Washington moved to close the last big workaround in its AI chip controls on Monday — while Beijing published the rulebook for the robot industry it wants to dominate. The squeeze and the build-out arrived on the same day.
Washington closes the remote-access loophole for AI chips
The US Commerce Department’s Bureau of Industry and Security issued an interim final rule on November 16 that puts remote access to advanced American computing power under export controls, closing the workaround Chinese AI firms have used to rent restricted chips from data centres abroad instead of importing them. The rule, titled “Addressing Remote Access to Advanced Semiconductors”, is the direct sequel to the forensic evidence published last week that six of China’s newest AI models carry the numerical fingerprint of Nvidia’s banned Blackwell chips.
The rule has two parts. First, a licence is now required before advanced chips or servers go to any data centre worldwide that provides more than 12 million aggregate total processing performance (TPP) of computing power to companies headquartered in China, Macau or other restricted destinations — a threshold TechCrunch estimates at roughly 750 Nvidia H100 chips. Data centres already above that threshold get 90 days to obtain a licence. Second, licences are required for advanced chip shipments to five countries the rule calls higher-than-normal diversion risks: Indonesia, Malaysia, Singapore, Thailand and Brazil. The rule also rescinds the Biden-era AI Diffusion Rule, which the current administration had declined to enforce. The timing is pointed: it landed four days after Nvidia co-founder Jensen Huang resigned as chief executive, and three days before US and Chinese leaders meet at the APEC summit in Shenzhen. It is scheduled for publication in the Federal Register on November 19, when it takes effect. China rejected the move on November 17, and Chinese analysts quoted by the South China Morning Post have warned Beijing could answer with its rare-earths card.
Why it matters: the rule turns a hardware ban into a compute ban. Renting Blackwell-class power from a Singapore or Johor data centre was the escape valve for labs that could not buy the chips; that valve now needs a US licence. Analysts expect the restrictions to push Chinese developers faster toward domestic hardware such as Huawei’s Ascend chips — still weaker for training — and toward the cloud capacity Baidu and others are now selling at home. (Reuters via sofx.com; Profile News; Kurums; Seoul Economic Daily)
Beijing approves 274 industry standards for robots, AI and chips
On the same day Washington tightened the screws, China’s Ministry of Industry and Information Technology published 274 new industry standards covering humanoid robots, embodied AI, semiconductors, cloud computing, batteries and telecommunications. The ministry issued Announcement No. 30 on November 10 and published it on November 16; some standards took effect immediately, others come into force in 2027 or 2028.
Six of the standards cover humanoid robots and embodied AI directly: rules for collecting, processing and deleting robot-training data; design and safety requirements for robot-training sites; a framework for grading a humanoid’s intelligence; and performance and safety requirements for electric-drive joints and dexterous hands. The package also sets standards for chip-design (EDA) tools including chiplet interfaces, for AI computing services, and for interchangeable battery-swapping systems in heavy electric trucks.
Why it matters: this is the state codifying the robot race it has been funding, from training data to dexterous hands. Standards decide whose robots can be sold, serviced and insured — and whose training data counts. (MediaNama)
Baidu’s AI cloud nearly doubles as its ad business shrinks
Baidu reported third-quarter results on November 16 that show the company’s pivot in full: revenue from its AI Cloud Infra business rose 94% year on year to RMB 8.2 billion, with GPU cloud revenue up 191% — while online marketing revenue fell 22% to RMB 11.9 billion as Baidu rebuilds search around AI. Total revenue slipped 1% to RMB 30.8 billion.
The company also unveiled ERNIE 5.5 Flash Preview this month, a 280-billion-parameter foundation model which Baidu says ranks second among open-source models on the Code Arena WebDev leaderboard — a company claim. At Baidu World this month it also introduced Token Exchange, a marketplace connecting AI model providers with customers. Apollo Go, its robotaxi service, operated in 30 cities as of October and logged about 34 million autonomous kilometres in the quarter.
Why it matters: Baidu is becoming a compute landlord as its search ads fade — and the GPU cloud surge is the domestic side of the squeeze in today’s lead story. (TNGlobal; TradingView)
Pony.ai and Toyota start fully driverless bZ4X robotaxis in Guangzhou
Toyota and Pony.ai began fully driverless commercial bZ4X robotaxi service in Guangzhou on November 16, moving their partnership from vehicle development into passenger operations. The jointly developed Gen-7 bZ4X Robotaxi, built for Level 4 autonomous driving, has joined Pony.ai’s fully driverless commercial fleet; passengers can hail it around the clock through the PonyPilot app and mini-programs on WeChat and Alipay.
The companies are testing the model and preparing deployments in other major Chinese cities, working toward a planned fleet of around 1,000 bZ4X Robotaxis. The first mass-produced bZ4X Robotaxi rolled off the line in February 2026; the release gave no initial vehicle count, fares or financial guidance for the Guangzhou operation.
Why it matters: this is the world’s biggest automaker putting its badge on a fully driverless service in China — and it lands a day after Pony.ai’s CFO told TechCrunch the industry has practised its non-Nvidia chip backup plan for years. (ad-hoc-news)
Huawei Cloud open-sources an Ascend-native world model for robots
Huawei Cloud released CloudRobo-WM1.2 on November 16, an embodied world model trained on its own Ascend chips that predicts how the physical world will change in response to a robot’s action and generates the action itself. The company says the model ranks first on the RoboTwin 2.0 and LiberoPro 2.0 robotics benchmarks — company-reported results, not yet independently verified — and that it was trained on about 27,500 hours of data.
Huawei Cloud says it will release the model weights, inference code and post-training code on November 30, and that it will not build robot bodies — only compute, models and a development platform. A first commercial application is already named: a new generation of intelligent sweeping robots developed with sanitation equipment maker Fulongma, on pre-sale now and due in Shanghai on December 17.
Why it matters: open-sourcing a robot brain trained on Ascend is Huawei’s answer to the compute squeeze — build the developer ecosystem around domestic chips, the way Nvidia did with GR00T. (TechNode; The Robotics Media)
At APEC in Shenzhen, Xiaomi and CATL pitch the physical-AI economy
The APEC CEO Summit opened in Shenzhen on November 17, and China’s tech bosses used it to pitch the same theme from different angles. Xiaomi chief executive Lei Jun said the company’s embodied AI robots have completed practical trials in its factories, working full eight-hour shifts and recharging themselves twice during the shift, and called embodied AI “a massive industrial opportunity” with potential surpassing the automotive industry.
CATL founder Robin Zeng, speaking on the same stage, said battery swapping — which can swap a battery in 99 seconds — is the infrastructure innovation that unlocks EVs across Asia-Pacific. Baidu chief executive Robin Li added that Apollo Go’s safety record is about ten times better than a human driver’s.
Why it matters: Shenzhen is hosting APEC this week, and its biggest companies are selling the city itself as the model — factories that run themselves, cars that swap batteries in 99 seconds, taxis with no driver. (Longbridge; HK Businesswire)
The big picture
Today’s six stories are one story told six ways: the compute squeeze is now a legal fact, and China’s answer is to standardise, localise and open-source its way around it. Washington closed the rental loophole, Beijing published 274 standards for the industries it wants to lead, Baidu sold the domestic compute, Huawei open-sourced a robot brain trained on home-grown chips, and Toyota and Xiaomi put driverless cars and factory robots into real service.
Watch next
Nvidia reports quarterly results after the US market close today, November 17 — interim chief executive Colette Kress’s first public outing in the role, and the first hard numbers since Jensen Huang’s resignation. The Guangzhou Auto Show opens on November 27, with BYD’s next-generation Han and the Yangwang Altior’s China debut expected on media day.
Sources
- Reuters (via sofx.com), November 17, 2026 — US mandates licenses for AI chips shipped to foreign data centers serving China
- Profile News, November 17, 2026 — US AI chip curbs target China’s remote access
- Kurums, November 17, 2026 — BIS remote-access AI chip rule: who needs a licence
- Seoul Economic Daily, November 17, 2026 — US moves to block China’s access to AI chips via third countries
- MediaNama, November 17, 2026 — China approves 274 industry standards covering humanoid robots, AI, semiconductors and data security
- TNGlobal (Baidu PR), November 16, 2026 — Baidu announces third quarter 2026 results
- TradingView, November 16, 2026 — Baidu Q3 earnings call highlights
- ad-hoc-news, November 17, 2026 — Toyota Motor stock: Pony.ai launches driverless bZ4X service
- TechNode, November 16, 2026 — Huawei Cloud plans to open-source its Ascend-based world model for robots
- The Robotics Media, November 16–17, 2026 — Huawei open-sources CloudRobo-WM1.2 embodied world model
- Longbridge, November 17, 2026 — Lei Jun: embodied AI robots represent a massive industrial opportunity
- HK Businesswire, November 17, 2026 — Tech leaders say EV innovation, robotaxis key for APEC
