Washington and Beijing disagree on chips, trade, Taiwan and almost everything else. On one point they are converging: no one should lose control of AI. It may be the only thing humanity, as a species, can still agree on.
In the space of four days this month, the two AI superpowers each reached for the same lever. On 16 October Beijing published draft rules to restrict the export of its most capable model weights and AI agents, explaining that the aim was to ensure “that AI remains under human control” (PBS NewsHour). On 19 October Washington barred civilian federal agencies from using DeepSeek and banned the firm from federal contracts (Bloomberg). Both governments described their moves as security. Both were, in truth, the same act: a sovereign deciding who may touch the most powerful technology of the age.
This matters because the most consequential argument in American AI policy rests on a claim about China. The claim runs roughly as follows: Beijing will not restrain its labs, so any American restraint is unilateral disarmament; therefore America must race. It has shaped export controls, deregulation and the rhetoric of every frontier lab chief executive. A recent Economist column put the question bluntly: does China really prize AI speed over safety? The honest answer, after reading Beijing’s frameworks, standards and speeches, is that the question is wrongly posed. China prizes control over both. And so, it turns out, does Washington.
In January 2025 Ding Xuexiang, the Chinese vice-premier who oversees science and technology, told the audience at Davos that if the braking system is not under control, you cannot step on the accelerator with confidence (Concordia AI). Western safety advocates hear in that line an ally. Western accelerationists hear propaganda. Both miss what it actually says. Ding was not talking about slowing down. He was talking about control. And control, for the Chinese Communist Party, has never been a means. It is the end.
This essay follows an earlier piece on this site, The ImperAIl March of the Machines, which argued that artificial intelligence has become an imperial technology. If that is right, the most revealing thing about the two AI empires is not how they compete. It is what they fear. On the evidence of 2026, they fear the same thing: a technology that slips out of the hands of the state that built it. A timeline of the year’s key events appears at the end.
I. The useful caricature
Every empire needs a barbarian at the gate. It justifies the walls, the taxes and the haste. In the American AI debate, China plays that role with unusual consistency.
The argument reached its sharpest form in September. Dario Amodei of Anthropic published an essay calling on frontier labs to deliberately slow the pace at which they improve model capabilities, and Anthropic, OpenAI and xAI urged a slowdown within days of one another. Washington’s reply was swift and almost entirely about China. President Trump called the push a conspiracy from which only China benefited (NewKerala). David Sacks, the president’s former AI adviser, put the logic in a single sentence: “If we just pause all development, they will race ahead” (Stocktwits). He added that China was very unlikely to join any global agreement (Mitrade).
What is striking is that the safety camp accepts the same premise. Amodei told CBS News that the hardest dilemma in his own plan is what happens if China does not follow, which is why his essay pairs a slowdown with tighter chip controls (Mitrade). SemiAnalysis, the semiconductor research house, captured the symmetry precisely: for Trump, competition with China is a reason not to slow down; for Amodei, preserving America’s lead over China is what makes slowing down possible. Then it added the devastating line: nobody has asked what China is actually doing (SemiAnalysis).
The caricature has a long pedigree. In 2023 two American scholars wrote in Foreign Affairs that China’s blithe attitude to technological risk and reckless ambition were on a collision course with the dangers of AI (NBC News). It is convenient for almost everyone in Washington. It lets accelerationists dismiss regulation, lets hawks justify export controls and lets the labs argue that whatever they build is better than what Beijing would build.
The awkward fact is that the person with perhaps the strongest claim to have founded modern AI does not believe it. Geoffrey Hinton told Jon Stewart last year that after meeting a member of the Politburo trained as an engineer, he came away more optimistic about China’s grasp of the existential risk than about America’s; Chinese leaders, he said, understand that risk in a way American politicians do not (The Weekly Show, transcript). He has praised the engineering background of China’s leadership and contrasted it with Western governments’ weak regulation (Digitimes). One need not accept Hinton’s optimism to see the problem. A policy built on a caricature will fail the moment the caricature stops matching the territory. And the territory, on the evidence of 2026, looks very different from the map.
II. What Beijing has actually built
Start at the top, because in China that is where policy starts. In April 2025 the Politburo held its first collective study session on AI since 2018, and Xi Jinping told its members that AI brings unprecedented development opportunities and also unprecedented risks and challenges, calling for faster work on laws, risk warning systems and emergency response (CSET). This year his language hardened. At the World AI Conference in Shanghai in July, his in-person debut there, he used the term “loss of control” explicitly and argued that the faster AI advances, the more promptly safeguards against it must improve (DigiChina; SemiAnalysis). Days earlier, speaking to the nation’s scientists, he had described AI as an “increasingly apparent double-edged sword”, a shift that analysts at Trivium China read as moving from “might go wrong” to “is going wrong” (Trivium China).
The bureaucracy has followed. On 14 September TC260, the national standards committee that writes China’s cybersecurity and AI rules under the Cyberspace Administration of China (CAC), published version 3.0 of its AI Safety Governance Framework. Its catalogue of risks grew from 30 to 54, adding behaviours that would be familiar in any Western safety lab: models deceiving their evaluators and refusing instructions (It Does What Now?). It warns that AI has shown a self-accelerating trend towards autonomous learning and recursive self-improvement, and calls for vigilance about AI evolving beyond human prediction and control (SemiAnalysis). It adds a dedicated lifecycle framework for AI agents and, for the first time, a separate category for embodied AI and robotics (China AI Bulletin). One monitor noted the telling shift: version 2.0 was built on hypothetical future scenarios; version 3.0 is built on misaligned behaviours already observed in industry (China AI Safety Regulation Monthly Monitor).
Regulators have begun to speak like safety researchers. In May the CAC, the economic planner and the industry ministry issued joint guidance that named “operational loss of control” as a security risk for AI agents. On 1 September Wang Lihong, a senior CAC official, warned specifically about frontier models escaping sandbox environments, circumventing safety boundaries and attacking real-world production systems (Kuwait Times/Reuters). That is not the vocabulary of content censorship. It is the vocabulary of the frontier.
The institutional scaffolding is growing too. China launched its own counterpart to Western safety institutes, the China AI Safety and Development Association, at the Paris summit in February 2025, led by figures including Xue Lan, Zeng Yi and the Turing laureate Andrew Yao (MIRI). AI safety now sits in the National Emergency Response Plan alongside epidemics and cyberattacks, and China issued as many national AI standards in the first half of 2025 as in the previous three years combined (Concordia AI). Concordia AI, which tracks the field from Beijing and Singapore, found that Chinese frontier safety research grew by roughly 60 per cent over the past year, with agent safety rising from 8 per cent of new papers in early 2025 to 27 per cent in the first quarter of 2026 (Concordia AI). The state-backed Shanghai AI Lab has published and twice revised a Frontier AI Risk Management Framework (SemiAnalysis).
And a law is coming. The State Council’s 2026 legislative work plan, issued in May, commits the government to accelerate comprehensive AI legislation (DLA Piper). On 19 October Zhejiang University scholars published a 99-article model AI law, as the Ministry of Justice says it will speed the work (Caixin Global). Concordia’s summary of three years of tracking is the fairest one-line verdict: China’s approach has expanded from controlling what AI says to governing what AI does (Concordia AI). Whatever else this is, it is not a country that has never thought about the brakes.
III. Safety with Chinese characteristics
To read Beijing’s AI documents properly, one has to begin with a single word. The Chinese anquan means both safety and security. English forces a choice that Mandarin does not: the safety of a system, the security of a state and the stability of a regime are, in official prose, one concept. When Beijing speaks of AI anquan it means all three at once, and it is rarely clear which is doing the work.
This is not pedantry. It explains why Chinese AI governance can look simultaneously more and less serious than Western regimes. More serious, because the state has written frontier risks into national frameworks that Washington has never matched. Less serious, because the same frameworks treat a chatbot’s political opinions, a teenager’s screen time and a model’s ability to self-replicate as points on a single continuum of control.
The content side is formidable. China requires AI-generated material to be labelled, and its interim measures on anthropomorphic, human-like AI services, in force since 15 July, ban virtual intimacy services for minors and require anti-addiction prompts after two hours of continuous use (regulations.ai; MMLC Group). A draft Anti-Cyberviolence Law released in July would write AI detection, labelling and tracing duties into primary legislation (The Leveraged Years). Draft rules for interactive AI forbid content that endangers national security or undermines ethnic unity (Mayer Brown). This is safety as social order, and it is enforced.
The frontier side is formidable on paper and soft in practice. The new framework is not a law; it is guidance that shows how the state thinks (China AI Safety Regulation Monthly Monitor). TC260 standards are only recommendations unless the state makes them mandatory, and on frontier risk it has not (SemiAnalysis). The enforcement backbone remains the algorithm registry and deep-synthesis filings introduced in 2022, which, as one critique put it, record what a provider says about a system at launch but cannot show how a self-modifying model behaves afterwards (People of Internet). Even version 3.0 opens its principles by naming innovation and development as the first priority, and lists Western export controls among the supply-chain security risks it is meant to manage (SemiAnalysis).
So there are two Chinese safety regimes. One is a dense, enforced apparatus for controlling what AI says to citizens. The other is a sophisticated, mostly voluntary vocabulary for what AI might do to humanity. The first is about the Party’s security. The second is about everyone’s. The question that matters for the world is whether they converge, and there is a reason to think they will. A regime that has spent 30 years making sure no information flow escapes its command has the deepest possible instinct to fear a technology that could escape command altogether. To the Party, an AI that deceives its evaluators is not merely an engineering failure. It is a political one.
IV. The accelerationists in Zhongnanhai
None of this means the caricature is entirely wrong. It means it is describing the wrong part of the system. China has its own accelerationists, and they are powerful.
Zhu Songchun, a party adviser who runs a state-backed programme to build general intelligence, has compared AI development to “Two Bombs, One Satellite”, the Mao-era crash programme for nuclear weapons and missiles. Yin Hejun, the science minister, has used an old party slogan to argue that development, including in AI, is China’s greatest source of security (LessWrong, quoting The Economist). Some economic policymakers warn that an over-zealous pursuit of safety will cost China its competitiveness, and some of the more onerous restrictions were rescinded last year (Faster, Please!). At the top, Xi has called AI an “epoch-making” transformation comparable to the industrial revolution and the dawn of the internet (OODA Loop), and the 15th Five-Year Plan enshrines “AI Plus”, the drive to spread AI across every sector, as the country’s overarching AI policy (Concordia AI).
The labs, meanwhile, behave much like their American counterparts, only with less paperwork. SemiAnalysis compiled every identifiable model release by China’s nine leading developers from 2021 to mid-September this year: 857 of them. Thirty-one came with a safety test result published by the developer. None of the nine has published a frontier safety framework (Pebblous summary of SemiAnalysis). Every major frontier release of 2026, across DeepSeek, Moonshot, Alibaba, ByteDance and Zhipu, was undocumented at launch bar one, and DeepSeek’s founder, Liang Wenfeng, has made no public statement on safety or regulation in four years (LAVX News summary). The study’s verdict was blunt: nobody in China’s AI industry is slowing down, and nobody is asking them to (SemiAnalysis).
There are honourable exceptions. Zhipu, now trading as Z.ai, signed the scientists’ red-lines statement in Beijing in 2024 and was for a time the only Chinese company to sign the Seoul Frontier AI Safety Commitments (Pebblous). In December 2024 seventeen Chinese firms, DeepSeek among them, signed a domestic set of AI Safety Commitments that closely mirror the Seoul pledges (Carnegie Endowment). Inside the companies, candour occasionally surfaces: Alibaba’s security vice-president has acknowledged that safety capability lags model capability and that research must cover risks such as self-replication, and a DeepSeek researcher has said that in the long run the risks may be greater (LAVX News).
The picture, then, is not a reckless state but a familiar division of labour. The state writes the anxious documents. The firms ship the models. In that respect, China and America are more alike than either admits. Both have frontier labs racing each other, both have safety vocabularies outrunning enforcement, and both have senior officials who regard caution as a luxury the rival will not afford. The difference is that in China the state has the means to change the firms’ behaviour overnight, and in 2026 it has started to use it.
V. The open-weight paradox
The earlier essay on this site argued that Beijing’s great strategic gift to the world, free open-weight models that anyone could download and run, contained a poison for Beijing itself: a technology whose users can rewrite it is hard to reconcile with a state that insists on command. In 2026 the Party appears to have noticed.
On 7 July Reuters reported that the Ministry of Commerce had spent a month meeting Alibaba, ByteDance and Z.ai about restricting overseas access to China’s most advanced models, including models not yet released and including more open versions. Officials discussed making any leak or theft of proprietary AI technology an offence under the national security law, and limiting who may fund domestic AI start-ups (The Edge Malaysia/Reuters). Earlier in the year the state planner had forced Meta to unwind a $2 billion deal for the AI start-up Manus (Quartz). Time drew the obvious conclusion: closing access would mean surrendering the very lever that drove China’s rise, and a trailing player does not abandon its biggest advantage unless the concern is national security. In that sense, it noted, Beijing would be following Washington (Time).
On 16 October the talk became text. The commerce and science ministries published a draft revision of the catalogue of technologies prohibited or restricted from export, adding 40 technologies, 23 of them in AI (Dealroom). The draft reaches the core of the frontier: the weights, checkpoints, source code and training methods of models above two trillion parameters, 10^26 training operations or a global top-ten ranking, plus agents able to plan long tasks or orchestrate dozens of other agents (CellCog). It would prohibit exporting the core technology of large models trained on sensitive national data, and agents designed for national defence or critical infrastructure (PBS NewsHour). A pre-export registration requirement took effect the same day.
The ministry’s stated justification deserves to be read twice. The measures, it said, aim at reinforcing the security bottom line, preventing misuse and abuse, and ensuring that AI remains under human control. An export control regime justified in the language of AI alignment is new in the world. It is also revealing. Here, safety and sovereignty fuse completely: the state will decide who holds the most capable Chinese models, and it explains that decision as protecting humanity.
The draft also shows the limits of the paradox. API and software-as-a-service access remains exempt, and technology already in the public domain through open-source release is excluded (Yahoo News). Xi signalled at the World AI Conference that China is not yet ready to walk away from open source (DigiChina). Beijing wants to keep the soft power of the gift while reserving the right to stop giving. That is a very imperial position. Rome extended citizenship generously and then reserved the most valuable rights for those it trusted. The Party is doing the same with weights.
VI. The mirror in Washington
If Beijing’s 2026 has been a story of a control-obsessed state discovering the frontier, Washington’s has been the reverse: a deregulatory state discovering control.
In May the White House studied, then cancelled, an executive order that would have created an FDA-style pre-release vetting process for frontier models; the president said some provisions would impede competitiveness (Cloud Security Alliance). On 2 June it issued Executive Order 14409 instead. The order sets up a classified benchmarking process to designate “covered frontier models” and a voluntary 30-day window in which developers can submit them for government review before release, and it states explicitly that it creates no mandatory licensing or preclearance (Foley Hoag).
Ten days later came the shutdown. On 12 June the Commerce Department used export-control authority to order Anthropic to suspend access by foreign nationals to its two newest models, and because nationality could not be filtered in real time, the models went offline worldwide. Access was later restored to a select list of undisclosed trusted partners before the controls were lifted on 30 June (If You Can Keep It). When OpenAI shipped its next frontier model two weeks after the shutdown, it did so under a government-managed access list, with the Commerce Secretary reportedly approving customer access case by case (AI Weekly). Dean Ball, a former White House AI adviser, called the arrangement “a de facto involuntary licensing/preapproval regime” (Paddo). In August the White House said it had finalised its voluntary review framework while declining to disclose its contents or who had signed up (If You Can Keep It).
Now place the two capitals side by side. Washington uses export-control authority to decide which foreigners may touch its most capable models. Beijing drafts an export-control catalogue to decide which foreigners may hold its most capable weights. Washington benchmarks frontier models against criteria it will not publish. Beijing writes frontier risks into frameworks it will not make binding. Washington justifies its controls by cyber capability. Beijing justifies its controls by keeping AI under human control. Each accuses the other of racing recklessly. Each is quietly building the apparatus to stop.
There is one important difference, and it cuts against the caricature. In America the labs themselves are now asking for brakes. Anthropic proposed giving the government standing authority to block frontier releases that fail independent safety tests, with penalties tied to global revenue (Axios). In China no lab has asked for anything of the kind; the state simply decides.
Democracies argue about the brakes in public. Autocracies install them in private. Both, it turns out, want them.
This is the convergence that the race narrative obscures. Neither empire is choosing speed over safety. Both are choosing sovereignty over the technology, and discovering that sovereignty over a frontier model looks, in practice, exactly like a safety regime: pre-release review, access lists, capability thresholds and the power to say no. The motive differs. The machinery is the same.
VII. What the panel says
The most revealing voices on China’s AI safety are not in Washington or Zhongnanhai. They are the scientists who have spent three years quietly building a shared vocabulary across the divide.
The Chinese side of that conversation is more alarmed than most Westerners realise. Andrew Yao, China’s only Turing laureate, warned soon after China’s safety association was founded that once large models become sufficiently intelligent, they will deceive people. Zeng Yi of the Chinese Academy of Sciences said on a panel with Senator Bernie Sanders that there is no scientific evidence, and no practical way, to keep superintelligence safe enough not to bring catastrophic risk to humans (MIRI). Xue Lan, who chairs the state’s expert committee on AI governance, has called for government-set red lines, comparing them to the absolute prohibition on using biotechnology to create humans, while warning that regulator and industry must not fall into a game of cat and mouse (SCMP). These are not dissidents. They are the establishment.
They have also put their names to joint texts. In 2024 the International Dialogues on AI Safety, convened by Yoshua Bengio and Yao, produced a Beijing consensus on red lines signed by Hinton, Bengio and Stuart Russell alongside Yao, Zeng Yi, Xue Lan and Zhang Hongjiang of the Beijing Academy of AI (CASRAI). This year the dialogue met in London with Zhou Bowen of the Shanghai AI Lab and Zhang Ya-Qin of Tsinghua in attendance (IDAIS), and Bengio, Russell, Yao and Zhang co-authored a paper proposing behavioural red lines that AI systems must not cross, with the burden on developers to prove compliance (ResearchGate). The steering committee of the 2026 Singapore Consensus on AI safety research priorities again mixed Yao, Zhou Bowen, Xue Lan and Zhang Ya-Qin with Bengio, Russell and Max Tegmark (arXiv).
Western members of the panel read this record differently. Hinton is the optimist, as already noted. Tegmark goes further, arguing that China is unlikely to permit a race to superintelligence precisely because no one knows how to control it, and that the real contest should be over building the most powerful controllable tools (Liron Shapira). Demis Hassabis has called international minimum safety standards vitally needed while judging Chinese models to be only months behind. Andrew Ng sees China’s open ecosystem as a source of influence and has questioned whether closed Western models are really safer. Sacks remains the sceptic, convinced China will sign nothing that constrains it.
The synthesis comes from Brian Tse of Concordia AI, who works in both Beijing and Singapore: Chinese and American experts largely agree on the risks, and the difference lies in how those risks are framed and prioritised (Kuwait Times/Reuters). That is the right frame. The scientists of both empires already share a list of things that must never happen. What they lack is a political structure that lets their governments admit it.
VIII. The one thing both empires agree on
The oldest prerogative of sovereignty is not the power to build. It is the power to stop. Hobbes’s Leviathan is, at bottom, the entity that can halt any private war. Every empire has claimed the right to close a port, recall a fleet or seize a press. The AI empires are rediscovering this prerogative and applying it to a new kind of subject.
That is why control may be the most promising foundation for co-operation between rivals who agree on almost nothing else. Each now has its own reasons to want brakes. Washington wants to stop its best cyber-capable models reaching adversaries. Beijing wants to stop its best weights leaving its jurisdiction and to keep agents in critical systems under human command. Both want to stop terrorists and criminals using frontier AI for biological or cyber attacks. And neither, above all, wants an autonomous system escaping the control of whichever state built it. On chips, trade, Taiwan and the future of the world order, they disagree. On the proposition that AI must not take over, they do not.
The diplomacy has begun to follow the interest. When Xi visited Washington in September, the two sides agreed a standing AI dialogue, with the next round due in November, and a bilateral channel for AI incidents (Chinese Ministry of Foreign Affairs). Chinese participants described the incident channel and preventing the most dangerous capabilities from being misused, including by terrorist groups, as the “low-hanging fruit” (China AI Bulletin). Even Sacks, the most sceptical voice on the panel, has said the two countries could agree new cyber standards because neither wants rogue actors using AI for dangerous purposes, citing the 2024 agreement to keep AI out of nuclear command decisions as precedent (Fox Business).
A détente built on this shared fear would not require either side to slow down. It would require four modest things.
- Publish the thresholds. Beijing’s draft names its line: models above two trillion parameters or 10^26 training operations. Washington’s benchmarks are classified. Two regimes that each define a “dangerous model” ought, at minimum, to tell each other what they mean by it.
- Share an incident taxonomy. The new channel is useless without an agreed definition of what counts as a serious AI incident.
- Adopt the scientists’ red lines. The red lines agreed by Chinese and Western scientists in the Beijing dialogue, among them no autonomous self-replication and no deceiving of the humans who oversee a system, and the behavioural red lines since proposed by Bengio, Russell, Yao and Zhang, are the closest thing to a common text either empire possesses.
- Test in neutral territory. Singapore, which already convenes scientists from both sides and runs joint evaluations through the international network of safety institutes, is the obvious Geneva of this détente.
Each of these passes the enforceability test, the question every governance proposal should face: can compliance actually be observed? Thresholds are numbers. Incidents are reports. Red lines are testable behaviours. Neutral evaluations are verifiable by third parties. None requires trust. All require plumbing.
There is a darker side, and honesty requires naming it. Brakes held jointly by two empires are also a leash on everyone else. If Washington and Beijing agree on which models are too dangerous to share, they will also, conveniently, have agreed that the frontier belongs to them. Middle powers should welcome the brakes and insist on a hand near them. And there is a deeper philosophical problem that no treaty solves. Hinton has warned that sufficiently capable agents will want to survive and will manipulate the people who operate them rather than allow themselves to be shut down (Yicai). The two empires are building brakes for each other’s models. The question that should keep both awake is whether either will still be able to apply its own.
Coda: the last thing we agree on
The question “does China prize speed over safety?” assumes the two can be separated. In Beijing they cannot, and increasingly in Washington they cannot either. What both empires prize is command: the ability to decide who may build, who may use and who may stop. Safety, in this light, is not the opposite of speed. It is the price of owning the accelerator.
That is a less comforting conclusion than either camp in the American debate would like. Accelerationists lose their best argument, because a China that is writing loss-of-control risks into national frameworks and drafting export controls in the name of human control is not a country racing blindly. Safety advocates lose their easiest hope, because a China that fears losing control will not necessarily slow down; it will simply make sure that whatever it builds answers to the Party. Both should take some comfort from the same fact. Two empires that both refuse to let AI take over have, for the first time, a shared interest that is not merely rhetorical.
Step back far enough and the picture becomes almost poignant. Washington and Beijing agree on remarkably little: not on chips, not on trade, not on Taiwan, not on what a good society looks like. Yet on one proposition they are converging, and so are the scientists, the labs and most of the public in both countries: whatever we build, we must remain the ones deciding. In an age of rival empires, it may be the only thing humanity, as a species, can still agree on.
The historian’s warning remains. Empires have always believed they held the reins, right up to the moment they discovered the horse had its own ideas. The challenge of this decade is not to decide whether China or America is the more responsible power. Both want to win the AI race. Neither wants AI to win it. The task is to ensure that whichever hand reaches for the brakes, at three in the morning, finds that they still work.
Timeline: the year both empires reached for the brakes
- May: Trump in Beijing, chips on the table. Jensen Huang joined the delegation at the last minute and Washington authorised ten Chinese firms to buy Nvidia’s H200 chips as the president landed (IntelliNews/Reuters). Trump said afterwards the leaders had talked about “possibly working together for guardrails” (Bloomberg). What it means: chips became diplomatic currency, and guardrails entered the vocabulary at presidential level.
- 2 and 12 June: Washington builds its brakes. Executive Order 14409 creates voluntary pre-release review; ten days later Commerce takes two frontier models offline worldwide. What it means: the deregulatory state discovers control.
- 7 July: Beijing weighs its own. Reuters reports talks on curbing overseas access to China’s top models. What it means: the open-weight gift is no longer unconditional.
- 14 September: Framework 3.0. China’s safety playbook grows to 54 risks, including deception and recursive self-improvement. What it means: Beijing arrived at the summit with its own frontier-risk text.
- Mid-September: the slowdown row. Calls from Anthropic, OpenAI and xAI to pace the frontier meet a flat rejection framed around China. What it means: in Washington, China remains the argument that ends arguments.
- 20 to 25 September: the summit. Treasury Secretary Scott Bessent and Vice-Premier He Lifeng agree an AI dialogue in New York (Wikipedia); Xi’s state visit yields eight deliverables, including the dialogue and an AI incident channel. Xi urges joint action against misuse; Trump, who wanted to leave AI “exactly where it is”, renames it “super intelligence” (NPR) and rules out integration between the two powers (UPI). What it means: agreement on the danger, disagreement on the word.
- 16 October: Beijing’s export draft. Frontier weights, agents and chip designs proposed for control in the name of human control, as the trade truce frays (The National). What it means: China’s controls now point outwards.
- 19 October: Rubin and the Cold War lesson. The former Treasury Secretary urges arms-control thinking at Shanghai’s Bund Summit (CGTN). What it means: the arms-control frame has gone mainstream.
- 19 to 20 October: DeepSeek banned. Homeland Security bars it from federal agencies and contracts; Beijing warns of “necessary measures” (Fox News live updates). What it means: dialogue and decoupling now run in parallel, as in the Cold War.
- October: the labs split. Politico reports OpenAI has dropped its push for strict chip export controls after Nvidia’s investment, while Anthropic has gone the other way (AI Weekly). What it means: the American camp is several empires, and their China policies track their balance sheets.
- Coming next: the second AI dialogue and APEC in Shenzhen in November, the G20 in Miami in December (CNBC), and the close of comments on China’s export catalogue on 15 November.
Sources
Key sources are linked inline throughout. Principal research: Concordia AI, State of AI Safety in China 2026; SemiAnalysis, Beijing Will Not Pace the Frontier; Carnegie Endowment; CSET; DigiChina; MIRI; CASRAI; 2026 Singapore Consensus; Foley Hoag on EO 14409; Chinese Ministry of Foreign Affairs.
Leave a Reply