|
|
|
|
Bottom Line Up Front Moore Threads reported training a frontier model on a 10,000-chip cluster of its own GPUs, saying its 236-billion-parameter mixture-of-experts system used more than 25 trillion tokens, despite the company's placement on the U.S. Entity List. Two Chinese AI systems scored perfect IMO golds, with Xiaohongshu's dots-note-3.0 and Huawei's Xiaoyi each earning a 42-out-of-42 at the 2026 International Mathematical Olympiad, the first Chinese models to reach that mark. China's foreign ministry announced a UN position paper on global cyber governance, filed with the UN Permanent Mechanism on Global Cybersecurity, that casts "digital sovereignty" as the successor to "cyber sovereignty" and asserts each state's right to choose its own AI development path without being coerced into taking sides. |
| Sanctioned GPU maker Moore Threads says it trained a frontier AI model on 10,000 domestic chips Moore Threads is a Chinese GPU designer on the U.S. Entity List, the Commerce Department blacklist that bars American suppliers from selling to it without a license. At WAIC 2026 in Shanghai, the company said it trained a 236-billion-parameter mixture-of-experts model on a 10,000-chip cluster built entirely from its own GPUs, using more than 25 trillion tokens of training data. A mixture-of-experts model routes each input token to a small subset of its parameters, so the 236-billion-parameter total runs like a smaller model at inference. Moore Threads says the cluster kept effective training time above 90 percent, the wall-clock share spent training rather than recovering from failed chips. |
| Two Chinese AI systems score perfect 42-out-of-42 gold at IMO 2026 Xiaohongshu's dots-note-3.0 reasoning model and Huawei's Xiaoyi Agent each scored 42 out of 42 at the 2026 International Mathematical Olympiad (IMO), the annual high school math competition. Both results were announced July 22 and were certified as gold medal performances by IMO's own graders, who apply the same rubric to models as to human contestants. Xiaohongshu is a Chinese lifestyle and shopping social platform; Xiaoyi is Huawei's on-device AI assistant, and Huawei sits on the U.S. Entity List. Xiaohongshu says dots-note-3.0 is the first Chinese large language model to receive IMO official gold certification, and the first globally to receive a perfect score. IMO president Gregor Dolinar separately sent Huawei written congratulations confirming its 42-point perfect score. |
| China's foreign ministry announces UN position paper on 'digital sovereignty' and self-directed AI paths China's Ministry of Foreign Affairs said on July 22 that its delegation had filed a formal position paper on global cyber governance with the UN Permanent Mechanism on Global Cybersecurity. The mechanism's first session runs July 20-24 at UN headquarters in New York. The paper's central proposition is that each state has the right to choose its own development path for AI and other digital technologies, and should not be coerced into taking sides. It introduces "digital sovereignty" as the successor to "cyber sovereignty," extending the earlier Chinese principle from regulating the internet inside a country's borders to the AI systems, products and services built on top. |
| Chinese electronics analyst reframes chip independence progress as three system-level tests Guolian Minsheng Securities electronics analyst Xue Hongwei said in a widely circulated note that Beijing's chip independence progress should be judged by three system-level tests, not single-chip parity with U.S. GPUs. Xue's three "arrows" are whether Chinese AI chips ship in volume, whether advanced fabs can keep up with demand and whether multi-chip super-node systems turn many cards into usable compute. A super node is a rack-scale chassis that bundles many accelerators with tight interconnects so software treats them as one large compute pool. On volume, Xue said 2025 Chinese AI accelerator card shipments reached roughly 4 million units, of which about 1.65 million came from domestic vendors, putting domestic share above 40 percent. |
| Chinese government AI agent platform launches at WAIC with Alipay, CAICT and state telecom units The Internet Society of China, Alipay, the China Academy of Information and Communications Technology (CAICT), and subsidiaries of the state-owned telecom carriers launched a Trusted Government AI Agent Public Service Platform at WAIC 2026 in Shanghai, with seven institutions and companies taking part. CAICT is the Ministry of Industry and Information Technology's in-house think tank, which sets many of the technical benchmarks Chinese regulators cite. A government AI agent answers citizen queries, walks users through administrative procedures, and executes tasks such as pulling records or filing forms. Alipay separately disclosed that its own government AI assistant "Xiaozheng," launched in September 2025, has passed 100 million cumulative service interactions. |
|
| Chengdu sets 2027 target of 260 billion yuan for core AI industry under new city action plan The Chengdu Municipal People's Government General Office issued an "AI+" Action Plan on July 22, targeting a core AI industry above 260 billion yuan (about $36 billion) by 2027, with annual growth above 30 percent. Chengdu is a large inland provincial capital in Sichuan; the plan places it in the same tier of city AI industrial policy as Shanghai's July 17 subsidy scheme. Core AI industry means revenue from firms whose primary business is AI chips, models, software or dedicated hardware, as opposed to broader AI enabled sectors. The plan also includes a high-performance compute buildout and a connected autonomous vehicle mass production push. |
| MOFCOM protests EU's 550-million-euro AliExpress fine as 'digital barrier' targeting Chinese platforms China's Ministry of Commerce (MOFCOM) said on July 22 that it registered "strong dissatisfaction and grave concern" over the European Commission's 550-million-euro (about $629 million) fine on AliExpress, Alibaba's cross-border shopping platform, and firmly opposed the decision. MOFCOM said the EU was using platform regulation as a pretext to erect "digital barriers" and enact discriminatory measures against Chinese e-commerce companies. The European Commission imposed the fine on July 20 under the EU Digital Services Act, for AliExpress's failure to curb sales of illegal and counterfeit products. The Commission set an October 20 deadline for AliExpress to file remedial measures, and AliExpress separately called the fine disproportionate. |
| China's finance ministry raises 2026 central basic research budget 16.3 percent under tight fiscal stance China's Ministry of Finance said on July 22 that its 2026 central government budget for basic research grew 16.3 percent from 2025, an increase under overall belt-tightening. Within the 426.4-billion-yuan (about $62 billion) central science and technology envelope for 2026, basic research is set at about 116.9 billion yuan (about $16 billion). Applied research is set at 156.9 billion yuan (about $22 billion), down 13.3 percent. |
|
| III | CHIPS & SEMICONDUCTORS | |
| Arm China unveils new NPU architecture and open source edge AI operating system alliance at WAIC Arm China is the joint venture that licenses UK-founded Arm's chip designs into China. It has operated on an unusually independent footing since a 2020 governance dispute that Beijing did not resolve in Arm Ltd.'s favor. At WAIC 2026 in Shanghai, Arm China unveiled its next-generation neural processing unit (NPU) architecture, called Zhouyi X3-Pro, and led the launch of an open source AI operating system alliance. An NPU is the dedicated on-chip accelerator that handles AI model inference on phones, cameras and other edge devices. The Zhouyi line is Arm China's own NPU design, sold alongside its Arm-derived CPU cores. Founding chip members of the alliance include Unisoc, Rockchip and Huixi, with ModelBest joining on the model developer side. |
|
| IV | ROBOTICS & AUTONOMOUS SYSTEMS | |
| Momenta wins Shenzhen robotaxi road-testing permit as SAIC readies mass production model for 2027 Chinese autonomous driving developer Momenta said recently that it received a Shenzhen intelligent connected vehicle road test permit for its Robotaxi, an L4 self-driving vehicle that operates within a defined area without a human safety driver. Shenzhen becomes the fourth Chinese city where Momenta runs Robotaxi tests, after Shanghai, Suzhou and Wuxi, alongside overseas testing in Munich and Abu Dhabi. The city's revised intelligent connected vehicle rules took effect in July, lowering application thresholds and supporting fully driverless testing. At WAIC 2026, SAIC's Xiangdao Mobility ride-hailing arm launched a custom mass production Robotaxi project with Momenta, with the vehicle planned to debut in 2027. SAIC Motor will build the car, Momenta will provide the driving stack, and Xiangdao will operate the fleet. |
| Chinese and international embodied AI leaders peg physical AI 'ChatGPT moment' at two to five years Six leaders of the Chinese and international embodied AI industry sat on a WAIC 2026 panel and gave two-to-five-year estimates for when robotics reaches a "ChatGPT moment." Embodied AI pairs an AI system with a physical body such as arms, grippers and sensors, so it acts on the world rather than only reasoning about it. AgiBot's Yao Maoqing defined the milestone as a general-purpose robot working out of the box on common tasks with 70 to 80 percent basic success. Yao gave the two-year lower bound, and Georgia Tech's Xu Danfei gave the five-year upper bound. Panelists from Physical Intelligence, Tencent's Robotics X Lab and others placed themselves in between. |
|
| V | QUANTUM & FRONTIER SCIENCE | |
| Chinese researchers show LLM agents can help bypass DNA synthesis screening in wet lab test A team from the Beijing Academy of Artificial Intelligence (BAAI), a state-adjacent AI research institute in Beijing, and Peking University reported that widely available large language model agents can help non-experts bypass automated DNA synthesis screening. Gene synthesis vendors use that screening to catch customer orders matching dangerous pathogens or toxins on export control lists. The researchers tested a fragmentation attack, where a hazardous sequence is broken into short fragments that individually look benign and only reassemble in the lab. Across 11 commercial models, all 11 produced fragmentation plans that passed computational screening against the IBBIS Common Mechanism baseline. Four benign proxy constructs were then assembled from model-generated instructions in the wet lab. |
|
|
|