 7% who said it will narrow (Source: pewresearch.org). See Public Opinion on AI.

The annual Gallup–Special Competitive Studies Project survey asks about competition with "other countries" and does not name China. Its first wave, 3,128 U.S. adults fielded April 25 to May 5, 2025, found 79% saying it is important for the United States to have more advanced AI technology than other countries and 85% agreeing that countries are already competing for the most advanced AI. On current standing, 34% said the United States is keeping pace, 32% did not know, 22% said it is falling behind and 12% that it is moving ahead, and 15% thought it very likely that the United States will have the world's most advanced AI in ten years. The share calling U.S. AI superiority very important ran from 26% of adults aged 18 to 29 to 64% of those 65 and older. More favored developing AI with a large group of allies (42%) than a smaller group (19%) or alone (14%), and 87% said foreign governments are likely to use AI to attack the United States within 20 years (Reward, Risk, and Regulation: American Attitudes Toward Artificial Intelligence (Gallup and SCSP, 2025)). The 2026 wave's report does not include these items (American AI Sentiment in 2026 (Gallup and SCSP)).

Public trust in the US, China and the EU as AI regulators

Outside the United States, the Pew Research Center asked adults in 12 countries in spring 2026 how much they trust the United States, China, the European Union and their own country "to regulate the use of artificial intelligence effectively." Across the 12, which are mostly middle-income (Bangladesh, Chile, Colombia, Ghana, Malaysia, Pakistan, Peru, the Philippines, Singapore, Sri Lanka, Thailand, and the West Bank and East Jerusalem), median trust was 43% for China, 35% for the United States and 34% for the EU. China was the most trusted of the three in five places: Bangladesh, Malaysia, Pakistan (87% against 28% for the United States), Sri Lanka, and the West Bank and East Jerusalem. The Philippines was the only country where the United States was trusted more than both China and the EU (53% against 36% and 44%); in Ghana, Peru and Colombia the United States and China were trusted about equally, and in Chile and Thailand all three were trusted at about the same rate. In half the countries respondents trusted their own country more than any of the three. Trust tracked general favorability toward each power: 57% of Peruvians with a favorable view of China trusted it to regulate AI, against 26% of those with an unfavorable view (Globally, More People Expect AI to Cause Job Loss Than Growth (Pew Research Center, 2026)).

The result differs from Pew's 2025 survey of 25 countries, in which people tended to trust the EU most, then the United States, then China. The two surveys cover different sets of countries, and Pew notes that people in middle-income countries tended to view China more positively than those in high-income countries in its separate 2026 analysis of views of China (Globally, More People Expect AI to Cause Job Loss Than Growth (Pew Research Center, 2026)).

Related domestic externalities

The NTSB-Codex pilot-voice-reconstruction incident (May 21, 2026) sits adjacent to the US-China thread. Internet users used OpenAI's Codex to reconstruct cockpit-voice audio from spectrograms the NTSB had released for the UPS Flight 2976 investigation, forcing the agency to suspend public access to its entire civil-accident docket database (Source: arstechnica.com). It is a US-domestic accountability and transparency externality, distinct from the US-China competition but evidence that AI capability is propagating faster than US institutional response cycles, whether the institutional layer is BIS export controls or NTSB document-release procedures.

Sources

• China and the US Are Running Different AI Races, (Source: epochai.substack.com), Geopolitics in the Age of Artificial Intelligence, Export Controls (AI), Eight Worlds Framework
• Fast-Follow Problem, AI Software Progress, Distillation, Compute Governance
• Broad Timelines, Clawed, America's AI Action Plan
• DeepSeek, OpenAI
• The Incentive Architecture Export Controls Cannot Reach (Sun, Lawfare, May 6 2026) — Sun PEAT framework, da moxing bei'an, and Manus saga (May 6, 2026)
• PEAT — Proactive Elite Alignment Theory, China Generative AI Registration (da moxing bei'an / 大模型备案) (anchored by Sun)
• 2028: Two Scenarios for Global AI Leadership (Anthropic) — Anthropic four-fronts framework and two 2028 scenarios (May 13, 2026)
• (Source: axios.com) — Bessent statement on AI talks with He Lifeng (September 16, 2026)
• Advancing human control of military AI (Sisson and Jiang, Brookings, September 2026) — Sisson and Jiang, Brookings–Tsinghua CISS Track II recommendations on military AI (September 9, 2026)
• The current balance of power in open models (Nathan Lambert, September 2026) — Lambert's prepared remarks for a congressional briefing on open-weight models (September 21, 2026)
• Khanna letter to Alibaba on pacing the frontier (September 16, 2026) — Khanna letter to Alibaba urging a binding frontier-pacing agreement (September 16, 2026)
• (Source: washingtonpost.com) — Washington Post on the Khanna letters and hearing (September 17, 2026)
• (Source: ft.com) — Financial Times on Eric Xu's call for Chinese labs to accelerate (September 17, 2026)
• (Source: reuters.com) — Reuters on the Track II proposals and official positions (September 17, 2026)
• (Source: reuters.com) — Huawei Connect 2026 Ascend schedule and Xu remarks (September 17, 2026)
• (Source: huawei.com) — Huawei Connect 2025 Ascend roadmap (September 18, 2025)
• Who's Ahead in the Global AI Talent Race? (Ma and Yang, Carnegie China, 2026) — Ma and Yang, Carnegie China AI Talent v3.0 study of NeurIPS author career paths (September 23, 2026)
• Chinese MFA readout: President Xi Jinping Holds Talks with U.S. President Donald J. Trump (September 24, 2026) — Chinese Ministry of Foreign Affairs readout of the Xi–Trump White House talks (September 24, 2026)

Relationships

• related: Liccardo–Kiley U.S.–China frontier AI safety negotiations bill (2026), Negotiating AI Safety Globally — Liccardo–Kiley outline of U.S.–China frontier AI safety legislation (September 2026) — congressional proposal for State Department negotiations with China on frontier AI safety.
• related: Building Standards for the Next Phase of AI (OpenAI, September 2026) — OpenAI's September 2026 call for US-led international frontier-AI standards and US–China security channels
• related: China and the US Are Running Different AI Races
• related: Export Controls (AI)
• related: Eight Worlds Framework
• depends-on: PEAT — Proactive Elite Alignment Theory
• related: Scott Bessent
• related: Huawei — Ascend AI Accelerators
• related: Autonomous Weapons
• related: Brookings Institution AI Initiative
• related: Advancing human control of military AI (Sisson and Jiang, Brookings, September 2026) — Track II proposals on human control of military AI
• related: Khanna letter to Alibaba on pacing the frontier (September 16, 2026) — congressional request to Chinese developers to join a pacing agreement
• related: Ro Khanna, Melanie W. Sisson, Tianjiao Jiang
• related: Quad Critical Minerals Initiative Framework (May 26, 2026)
• related: Quad Critical Minerals Initiative Framework
• related: Chip Smuggling and Export-Control Evasion
• related: 2028: Two Scenarios for Global AI Leadership (Anthropic)
• related: Superpowers and AGI: How U.S.-China Competition Could Be Shaped by the Emergence of Artificial General Intelligence (CNAS, August 2026)
• related: A New Counterintelligence Map for the AI Age (Feith, HPSCI, 2026-09-02) — counterintelligence framing of the competition's private-sector surface
• related: Globally, More People Expect AI to Cause Job Loss Than Growth (Pew Research Center, 2026) — trust in the US, China and the EU to regulate AI in 12 mostly middle-income countries
• supports: Who's Ahead in the Global AI Talent Race? (Ma and Yang, Carnegie China, 2026) — research-talent balance
• related: AI Talent Flows
• related: Chinese MFA readout: President Xi Jinping Holds Talks with U.S. President Donald J. Trump (September 24, 2026) — Chinese account of the September 24, 2026 summit