|
|
|
|
Bottom Line Up Front China's Ministry of Commerce tightened drone export controls today, placing U.S.-bound drones, key components and related technology under case-by-case review, barring transactions with six U.S. entities, and citing U.S. actions since the Busan summit such as the July 31 addition of more than 40 Chinese entities to an import ban list. The ministry opened a national security investigation into imported printing and copying equipment, focused on foreign-developed driver and embedded software, with a 12-month deadline that can be extended and a 30-day window for written comments to the Trade Remedy and Investigation Bureau. The FCC is drafting an import restriction on newly introduced Chinese-made optical modules, leaving already certified products in place as the industry shifts from 1.6-terabit to 3.2-terabit links, in a datacom optical module market LightCounting sized at $23.8 billion in 2025 where Chinese suppliers including Innolight and Eoptolink hold seven of the top ten positions. |
| Beijing puts listed dual-use drone exports to the U.S. under case-by-case review China's Ministry of Commerce tightened export controls on U.S.-bound drones today, ordering strict case-by-case review effective immediately. The measure covers drones, their key components and related technology on China's dual-use export control list, meaning goods with both civilian and military uses. Those items no longer qualify for the license facilitation measures that had applied, so each shipment goes through individual approval. The ministry also listed six U.S. entities that it said support U.S. sanctions over Xinjiang: Applied DNA Sciences, Stratum Reservoir, Altana Technologies, the Responsible Business Alliance, Verite Group and Human Rights in China. Organizations and individuals in China may not transact with them, and the ministry said it will act again if Washington issues new restrictions. It linked the package to U.S. steps taken since the two presidents met in Busan, including more than 40 Chinese entities added to a U.S. import ban list on July 31. |
| China opens a 12-month security review of imported printers running foreign software China's Ministry of Commerce opened a foreign trade national security investigation into imported printing and copying equipment. The subject is the software rather than the machines, including foreign system software as driver and embedded software developed, tested or maintained by foreign individuals or entities. Driver software is what lets a computer operate a device, while embedded software is the code that runs inside the machine itself. The ministry said it will examine import volumes, how far domestic demand depends on foreign goods and technology, and the capacity of Chinese industry to meet it. The investigation must conclude within 12 months, though it can be extended in special circumstances. Interested parties have 30 days from publication to file written comments with the ministry's Trade Remedy and Investigation Bureau. |
| FCC draft rule would bar new Chinese optical modules from U.S. data centers The Federal Communications Commission is drafting a rule to bar imports of newly introduced optical modules made in China, per a report yesterday citing people familiar with the plan. Modules already certified and shipping in volume would be untouched, so the rule would land on the industry's shift from 1.6-terabit links to 3.2-terabit ones rather than on installed equipment. The author argues most of the world's high-speed modules come from Chinese makers, and non-Chinese suppliers cannot meet North American demand on their own. Innolight and Eoptolink also built plants in Thailand years ago and now ship to North America from there, so a ban on China-made goods could miss its target unless Washington judges origin by ownership rather than factory location. |
| Chinese system built on one open model ranks third on an AI bug-finding benchmark CyberGym published new rankings, and a Chinese system called DoGNAVY placed third globally and first among open-source entries with a 90.8% pass rate. DoGNAVY runs on one open-weight model that anyone can download. The higher-ranked Microsoft and Google entries route each stage of the work to whichever top closed model performs best at it. CyberGym is a benchmark built at the University of California, Berkeley that scores whether AI agents can reproduce known software flaws. Its 1,507 test cases come from past vulnerabilities in 188 large software projects, and an agent must produce a working test that triggers the flaw. DoGNAVY's model is GLM-5.2, which Zhipu AI open-sourced in June. The system was built with the security team DARKNAVY. Agents evaluated on CyberGym have found 34 previously unknown vulnerabilities and 18 incomplete patches in the tested projects. |
|
| Chinese startup Sand.ai open-sources a video model estimated to cut generation cost tenfold Sand.ai released the code and weights for MAGI-2-preview, a video generation model. The model holds 114 billion parameters but activates about 6 billion on each pass. A mixture-of-experts design routes each piece of input to a small subset of specialized sub-networks, so a large model runs at small-model cost. By QbitAI's calculation at current rates for eight Nvidia H100 chips, 10 seconds of 1080p video costs about 0.5 yuan (about 7 cents), roughly a tenth of mainstream models. The model feeds text, video and audio into one shared network that exchanges information between the three at every layer, instead of processing each type separately and joining them at the end. MAGI-2-preview ranks sixth on the video generation leaderboard run by Artificial Analysis, an independent model performance tracker. |
| ModelBest open-sources an AI system that speeds up industrial simulation code 2.35 times ModelBest and the OpenBMB open-source community released ForgeStencil on Tuesday, an AI system that rewrites the compute routines at the core of physical simulation software. Those routines, called stencil computations, update each point on a grid from its neighbors. They run over and over through a simulation, so their speed sets the length of the whole run. At 32-bit precision on identical hardware, it runs 2.35 times faster on average than mainstream code generation frameworks including Halide and Devito. The system uses two cooperating agents, one tuning the low-level routines and the other handling deployment into a working application, with no human expert in the loop. ModelBest says it ran the system on more than 100 real industrial and scientific codes in one week, spanning oil exploration, medical equipment and weather simulation. It published the code on GitHub. |
| ByteDance ships a model that can listen and speak at the same time ByteDance's Seed research team launched SeedRealtime today, rolling it out at once to all users of Doubao, its consumer chat assistant. The model handles audio, video and text in one architecture, replacing the usual chain of separate speech recognition, language model and text-to-speech steps that forces strict turn-taking. It is natively full duplex, meaning it can listen and speak at the same time rather than waiting for the user to finish. It also keeps watching its camera feed for changes so it can speak up when a target appears. It filters out background chatter to avoid false triggering. ByteDance's own human evaluation found dialogue pacing problems cut in half against a conventional cascaded system. |
| Tencent ships a speech recognition model with error rates near 3% in three languages Tencent's Hunyuan team released Hy ASR 3.0 preview, a speech recognition model built on its Hy3 large language model. Tencent put it on its cloud programming interface at launch, so outside developers can use it rather than only Tencent's own products. On open-source evaluation sets Tencent reports word error rates, the share of words a transcript gets wrong, of 3.34% for Mandarin, 2.62% for English and 3.12% for Cantonese. The company says those figures lead competing products. It reports similarly low rates on its in-house sets for dialects and hard acoustic conditions such as high noise and whispering. Tencent's Yuanbao assistant co-developed the model and shipped it first, offering dialect recognition and context-aware correction at no charge. |
|
| III | CHIPS & SEMICONDUCTORS | |
| Samsung and SK Hynix test Chinese etching tools inside their China memory plants Samsung Electronics and SK Hynix are evaluating Chinese-made semiconductor etching equipment at their fabs in China, per Reuters. The tools would keep installed lines running if U.S. controls extend to maintenance, repair or replacement parts for their American equipment. Etching removes material from a wafer to carve circuit patterns into it, one of the core steps in chipmaking. Samsung's Xi'an plant and SK Hynix's plants in Dalian and Wuxi make memory chips and depend heavily on Applied Materials and Lam Research for etch tools. Washington revoked the two companies' license-free import status for controlled U.S. equipment in 2025, then granted both 2026 licenses covering continued imports to their China sites. Samsung said it had not tested the equipment, and SK Hynix declined to comment. |
| Tsinghua spinout Meike Shensi runs light-based AI chips in a 1,000-card pilot Meike Shensi has completed a commercial pilot of its optical computing chips at a domestic AI computing center running about 1,000 accelerator cards, co-founder and CEO Zou Tianqi said. The chip is meant to work alongside GPUs rather than replace them, taking over the stage where a model generates output one token at a time. GPUs keep the first stage, which processes the entire input prompt in one pass. The company uses free-space optical computing, sending light through open space instead of through waveguides etched onto a chip. Zou said the optical array has reached 400 by 400 elements, a scale he credits with a hundredfold gain in parallel processing under this design. In July 2026 the company launched a joint computing platform with domestic GPU maker Biren Technology and model developer StepFun. |
|
| IV | ROBOTICS & AUTONOMOUS SYSTEMS | |
| Chinese robot makers redirect exports to Europe and Asia after the U.S. certification block Chinese humanoid and four-legged robot makers are redirecting exports toward Europe, Japan and emerging markets in Southeast Asia and the Middle East. The FCC block on U.S. certification for new foreign-made robots, in force since July 28, lets brands keep selling older authorized models but not their successors. The listing covers mobile machines rather than fixed robotic arms. China exported 10.377 million robots of all categories in the first five months of 2026, worth nearly 20 billion yuan ($2.8 billion), to more than 150 countries. A July video appearing to show a Chinese robot attacking people drew more than 120 million views abroad. The machine was a Unitree G1 running an operating system built by Poland's MERA Robotics, filmed by an Indonesian marketing team. MERA said it was under human control. |
| X Square Robot open-sources a framework that teaches a robot a task from one video X Square Robot open-sourced a framework called HOST on Monday with Beijing Institute of Technology and Tsinghua University. A robot running it acquires a new task from one human demonstration video, averaging 29 seconds. That removes the per-skill robot data collection and fine-tuning the standard approach requires. X Square Robot reports 62% success across 50 tabletop manipulation tasks the model never saw in training, including placing fruit, stacking bowls and wiping plates. HOST aligns video frames to robot action steps by stretching and compressing timing rather than by clock time. Because the weights stay fixed, old skills survive: the strongest fine-tuning baseline fell to 43% of its original performance on earlier tasks after learning new ones. |
| Xiaomi open-sources its robot foundation model with the pipeline that trains it Xiaomi said on its official Weibo account that it has open-sourced Xiaomi-Robotics-1, its embodied foundation model. The release covers the whole pipeline from real-robot post-training through deployment, so outside developers can reproduce Xiaomi's workflow rather than only run its weights. An embodied foundation model turns camera input and a written or spoken instruction into robot motion. Pretraining used more than 100,000 hours of data from the universal manipulation interface, a standard way of recording human manipulation demonstrations. Post-training used more than 10,000 hours of data gathered on robots with different bodies, so one model can drive more than one hardware design. The weights and the evaluation code for the relevant benchmarks are published on GitHub and Hugging Face. |
| China's robot touch sensor suppliers draw 7 billion yuan as orders shift to precision handling Chinese makers of tactile sensors for robots drew more than 7 billion yuan ($980 million) in new financing in the first seven months of 2026, by Securities Times' incomplete count. Hardware orders are climbing alongside a data shortage the industry has not solved, since touch data carries deformation, force and contact dynamics that other data types cannot supply. A Guolian Minsheng Securities report puts the current gap between the data robot developers have and what they need at 20-fold. Xense Robotics founder Ma Daolin said orders rose sharply this year as demand moved from basic perception to precision manipulation, with more than 300 paying sensor customers. Tashan Technology said its first-half 2026 shipments doubled its full-year 2025 total. Tashan recently released a dynamic tactile sensing chip, the Emerald E10A, which it says lets a robot detect an approaching object and begin responding before contact. |
|
|
|