Topics / AI Policy & Regulation
AI Policy & Regulation
Last updated
Two governments are drawing lines around AI at once, and in 2026 the lines moved from chips toward almost everything else. Washington extended its export controls from hardware toward model access and whole product categories, ordering a top US lab to cut its best models off from foreign nationals and barring Chinese robots from the market before they arrived; Beijing rewired procurement toward domestic silicon, blocked the offshore sale of Chinese-built AI, banned Nvidia purchases at home, and began writing the safety standards its own market will run on. The policy layer often moves the industry more than any single model, so we track both capitals with dates and sources.
The state of play
- The clearest 2026 shift is that US controls moved past chips to the models themselves. On 12 June 2026 the US ordered Anthropic to suspend its two highest-capability models for every foreign national in the world, including those physically inside the United States, the first time a model rather than a chip was treated as an export-controlled item; Anthropic disputed the severity of the cited vulnerability but complied. The immediate effect ran the other way, pushing the Chinese developer community, which leans heavily on Claude for coding, toward open-weight domestic alternatives such as Moonshot’s Kimi K2.6, DeepSeek V4-Pro and Zhipu’s GLM-5.2 (Anthropic newsroom statement; V2EX, via CAD, “The Cutoff”).
- Washington then fenced out an entire hardware category before it arrived. On 28 July 2026 the FCC added “foreign-produced advanced robotic devices” to its Covered List, so new humanoid and four-legged robots can no longer get the authorization they need to enter the US market, effective immediately, with Chinese-made power inverters swept in on the same national-security argument. Unlike the DJI drone controls, which came after Chinese drones already owned the market, this gate came down before the product did: China builds more humanoid robots than any country but almost none are inside US factories yet. Many non-Chinese makers are expected to win Pentagon exemptions, and in the December 2025 drone precedent none of the four exempted foreign devices was Chinese (BBC; Reuters; CNN, via CAD, “The Preemption”).
- Beijing has built its own gate on the way out. The National Development and Reform Commission blocked Meta’s roughly 2-billion-dollar acquisition of Manus, a Singapore-incorporated but Chinese-built AI-agent company, and ordered it unwound; state broadcaster CCTV named the structure 洗澡式出海 (“Singapore-washing”), and Beijing’s stated test is where the technology, people and data came from, not the legal address. It is a functional equivalent of the United States’ CFIUS, pointed at Chinese-origin technology trying to leave the country (NDRC; CCTV; Wall Street Journal, via CAD, “The Singapore Clause”).
- China is now turning its chip and model rules outward too. Its internet regulator formally banned major domestic tech companies from buying Nvidia AI chips after summoning Huawei, Cambricon, Alibaba and Baidu and concluding domestic accelerators had matched or passed the Nvidia parts still allowed under US controls; and the export controls on its own AI models and chip technology that the Financial Times reported Beijing was weighing arrived on 16 October 2026 as a draft export list from the commerce and science ministries. Its restricted tier covers the weights and source code of models trained with at least 10^26 floating-point operations or with at least 2 trillion parameters, and AI chip designs meeting any one of six conditions, with a filing to the commerce ministry required immediately before any such export; open-sourced weights and API, app and SaaS access by foreign users are outside it, so what it reaches is who holds the weights. Read next to the US moves, both sides are raising walls at once, and the shared global AI stack that everyone built on starts to come apart (Financial Times, via CAD, “The Exit” and “The Preemption”; Ministry of Commerce Announcement No. 48 and spokesperson Q&A, via CAD, “Sending a Chinese Model’s Weights Abroad Now Needs a Commerce Ministry Filing”).
- Beijing also has an industrial-policy instrument that has no Western equivalent, and it is not a subsidy. On 3 June 2026 MIIT and SASAC jointly issued 工信厅联科函〔2026〕256号, which does not pay for humanoid robots but instructs the organisations the state owns to open their work stations to them: at least ten real scenes per designated central enterprise, at least twenty per named province, targeting more than a hundred high-value scenarios and ten-thousand-unit deployment capability by end-2026, with results due 30 November. Subsidies and cheap credit are the tools most coverage tracks; this one conscripts workplaces, an asset the Chinese state owns at a scale no Western government does, and the output it is buying is the real-world operating data the field is short of. Whether that produces working deployments or paperwork is the open question, and China’s own 2021 autonomous-driving buildout is the cautionary precedent rather than the encouraging one (MIIT and SASAC notice, read in full; via CAD, “The Internship”).
- At home, Beijing is writing the rulebook the market will run on. The Ministry of Industry and Information Technology issued GB 44721-2026, China’s first compulsory safety standard for L3 and L4 automated driving, upgrading a 2024 recommended standard to a required one that takes effect in July 2027 and covers both passenger and goods vehicles, and the same window brought a first national standard system for humanoid robots and embodied intelligence. On the consumer side the two-year AI price war ended as all three major clouds raised prices in one quarter, a sign the market is maturing from subsidy toward industrial deployment, where IDC data cited by the NDRC puts industrial-enterprise AI penetration up from about 9.6 to 47.5 percent in a year (ithome, via CAD, “The Cheap End”; CAD, “The Preemption”; IDC via NDRC, via CAD, “The Price War Is Over”).
Best issues on this topic
- China Will Screen Some AI Projects for the Jobs They Could Cut
A State Council General Office opinion on AI and employment, document No. 29 of 2026, signed on 16 October and posted on gov.cn on 23 October, says AI application projects in fields subject to market access that may cause a relatively large employment shock are to get an employment risk assessment alongside their review. Other AI projects are assessed according to actual circumstances. Most of its eighteen points are about creating jobs, through new occupational standards, training, start-up support and job-stabilising loans. Point 6 asks authorities to handle the pace of AI in labour-intensive industries steadily and cautiously and encourages firms to explore an unemployment buffer period, and point 17 asks for a special survey of AI's employment effects. The opinion names no sector and contains no number, deadline or budget. The June employment plan for the 15th Five-Year period had asked for job-impact assessments of major projects and an AI employment risk early-warning system; the October opinion turns the second into a review step.
- Sending a Chinese Model's Weights Abroad Now Needs a Commerce Ministry Filing
Commerce Ministry Announcement No. 48 makes a filing mandatory, effective immediately, before a Chinese company licenses restricted model weights or an AI chip design abroad. The draft export catalogue, open for comment until 15 November, restricts weights of models trained with at least 10^26 FLOPs or with at least 2 trillion parameters. The ministry's Q&A puts API, app, web and SaaS access outside its scope, so the rule governs who holds the weights, not who calls the model.
- A Beijing Court Reinstated the Compliance Officer Her Employer Fired for Using AI
The Beijing Chaoyang District People's Court found an insurer's dismissal of a compliance officer unlawful after she uploaded an internal audit table to a chatbot her own manager had recommended. The insurer had confidentiality rules but none that prohibited third-party AI tools, and it could not show the file was core confidential information or that anything leaked. The court ordered her contract resumed and her lost pay made up.
- Qualcomm Is Buying Huawei Patents in Compute, AI and Networking
Qualcomm's press release of 5 October 2026 announces a multi-year, broad cross-licence with Huawei spanning 5G, compute, AI and networking, and in the same clause Qualcomm's purchase of certain Huawei US patents in compute, AI, networking and other technologies. A purchase on those ones, not a licence, closing after regulatory approvals, with FRAND principles invoked. Neither company disclosed a price, a term length, or which way a balancing payment runs, and the two quotes point in opposite directions: Qualcomm's John Han reads the agreement as industry recognition of Qualcomm's 5G technology leadership and of its 5G SEP licensing programme, while Huawei's Alan Fan reads it as demonstrating the value of Huawei's innovations and name-checks polar codes. Both can hold in a cross-licence with a net payment, and neither side will say which direction the balance points. The only figure came from Huawei to the Chinese press rather than from the joint release: ithome reports Huawei saying that once the transaction closes the cumulative total of all its patent licence agreements is expected to exceed 6.9 billion US dollars, a 累计总金额 that covers contracted value and cash received alike without separating them, and that its IP licensing business has run net revenue-positive since 2021. The dates Huawei volunteered are firmer than the total: it first paid Qualcomm a patent licence fee in 2001 and took its first licensing income, from Motorola, in 2011. Bloomberg, relayed by ithome, headlined the deal as Qualcomm licensing patents on Huawei's LogicFolding chip technology, and ithome flags on its own initiative that the English phrase is directionally ambiguous where the Chinese is not, reading it in context as Qualcomm obtaining the licence rather than granting one. LogicFolding sits inside 韬定律, the scaling law Huawei semiconductor president 何庭波 proposed in May 2026, which gives up on geometric shrink and goes after the time constant instead; Huawei says 381 chips have been designed and mass-produced against it. The structural point is that a patent licence does not clear customs. The rules that stop ASML selling Huawei a lithography scanner do not stop Qualcomm buying a Huawei patent, subject to the approvals the release names, so six years of working around an equipment blockade now has a buyer in San Diego. Qualcomm shares were up as much as 4 percent premarket.
- Beijing Will Lend Against Token Counts Its Own Official Says Are Idling
Beijing issued the Action Plan for Accelerating the Development of the Token Economy (2026-2028) jointly through the Municipal Bureau of Economy and Information Technology and the Development and Reform Commission, dated 15 September, published on the 18th and running to the end of 2028. It descends from State Council Document Guo Fa [2025] No. 11, the "AI Plus" directive, and is the first policy at provincial level in China written around the token as an economic object rather than a technical one. Ten numbered measures follow. Beijing will write a graded evaluation standard for what it calls token factories, a compute centre bolted to a model-tuning operation, grading them on models adapted, token throughput, time to first token, cache hit rate and power usage effectiveness, and back the leading ones with land, energy quotas and network guarantees. It will build a token distribution platform with a unified interface, model aggregation and token metering and settlement, then fold that platform into the city new-infrastructure plan. It will have financial institutions develop products for token consumption scenarios in order to extend medium and long-term credit. And it will retrain working software engineers as token engineers. At the press briefing the next day, Chen Guangming, deputy director of the bureau digital industry division, said standards and evaluation systems for token factories are not yet unified and that the problem of token idling still exists. The Chinese term is 空转, the word for an engine running in neutral, and it carries the weight of 资金空转, capital idling inside the banking system without reaching a factory, which regulators spent most of the last decade on. His colleague Zhang Jinrui said Beijing will schedule compute across Beijing, Tianjin, Hebei and Inner Mongolia and build a chain running from electricity through compute invocation to token production, distribution and application, which the bureau expects to pass a trillion yuan during the current five-year plan. The lending is already happening: Beijing E-Town arranged close to 2 billion yuan of credit lines partly on the basis of token consumption, with the Agricultural Bank of China, CITIC, Industrial Bank, Bank of Beijing, Minsheng and Huaxia participating, and Bank of China Guangzhou ran the first token loan in August. Lou Feipeng of Postal Savings Bank named exaggerated token-use data as the first obstacle to scaling it. Leiphone reported from the cloud sales floors at the big three platforms, where model-as-a-service quotas run above 10 million yuan per salesperson against a few hundred thousand to a million last year, one regional team resigned together in an episode classified internally as a management accident, and vendors resell Zhipu models at 60 percent of list and hand over cloud-service vouchers when the discount runs out of room, so the customer cost falls while the model revenue line does not move. At Volcano Engine token call volume is a single-veto item. The token count Beijing wants banks to lend against is, at three of the largest producers of it, a quota with a subsidy attached. In the Briefing: Huawei set commercial dates for the Ascend 950 cluster cloud service, 30 September in China and 30 November globally; the Ascend 960 SuperPoD takes 4,096 NPUs on near-packaged optics; CXMT posted an 82 percent operating margin for April to June, the highest of any memory maker in the world; and the securities regulator gave banks informal window guidance raising the bar for humanoid robot listings.
- App Registrations Doubled While Model Filings Went Flat in China
The Cyberspace Administration of China published its July to August 2026 generative-AI bulletin on 14 September: 124 new services completed national filing (备案), seven of them running on phones rather than in a data centre, while 133 applications and functions completed the lighter provincial registration (登记). The registration column has never been the larger of the two before. Cumulative totals to 31 August are 1,112 services filed and 731 applications registered. The distinction between the two registers carries the story. 备案 is what a company does when it trains or operates a generative model and offers it to the public, so that column counts model supply; 登记 is what a company does when it calls somebody else’s already-filed model through an API and wraps a product around it, so that column counts what is built on top. Read off every bulletin CAC has published, the pairs run 238 and 105 across 2024, 93 and 77 in April to June 2025, 99 and 27 in July to August 2025, 73 and 35 in September to October 2025, 48 and 46 in January to February 2026, 72 and 49 in March to April, 120 and 68 in May to June, then 124 and 133. The model column added four over the summer window; the application column roughly doubled from 68. A deflationary reading is available and cannot be ruled out from the bulletin, because 登记 is a cheap provincial procedure and a backlog clearing at a couple of dozen local offices would produce the same shape, and CAC publishes counts rather than queue times. The bulletin also carries an obligation rather than a count: a live application or function has to display which filed service it uses, naming the model and quoting its filing or online number. One link does not reconcile, and it is worth saying rather than printing a clean series: the filing column chains exactly from 302 through to 1,112, while the registration column shows 263 as of August 2025, a next bulletin adding 35, and a printed cumulative of 306 rather than 298.
- Spain and Croatia Opened Public Roads to Chinese Robotaxis
Spain’s Dirección General de Tráfico issued the country’s first operating licence for a Level 4 autonomous passenger vehicle on 10 September, and it went jointly to three companies: WeRide, Uber, and AVOMO, the self-driving unit of the Spanish rental group Moove Cars, with the Madrid city government supporting the application. The permit was granted under Spain’s ES-AV framework and clears the three of them to put vehicles on public roads, with commercial service to the public expected before the end of 2026 once the operating and regulatory conditions are met. Madrid is the fourth city in a WeRide and Uber plan the two say will reach 15 cities and tens of thousands of vehicles before 2030, and WeRide says it is now the only company whose products hold autonomous-driving permits in nine countries, naming China, Spain, Switzerland, Belgium, France, Singapore, the United Arab Emirates, Saudi Arabia and the United States. The same morning Pony.ai and the mobility operator Verne said they had begun fully driverless robotaxi passenger testing on public roads in Zagreb, which they call the first anywhere in Europe. Set against both, Anfia chairman Roberto Vavassori asked the EU for an 80 percent tariff on Chinese vehicles and parts above a quota of 8 percent of annual new-car registrations, per Reuters, on top of the combined 18 to 45 percent the 2024 measure already charges Chinese electric vehicles. The proposal covers vehicles and parts and says nothing about software, fleet operations or a permit: a tariff is charged on an object crossing a border, and an operating licence is granted for a service performed inside one. In both deals that day the Chinese company supplied the driving stack while a European operator supplied the cars, the depot, the insurance and the relationship with the regulator, which is a template that travels without a vehicle being imported. Nine permits is nine permissions and no commercial service, Madrid is an expectation for later this year, and Zagreb is a test with passengers in it.
- China’s Compute Grew 177 Percent. Its 2030 Plan Asks for 40.
MIIT issued its five-year plan for the information and communications industry on 7 September, thirteen headline indicators across twenty-six tasks and six areas, and set intelligent computing capacity at 9,800 EFLOPS by the end of 2030. The base it grows from is 2,185 EFLOPS at the end of June 2026, measured at FP16 and up 177 percent year on year, so moving to 9,800 over four and a half years works out at roughly 40 percent a year compounded against a trailing twelve months that ran at 177. The plan asks for the orderly deployment of ten-thousand-card and hundred-thousand-card clusters, and MIIT put occupancy across national computing facilities at 71.4 percent at the same July briefing, so close to three in ten of the capacity already built is not in service. Three of the thirteen indicators are efficiency or carbon measures, including the power usage effectiveness of newly built large and hyperscale facilities. The money side puts cumulative information infrastructure investment at 3.8 trillion yuan across 2026 to 2030 and industry revenue at 4.1 trillion by 2030; at the onshore rate of 6.7105 yuan to the dollar on 8 September that investment figure is about 566 billion dollars, where English coverage rendered it as 532 billion, implying a rate near 7.14 that nothing traded at in that week. Chinese banks spent August building collateral for it: Agricultural Bank put a Token loan scheme into Shanghai and Zhejiang, Bank of Chengdu ran a pure credit product capped at 5 million yuan per borrower and backed by compute vouchers, and Zhangjiagang Rural Commercial Bank set aside a 2 billion yuan facility, underwritten on token consumption, compute service contract value and receivables from compute work.
- Doubao, Qwen and ERNIE Now Answer for Their Training Corpus
The Cyberspace Administration published a phase-two note on its Qinglang campaign against chaos in AI applications on 2 September, and it places obligations at three points in the pipeline rather than one. Platforms including Douyin, Weibo, Tencent and Xiaohongshu are named for expanding face, voiceprint and violation sample libraries, with 46 governance announcements published between them. Four model products, Doubao, Yuanbao, Qwen and ERNIE, are told to guard the content review pass on original training corpora and to label AI-generated synthetic content so as to prevent illegal information at source, which is an obligation on the pre-training set rather than on a feed. The third layer is distribution: Huawei, Xiaomi, OPPO and vivo are named as setting developer admission standards for AI applications and spot re-testing apps already listed, so four handset makers become the admission gate. The note landed one year and one day after the labelling Measures took effect on 1 September 2025. No published standard, audit or methodology sits behind the corpus clause, so what changed this week is scope rather than enforcement capability.
- China’s L3 Rules Turn Mandatory on July 1, 2027
China’s L3 autonomous-driving standard becomes a condition of sale. GB 44721-2026, issued by the State Administration for Market Regulation and the Standardization Administration on 30 July 2026 and effective 1 July 2027, drops the T from the recommended GB/T 44721-2024 and so turns mandatory, enforced through type approval: after that date a passenger or goods vehicle whose L3 or L4 system falls short of it does not get sold in China. The L3 clauses reach past the vehicle to the person in it, requiring seatbelt monitoring, occupancy monitoring and monitoring of the driver’s capacity to take the wheel back, which makes interior sensing a compliance item rather than a feature. Huawei’s Luxeed brand said the RX has already entered L3 type-approval testing, ten months before the rule binds, and Horizon Robotics put its share of basic ADAS on domestic-brand passenger cars at 50.02 percent in the first half, its first time above half, on 2.055 billion yuan of customer contract revenue.
- The Register
The Central Cyberspace Affairs Commission published a 2026 to 2030 action plan for cyber-information enterprises on 21 August: seven actions and twenty-one measures. Measure 1 orders a continuously updated national database of these firms, with classification and grading standards sorting them into three named tiers. Measure 19 tells regulators to speed up domestic listing review and to open non-IPO exits through S-funds and M&A funds, and tells banks to treat patents, trademarks and data as credit-assessment inputs. Measure 16 orders administrative inspections regulated so as to minimise disruption to normal operations, and Measure 5 names high-end AI chips, high-performance training clusters, world models, agents and embodied intelligence.
- The Internship
Forty-eight central state-owned enterprises appeared together at the World Robot Conference in Beijing on 19 August, showing twelve application scenarios under a joint MIIT and SASAC notice that orders each central enterprise to open at least ten real work stations and each designated province at least twenty, targeting more than a hundred high-value scenarios by the end of 2026. The next morning on the same stage Unitree founder Wang Xingxing said robots are not yet efficient or general enough to be rolled out into factories at scale, and put embodied AI’s ChatGPT moment two to three years away at best. His stock fell 16.05 percent that day.
- The Preemption
On 28 July 2026 the FCC added foreign-produced advanced robotic devices to its Covered List, barring new Chinese humanoid and four-legged robots from the US market before they had arrived, with power inverters swept in on the same national-security argument. Beijing called it protectionism and, per the Financial Times, weighed export controls on its own AI models, training data and chip technology.
- The Cutoff
US Commerce ordered a leading AI lab to cut its two most capable models off from all foreign nationals worldwide, then reversed it a week later. Export controls reached model access, and retreated.
- The Singapore Clause
China’s NDRC blocked Meta’s roughly 2 billion dollar acquisition of an agent startup that had reincorporated in Singapore, ordering the deal unwound. State media called it "Singapore-washing."
- The Exit
China’s Cyberspace Administration barred major domestic firms from buying Nvidia AI chips, citing an assessment that domestic parts now match them. The control now cuts both ways.
- The Supplier
An interagency committee cleared CXMT for the US Entity List, but Commerce has not updated it since October 2025, the longest pause in a decade, because the two sides are mid-negotiation.
- The Cheap End
China’s delivery robots crossed the cost line: Neolix says its per-vehicle cost fell more than half in two years to below 10,000 dollars, moving a parcel 25 to 35 percent cheaper than a human courier. The logistics industry projects 860,000 units a year by 2030, and the startup AIROBO closed the field’s largest first round the same week a Shenzhen circuit-board and chip-design platform, JLC, opened more than 170 percent above its offer price.
- The Price War Is Over
The US House Select Committee documented the workarounds, legal procurement, smuggling and distillation, and noted China remains the largest market for chipmaking equipment despite the controls.
These link to the full issues on the newsletter. New pieces on this topic go out in the daily first.
Common questions
How is the United States restricting Chinese AI in 2026?
On more layers than chips. It kept the hardware controls that bar advanced accelerators, then extended them into software and finished products: in June 2026 it ordered Anthropic to cut its two top models off from all foreign nationals, treating a model itself as an export-controlled item, and in July the FCC barred foreign-made advanced robots and power inverters from the US market before they had arrived. The direction of travel is from chips toward model access and whole product categories.
What is the FCC Covered List robot ban?
On 28 July 2026 the FCC added foreign-produced advanced robotic devices to its Covered List, the register of products deemed a national-security risk, so new humanoid and four-legged robots cannot get the authorization they need to be sold in the United States. It is preemptive: China builds more humanoid robots than any country but almost none are in US factories yet, so unlike the earlier DJI drone case the restriction arrived before the product did. Many non-Chinese robot makers are expected to win Pentagon exemptions, which is how a rule that reads as global can be closed to one supplier and left open for the rest.
Does China control the export of its own AI?
Increasingly, yes. Beijing blocked Meta’s acquisition of Manus, a Singapore-incorporated but Chinese-built AI company, and named the “Singapore-washing” exit it now prohibits, giving it a CFIUS-like gate on Chinese-origin technology. It also banned domestic tech giants from buying Nvidia chips, and on 16 October 2026 its commerce and science ministries published a draft export list that restricts the weights of China’s largest models and most AI chip designs, with a filing required immediately before any such export. Open-sourced weights and API access by foreign users are outside the list, so it governs who holds a model’s weights rather than who may use it.
How does China regulate AI at home?
Through procurement, deal review and safety standards rather than one omnibus AI law. It steers state and industry buying toward domestic chips and models, reviews and blocks sensitive foreign deals, and is writing the technical rulebook the market runs on, from GB 44721-2026, its first compulsory L3 and L4 self-driving safety standard, effective July 2027, to a new national standard system for humanoid robots. On the consumer side the multi-year AI price war ended in 2026 as the major clouds raised prices, a shift from subsidy toward industrial deployment.