|
|
|
|
Bottom Line Up Front China's data center power draw passed 60 billion kWh so far this year, the China Electricity Council told CCTV News, up more than 40% from a year earlier and already beyond the full-year 2024 total. Orienspace flew a sea launch today, placing eight Qianfan satellites and a test satellite in orbit and claiming national records for payload weight and altitude on a single launch from the water. Huawei published a vertical data center design yesterday at a conference in Wuhu, releasing outline drawings for a hall that stacks cooling, computing and power in layers and supports 168 MW. |
| China's data center power use tops 60 billion kWh this year, up more than 40% China's internet data services industry has drawn more than 60 billion kWh of electricity so far this year, up more than 40% from a year earlier and already past its full-year 2024 total. Power is what limits how fast new AI capacity can be built, since a hall needs grid supply before the chips inside it matter. CCTV News obtained the figure from the China Electricity Council, the body that compiles national power sector statistics. On the council's projection, set out by supervisor-general Pan Yuelong, computing draws about 800 billion kWh by 2030, or roughly 6% of all the electricity China consumes. |
Commercial rocket firm Orienspace sets China's sea-launch payload recordA Gravity-1 rocket lifted off from waters in the East China Sea, placing eight Qianfan networking satellites and a wireless communication test satellite into their planned orbits. It is the second Qianfan networking launch in two days, after LandSpace's Zhuque-2E put 10 of the satellites up in the September 15th edition. Developer Orienspace told Xinhua the flight set national records for payload weight and orbital altitude on a single sea launch, inserting more than 3 metric tons at about 800 km. Qianfan is China's low-earth-orbit broadband constellation, assembled from batches of small satellites, and this was the first time Gravity-1 has flown one of its networking missions. The payload mixed stacked satellite assemblies with flat-panel ones, a combination that raises the demands on separating spacecraft in sequence. |
| Huawei publishes a vertical data center design it calls a world first Huawei released what it calls the world's first 3D data center yesterday at an AI data center conference in Wuhu, Anhui, opening a library of outline design drawings. The design stacks a data hall vertically instead of spreading it sideways, with cooling at the bottom, the computing floor above it, then power supply and backup batteries on top. Stacking the layers shortens the power and cooling runs to a minimum, which Huawei says allows zoned cooling, keeps faults from spreading between zones and lets a single hall support 168 MW. Rotating chairman Wang Tao said clusters of more than 100,000 Ascend accelerators, Huawei's AI chip line, will become the baseline configuration. Huawei Cloud is already building to the design at three hubs, in Gui'an in Guizhou, Horinger in Inner Mongolia and Wuhu. |
| China's public security ministry says 50 million have enrolled in the national internet ID China's Ministry of Public Security says 50 million people have applied for its national network identity credential, with more than 160 million authentication checks delivered through it. The credential lets a user prove who they are to a website without handing an identity-card number to each platform separately. Li Tong, political commissar of the ministry's 11th Bureau, gave the numbers Monday at the main technology forum of China's Cybersecurity Week in Jinan, Shandong. More than 400 government departments, agencies and companies have adapted their services to it, among them provincial government platforms, the large state-owned banks and courier firms such as SF Express. |
| Mozilla puts the gap between Chinese open models and the U.S. frontier at 4.4 months Mozilla's second State of Open Source AI report puts the gap between the best Chinese open-weight models and the U.S. closed frontier at 4.4 months. The yardstick changed with it, from a percentage performance difference in the July edition to how long a task a model can finish reliably, so the two readings are not directly comparable. Open-weight models are ones whose trained parameters anyone can download and run on their own hardware. On the new measure the best closed models manage tasks of about 12 hours, against roughly 7 hours for the best open ones. DoorDash now runs Moonshot's Kimi, a Chinese open-weight model, for routine tasks and keeps Anthropic's Fable 5 for more complex work. |
|
| Hong Kong's first five-year plan names AI a strategic industry Chief Executive John Lee published Hong Kong's first Five-Year Plan for Economic and Social Development, covering 2026 to 2030 and naming artificial intelligence a strategic industry. The plan gives the territory a role in China's technology strategy running well beyond finance, exercised through the separate standing Hong Kong keeps in international regulation. It calls for a Sha Ling Data Park to supply compute to that industry, and for a review of the legal framework covering AI and intelligent systems. Hong Kong is also to join mainland Chinese aircraft development and certification work. A cross-border logistics network for drones and small aircraft flying below conventional air traffic is to span the territory, Macao and nine Guangdong cities. Lee set a target for total local innovation spending to reach 3% of GDP after 2030. |
|
| III | AI & FOUNDATION MODELS | |
| iFlytek releases a speech model trained on what it calls all-domestic compute iFlytek put its Spark-Audio-1.0-Preview speech foundation model into service today, trained on what the company calls a purely domestic compute cluster. The domestic-compute claim is particularly relevant, since export controls are built to limit China's training capacity. Behind it sits 13 million hours of audio and a mixture-of-experts design that activates only a fraction of its parameters on each token, holding down what it costs to run. One model covers recognition of 99 languages and 202 dialects, alongside translation, speaker identification, emotion analysis and audio question-answering. iFlytek's own benchmark table puts it above Google's Gemini-3.1 Pro on Fleurs, Kespeech and LibriSpeech, three standard speech recognition test sets, and ahead of Alibaba's similarly sized Qwen3.5-omni-flash on dialects. The company says the model still trails on instruction-following and dialogue, and the release carries a preview label. |
Alibaba's Qwen tops the all-time popularity ranking on model hub Hugging FaceAlibaba's Qwen3.8-27B has become the most-liked open-weight model in the history of Hugging Face, the main public repository where developers publish and download model weights, ahead of FLUX.1 and DeepSeek-R1. It is a different model from the Qwen3.8-Max snapshot in the September 2nd edition. Alibaba has open-sourced more than 460 Qwen models, with downloads past 3 billion and more than 300,000 derivative models built on them. Overseas developers have taken to calling the model a "local Opus 4.6," crediting it with the best cost-effectiveness on Code Arena, a head-to-head comparison of coding models. Reuters, Airbnb and Pinterest switched earlier to Qwen models to cut their dependence on closed systems, and Qwen3.8-27B is already the default in Perplexity's local-agent business. |
| Tsinghua-affiliated team's table-data model tops three structured data benchmarks Stable AI and Tsinghua University professor Cui Peng released LimiX-2, a foundation model for structured data scaled to 400 million parameters. Enterprise AI work sits mostly on this kind of data, so first place on these leaderboards puts a Tsinghua-affiliated group ahead of comparable models from major Western tech companies. Structured data means the rows and columns held in databases and spreadsheets, rather than the web text general language models train on. The team reports first place on TabArena, BCCO and TALENT, three international structured data benchmarks, ahead of comparable models from Google, SAP and Amazon. Rather than predict one pre-specified target column, the model represents every cell and learns the joint dependency of all the variables in the context. The code is published on GitHub and the weights on Hugging Face. |
|
| Huawei's computing unit aims to run any AI model efficiently on its own chips Huawei supervisory board chairman Guo Ping told new employees that the company's ICT and computing business aims to become "Nvidia," in meeting minutes posted to Huawei's internal Xinsheng forum. Guo named compute, data and talent as AI's three bases and put China behind on compute, which U.S. export controls target. For that business, the aim is not to own an in-house large model, he said, but to help customers build strong ones and make any model in the world run efficiently on Huawei hardware. That hardware spans Ascend AI accelerators, Kunpeng server processors and SuperPoD, the tightly coupled multi-chip clusters Huawei sells as single machines. Huawei's core strategy is focus, with no plans to expand beyond connectivity and computing. For the Ascend 950 series, Guo said the company hopes to narrow the gap with competitors through architectural change rather than by stacking parameters. |
| Circuit board maker Shengyi starts small-batch output of 1.6-terabit switch parts Shengyi Electronics said in an investor relations record that its supporting products for 1.6-terabit switches have entered small-batch mass production. Those parts run ahead of the company's 1.6-terabit optical module board, still at sampling and testing, so the two halves of the same link generation are arriving at different speeds. Shengyi makes the printed circuit boards that sit under optical modules, the transceivers moving data between switches inside a data center. Its 800-gigabit module board is in volume production, with shipments scaling gradually. The company says optical modules are a priority area and that it will pursue new-customer qualification to scale up high-end products. It has also reached small-batch supply of satellite communications ground equipment. |
|
| V | ROBOTICS & AUTONOMOUS SYSTEMS | |
| Huawei-backed Luxeed RX wins a license to road-test conditional self-driving Luxeed, a brand in Huawei's carmaking alliance, said its RX has been approved for a Level 3 autonomous-driving road-test license and is testing on open urban expressways. Level 3 means the car drives itself within defined conditions while a driver stays available to take over. An automotive blogger's description lists redundancy across perception, braking, steering and computing, with a second brain chip that takes over and pulls the car aside if the first fails. The car carries 38 fusion perception sensors, among them an image-grade lidar, which measures distance by timing pulses of light. Presale pricing starts at 299,800 yuan (about $45,000), with the Level 3 version from 359,800 yuan (about $54,000). |
| Pony.ai unveils a driverless heavy truck with a 70% cheaper autonomy kit Pony.ai and GAC Commercial Vehicle gave the world premiere of their jointly developed Level 4 autonomous heavy truck Monday at the IAA Transportation show in Hannover, Germany. Full redundancy in steering, braking, power and computing is the engineering bill a truck has to pay before it can run without a safety driver. Level 4 means the vehicle drives with no human fallback inside its defined operating area. Pony.ai supplies the autonomy stack and GAC Commercial Vehicle the vehicle, a battery-electric T9 with a drive-by-wire chassis. The fourth-generation autonomy kit costs 70% less to build than the previous one, Pony.ai says. Volume production is set for this year, for trunk logistics, dedicated-line haulage and port transport. A globalization plan announced alongside the truck targets Europe and the Middle East within two years. |
JD.com adds lidar maker Hesai as a partner on its robot programJD.com has entered a strategic cooperation with lidar maker Hesai Technology on physical AI, an addition to the robot buildout its JDDiscovery conference set out in the September 10th edition. Physical AI pairs a model with a body and sensors so it acts on the world rather than only reasoning about it. Hesai's Picasso 6D chip platform lets a single lidar capture color and depth together, removing the synchronization and calibration steps between a camera and a scanner. Its AT-series lidar already rides on JD Logistics' sixth-generation self-driving delivery vehicle, and a JT-series unit built for robots carries a 360-by-189-degree field of view. |
| Robotics AI firm GigaAI will run its models on D-Robotics edge chips GigaAI and D-Robotics signed a cooperation Monday to pair GigaAI's robot models with D-Robotics silicon. What limits humanoid and mobile robotics is whether inference can run on hardware a product can afford to carry. D-Robotics' Sunrise S600 platform delivers 560 TOPS, or trillions of operations a second, enough to run vision-language-action models on the robot itself rather than in a data center. GigaAI's GigaBrain handles natural-language task understanding, visual-spatial perception and task decomposition on the device, while GigaWorld generates scenes a robot can train inside. The joint engineering list names converting the GigaBrain-0.7 model to run on the S600 and quantized deployment, which cuts a model's numerical precision to fit constrained hardware. Deployment targets include GigaAI's Maker and Shiguang robot bodies, plus robotic arms and mobile chassis. |
| Hesai launches what it calls the world's smallest high-precision laser scanner Hesai launched what it calls the world's smallest lidar in its spatial-resolution class at the InterGEO surveying fair in Munich. Size and weight decide what an instrument can be carried on, so holding accuracy while shrinking the body widens where the class can be used. The MT series body measures about 6 cm across and weighs 300 g, appreciably smaller and lighter than the generation it follows. Ranging accuracy is 3.5 mm, an improvement on that generation, in a class aimed at the scanning work behind survey maps and as-built records of buildings. Hesai says the smaller laser spot sharpens object boundaries and reduces the mixing of returns from adjacent targets. The series keeps adding samples as scan time increases, so a longer dwell yields a denser point cloud rather than a repeated one. |
|
| Chinese foreign ministry presses EU to drop planned network equipment restrictions Chinese foreign ministry spokesperson Guo Jiakun called on the European Union to avoid discriminatory restrictive measures, after an open letter from the telecom lobby Connect Europe. The proposal would let the EU designate suppliers as high-risk on cybersecurity grounds and give operators three years to strip their equipment out. The letter says the European Commission's Cybersecurity Act amendment could impose up to 40 billion euros (about $47 billion) in equipment replacement costs on European telecom firms. Guo said earlier assessments had put the amendment's cost to the EU at more than 360 billion euros (about $420 billion) over five years. Protectionism cannot buy competitiveness, he said. Enforcement of the earlier guidance has been uneven, with only 10 EU countries having excluded Chinese hardware makers after the Commission's 2023 recommendation, though nearly all had a framework in place. |
|
|
|