|
|
|
|
Bottom Line Up Front Pangoal Institution mapped out which metals give Beijing the most leverage over global AI chip production, ranking rare earths and gallium as the most direct chokepoint and tungsten, germanium and indium as secondary levers that touch memory and optical links. Taiwan's Investigation Bureau searched mainland-backed firms belonging to 17 companies, targeting firms concentrate in chip design, semiconductor equipment and advanced manufacturing. The bureau alleges the firms opened offices without permission or presented themselves as local or foreign-invested companies, and recruited island-based engineers at high salaries. Alibaba's Qwen Office cleared CAICT office agent assessment, becoming the first product to pass the framework run by the research institute of China's industry ministry. CAICT says the framework exists to screen agents who acting on their own risk prompt abuse, information leakage and untrustworthy results. |
| Chinese think tank ranks the metals that give Beijing the most control over AI chipmaking A report from the Pangoal Institution, a Chinese think tank, sorted the metals China dominates into three tiers by how directly each one constrains AI chip manufacturing. It treats China's October 2025 rare earth rule as the template by turning on when the buyer makes logic chips at 14 nanometers and below or memory chips at 256 layers and above. The top tier is rare earths and gallium, where a control measure would hit AI chip production directly, including lanthanum, cerium, yttrium, and lutetium that go into gate dielectrics, the insulating layer inside an advanced transistor. Tungsten, germanium and indium fall to a second tier, constraining memory and optical links only indirectly. Quoted 2026 prices for tungsten hexafluoride, the gas that deposits chip wiring, have risen more than 200%. |
| Taiwan searches 64 offices of 17 mainland-backed chip and equipment firms Taiwan's Ministry of Justice Investigation Bureau, the justice ministry's criminal investigation arm, said yesterday it searched 64 offices belonging to 17 mainland-backed companies. It questioned about 114 people. It is the second island-wide sweep of 2026, and it covers more companies and locations than the March round. That round hit 11 firms at 49 sites. The searches ran from July 13 to Tuesday. The bureau says the targeted firms concentrate in chip design, semiconductor equipment and advanced manufacturing. It alleges they set up offices without permission or presented themselves as local or foreign-invested companies. It also says they recruited island-based engineers through third-party headhunters at high salaries. Named targets include semiconductor equipment maker ACM Research, chip design firm Goke Microelectronics and lithium battery maker Zhuhai CosMX Battery. |
Alibaba's Qwen Office is first to clear China's new state test for office AI agentsAlibaba's Qwen Office became the first product to pass the office agent assessment run by the China Academy of Information and Communications Technology, the research institute of China's industry ministry. Coming three days after its August 3rd launch. The assessment sets a state bar for software that carries out multi-step work inside a company's documents and tools rather than only answering questions. CAICT says the framework exists because agents acting on their own risk prompt abuse, information leakage and untrustworthy results. It runs 16 tests across four areas: interaction and multimodal capability, execution of office tasks, security protection and management of the packaged abilities an agent can call on. Qwen Office met the standard on all 16. Alibaba says the system splits work between the device and the cloud, running the agent in a walled-off environment with runtime monitoring and one-click termination. |
| Envision opens a 120,000-square-meter AI data center built for domestic chip clusters Envision Group said its Ulanqab Galaxy Base has begun operating, built around a single 120,000-square-meter building. Concentrating that scale in one structure yields 10 times the compute output per unit area of a conventional data center, the company says. It says the base has million-card parallel capability, meaning accelerator chips numbering in the millions running as one system. Ulanqab is one of eight national hubs under East Data West Computing, the state plan that moves data processing inland from China's coastal cities. The general manager of Envision AIDC, Zheng Zihao, said the base offers domestic chip clusters a large enough deployment site, sufficient power and smooth energy logistics. Envision puts green power at 67% of the region's supply, delivered through wind farms it built and dedicated transmission lines, with more than 2 gigawatts planned for the park. |
| Huawei reports zero mismatch between training and inference on its Ascend AI chips Huawei Computing said its Ascend accelerators now run a full set of operators, the low-level math routines a chip executes, that return identical results in training and inference. That matters for reinforcement learning post-training, where a model is scored on outputs it generated and then updated, where a mismatch between the two engines corrupts the score. On a 30-billion-parameter Qwen3 model, Huawei measured zero difference between the two engines on the same sequence of tokens. Mismatch can occur when accumulation does not preserve order in low-level computation, that is amplified by a model being split across many chips. Huawei also reports a 20% to 60% speedup on an attention routine tuned for Ascend, and has published the supporting code on GitHub. |
|
| ByteDance tells staff it will keep building its own language models and accept falling behind ByteDance chief executive Liang Rubo told a midyear all-hands meeting that leading overseas models have widened their lead, and that the company will keep building its own. That rules out the cheapest catch-up route, since ByteDance internally bans distilling from open-source models and enforces the ban partly through detection on its programming interfaces. Founder Zhang Yiming said at a recent internal meeting that model building requires long-termism, not using other systems' outputs to trade for a temporary leaderboard ranking. Liang rejected outside reporting that the in-house strategy responds to external pressure, saying the company will accept a period of lagging behind. He said Doubao, ByteDance's consumer AI assistant, remains competitive, and that its Seedance video generation model remains among the best available. |
Alibaba's Qwen3.8-Max takes first place on an independent AI agent rankingThree days after Alibaba's Aug. 3 Qwen3.8 release, its Qwen3.8-Max variant took first place on the Agentic Index with 55.4 points, in rankings published by Artificial Analysis, an independent model performance tracker. The index scores how well a model calls tools to work through multi-step problems, a category the Claude and GPT families have long led. Qwen3.8-Max passed Anthropic's Claude Opus 5 and OpenAI's GPT-5.6 on that measure. The previous best score by a Chinese model was Moonshot's Kimi K3, at 50.1 points. |
|
| III | CHIPS & SEMICONDUCTORS | |
| Chip designer SmarCo shows a video AI processor built on the open RISC-V standard SmarCo, also known as Zhongke Tongliang, showed a video AI processor called the Jingang GC3 at the CCF Chip conference in Wuxi. It combines a dataflow architecture with RISC-V, the instruction set anyone can implement without paying a license fee. The company says that is a first for video intelligence chips. A dataflow chip has no program counter, the register that tracks which instruction runs next. Computation triggers when a node's inputs are ready, so intermediate results pass down the pipeline instead of returning to memory each step. The 12-core design runs at 2.0 GHz and delivers 200 trillion 8-bit integer operations per second, the precision used for most vision inference work. Shanghai Jiao Tong University professor Leng Jingwen said its peak capability on video AI inference is comparable to mainstream graphics processors. |
|
| IV | QUANTUM & FRONTIER SCIENCE | |
| China's new heavy-ion accelerator identifies an atomic nucleus never reported before Researchers at the Chinese Academy of Sciences Institute of Modern Physics produced and identified hafnium-153, a form of hafnium not previously reported. It is the first physics result from HIAF, the High Intensity heavy-ion Accelerator Facility, which is still in commissioning. The team fired a beam of bismuth-209 at a graphite target, shearing fragments off the beam nuclei to produce a spray of radioactive nuclides. A ring spectrometer then weighed the products by timing how long each took to circle the ring. Across the run the team observed 10 hafnium-153 ions, a nuclide that sits in the neutron-deficient region of heavy nuclei, where physicists probe the limits of nuclear stability. Japan's RIKEN reported observing hafnium-153 in the same period. Detection capability at HIAF is expected to improve by a factor of about 100 as performance rises. |
|
| Chinese fiber maker YOFC supplies hollow-core cable for a 3.2-terabit Hong Kong AI link YOFC said Hong Kong Telecommunications will run a 3.2 terabit-per-second data center interconnect for AI traffic on YOFC hollow-core fiber cable. The link quadruples an existing 800 gigabit-per-second AI channel, and its core segment runs from the Lok Ma Chau Loop to Tseung Kwan O. Hollow-core fiber carries light through air inside the cable rather than through solid glass, which cuts the delay light takes to travel. HKT is marketing the channel as ultra-low latency and plans to connect the territory's main data center clusters and a new AI building at Lok Ma Chau. YOFC says it has delivered more than 10,000 fiber-kilometers of hollow-core fiber, with attenuation as low as 0.04 decibels per kilometer, a measure of signal lost over distance. |
| State telecom institute says China's eSIM phone rollout will run on real-name identity checks The China Academy of Information and Communications Technology published its 2026 eSIM industry report, calling China the world's highest-potential eSIM market. That would extend to phones a technology China cleared for commercial trials only in October 2025. China's industry ministry approved China Telecom, China Mobile and China Unicom to run those trials. An eSIM is a subscriber module soldered into the phone and provisioned over the air, replacing the removable plastic card. The report says keys stay permanently inside a soldered security chip certified to EAL4+ or above, an international security rating, and that this physically rules out cloning. China then layers real-name verification, device and model identifier checks and behavioral risk control on top. Global eSIM device shipments reached 605 million units in 2025, up 18% year on year. |
| Huawei's NearLink wireless standard may add a battery-free radio mode in version 3.0 A blogger who tracks specifications for NearLink, Huawei's short-range wireless standard, said that version 3.0 will add a fourth air-interface mode supporting passive radio frequency. Passive devices carry no battery and draw power from the reader's radio signal, so the mode would put NearLink into territory held by contactless cards and inventory tags. Asked whether that would displace NFC, the near-field standard behind contactless payments, the blogger said it would probably reach all of RFID, the radio tag standard used in access badges and warehouse tracking. Huawei documents NearLink as evolving in three modes, each with its own generation number: E for low energy, B for broadband and P for positioning. NearLink E2.0 adds NearLink audio on top of the discoverability and device-to-device connectivity in E1.0. |
|
|
|