---
date: 2026-09-17
subject: "Huawei's Ascend 960 roadmap | Zhipu claims early self-improvement | Xi on advanced manufacturing"
---

**Huawei announced the Ascend 960 accelerator** at the opening of Huawei Connect 2026, along with a SuperPoD that runs 4,096 cards as one machine, and said its next two chips will keep doubling compute. **Zhipu said an AI agent** built its production inference software on a cluster of more than 100,000 domestic chips, claiming it as an early form of recursive self-improvement and raising end-to-end throughput to 3.2 times. **Xi Jinping issued instructions on advanced manufacturing** ahead of a national conference in Beijing, telling the sector to keep to intelligent, green and integrated development, raise the degree of self-control over supply chains, and build a modern industrial system around advanced manufacturing.

# Top Stories

- **Huawei sets an annual doubling schedule for its Ascend AI chips through 2029** — Huawei vice chairman Wang Tao set a delivery timetable for the Ascend 960 AI accelerator at the opening of Huawei Connect 2026. The 960DT arrives in the first quarter of 2027, three quarters ahead of plan, at 2 petaflops of FP8 compute and 4 petaflops FP4, up to 288GB of on-chip HBM and 9.6TB/s of memory bandwidth; the 960PR follows in the third quarter at 8 petaflops FP4. A SuperPoD wiring 4,096 cards into one machine, rated at 8 exaflops FP8 and 1PB of HBM, goes on sale in the second and third quarters of 2027, air-cooled then liquid-cooled. Wiring more cards into a single coherent machine cuts the share of training time lost to moving data between chips: in a Huawei simulation of a 100,000-card cluster built from conventional eight-card servers, that traffic can consume more than 40% of total training time. The Ascend 970 and 980 follow in 2028 and 2029, doubling compute again under what Huawei calls its Tau (τ) Scaling Law. [Jiemian](https://www.jiemian.com/article/15108082.html) [state-affiliated] [InfoQ China](https://www.infoq.cn/article/bmducufWEHZZRxEYjM4l)

- **Zhipu claims an early form of recursive self-improvement** — An AI agent running on Zhipu's GLM-5.3 model helped build and optimize the company's production inference system from scratch on a cluster of more than 100,000 domestic chips, founder Tang Jie said. Tang called it an early internal case of recursive self-improvement, an AI system improving the machinery that produces it. End-to-end throughput reached 3.2 times the initial baseline in under two weeks. The agent traced a concurrency bottleneck to Python's global interpreter lock, which lets only one thread run Python code at a time. In some test scenarios, that lock was costing more than 20% of performance when prompt processing ran alongside cached data transfer, a gap the agent cut to under 1%. Engineers still set the goals, the boundaries and the risk judgments, Tang said. Zhipu says it has not achieved recursive self-improvement. [QbitAI](https://www.qbitai.com/2026/09/491357.html)

- **Xi issues instructions on developing advanced manufacturing** — The National Advanced Manufacturing Conference met in Beijing on Sept. 16 and 17 by video link. Xi Jinping told the sector to keep to intelligent, green and integrated development, raise the degree of self-control over supply chains, and build a modern industrial system with advanced manufacturing as its backbone. Premier Li Qiang, who attended and spoke, named a new generation of intelligent manufacturing as the main line of attack and called for carrying the "Artificial Intelligence + Manufacturing" initiative deeper. That initiative is the factory-floor arm of a wider state campaign to drive AI into named sectors one by one, here by funding and guiding company digital transitions and building up domestic intelligent software and hardware. [China Military Online](http://www.81.cn/jwtt/16486998.html) [govt]

- **Xiaomi puts its MiMo-V2.6 language model training run on a public dashboard at about $31,000 an hour** — Xiaomi's MiMo-V2.6 large language model reached the halfway point of its reinforcement learning, team lead Luo Fuli said. She published a live dashboard of the run, putting the compute bill for a Chinese frontier training job on public view while it is still running. Reinforcement learning is the phase in which a model is tuned against feedback on its own outputs rather than on raw text. Combined training cost across the MiMo-V2.6-Pro and MiMo-V2.6-Flash runs had passed $1.35 million as of today, or about $31,000 an hour. The team is scaling three things at once: compute, the test environments the model practices in, and the compute spent grading its own answers. Xiaomi will open-source details of the run over the coming weeks, Luo said. [Jiemian](https://www.jiemian.com/article/15107931.html) [state-affiliated]

- **China orbits the 25th batch of satellites for its low-orbit internet constellation** — China launched the 25th group of low-orbit internet satellites on a Long March 12 rocket from the commercial spaceport at Wenchang, on Hainan island. Satellites in low orbit relay broadband from a few hundred kilometers up, close enough that signals travel back and forth faster than from the far higher geostationary satellites traditional satellite internet uses. The satellites entered their planned orbit and the mission was a complete success, per Xinhua. [China Military Online](http://www.81.cn/yw_208727/16486985.html) [govt]

# Policy & Regulation

- **China begins drafting its first automotive standard for RISC-V chips, with 34 companies taking part.** — Drafting has begun on China's first automotive industry standard for RISC-V chips, with 34 domestic and foreign companies taking part. RISC-V is an open, royalty-free instruction set, the basic command vocabulary a processor understands. Yan said the architecture gives carmakers a new route for automotive chips, but identified a number of obstacles, including fragmentation between incompatible versions, and the functional safety, security and real-time demands cars impose. Phase one constrains the base and extension instruction sets and defines CPU core requirements, with phase two covering system integration interfaces and compatibility testing. The work sits under a national automotive chip standards guide issued in January 2024 by the Ministry of Industry and Information Technology (MIIT), which sets industrial policy. [Gasgoo](https://auto.gasgoo.com/news/202609/16I70472190C601.shtml)

- **MIIT seeks comment on a plan for more than 50 national standards on EV battery recycling** — MIIT's science and technology department opened public comment Tuesday on a guide to the standards system for recycling power batteries. That puts a target and a deadline on the recycling piece of the battery-standards acceleration reported in the [September 3rd](https://chinatechdaily.org/archive/china-daily/2026-09-03/) edition. The draft covers the chain rather than the firms, Gasgoo writes. The 2018 white list settled which companies may recycle batteries but left transport safety, traceability and purity mostly to industry self-regulation. Cumulative power battery installations in China passed 1,200 gigawatt-hours by 2025. Because the first mass-market electric platforms reached production around 2018 and packs last about eight years, retirements climb year on year from 2025. The guide targets more than 50 national standards drafted or revised by 2030, and comments close Oct. 15. [Gasgoo](https://auto.gasgoo.com/news/202609/16I70472178C501.shtml)

# AI & Foundation Models

- **Unisound's U2-Flash model activates under 4% of its parameters and roughly doubles its predecessor's score on a coding benchmark** — Unisound released U2-Flash, a language model the company says took part in its own post-training loop, generating training data and repairing the training system. The gain is in how much work the model gets from a fixed amount of compute, not in raw scale. Of roughly 266 billion total parameters, under 4% are active per query, a sparse mixture-of-experts design that routes each request to a small slice of the network. On DeepSWE v1.1, a test that scores models on software engineering tasks, U2-Flash reached 64.6, roughly double the previous U2 model's result, by the company's own reported figures. U2-Flash is live on the front page of Unisound's model service platform, with an online trial and API access on request. [QbitAI](https://www.qbitai.com/2026/09/491091.html)

# Chips & Semiconductors

- **Third Paradigm releases an electromagnetic simulator that runs on graphics processors, aimed at advanced chip packaging** — Third Paradigm (Ningbo) released M3EM Studio 2026 R2 yesterday, calling it China's first cross-scale, system-level electromagnetic simulation package written to run natively on graphics processors rather than on ordinary server chips. The software predicts how electric and magnetic fields behave inside a product before it is built, moving those checks from prototype testing into the design stage. Third Paradigm says a whole-device model of a hundred million mesh cells, the small elements a design is divided into, now solves in hours rather than days. The company aims it at advanced chip packaging, intelligent vehicles and radar, among others, with validation cases including a 14-layer package substrate and a full 5G smartphone model. A built-in interface lets engineers run modeling, meshing and solving in plain language. The release follows a Sept. 11 MIIT briefing on pushing artificial intelligence into industrial software. [InfoQ China](https://www.infoq.cn/article/UZM5vKcN4CbNkPUS6Aid)

# Quantum & Frontier Science

- **Ningbo researchers report the first directional control of atomic-layer sliding in a two-dimensional ferroelectric** — Wang Hongwei of Ningbo University and colleagues at two other Ningbo institutions report the first directional control of interlayer sliding in a ferroelectric only atomic layers thick, in Physical Review Letters. Sliding ferroelectrics flip polarization by shifting whole atomic layers against one another rather than displacing ions, which makes them candidates for low-energy memory. The obstacle has been direction, since several equivalent sliding paths coexist and nothing fixes which the layers take. Using bilayer boron nitride, the team paired calculations starting from the equations of quantum mechanics with AI assisted simulation tracking more than a million atoms. Straining the material along one axis breaks its three-fold rotational symmetry, after which an electric field can slide the layers only one way. Simulations put one polarization reversal at about 1.7 picoseconds, three to four orders of magnitude faster than conventional ferroelectrics. [ITHome](https://www.ithome.com/1/003/591.htm)

# Digital Infrastructure

- **FiberHome draws 110 kilometers of hollow-core fiber onto a single spool** — FiberHome Telecommunication Technologies says it has drawn 110.6 kilometers of anti-resonant hollow-core fiber continuously onto a single spool, which it calls a domestic record. Hollow-core fiber carries light down an air channel instead of through solid glass, cutting the delay a signal takes to cross a link by roughly a third. Preparing the fiber at the hundred-kilometer scale had been the sticking point, and FiberHome says it kept the drawn product's structure uniform and stable at that length. [Yicai](https://www.yicai.com/brief/103368276.html) [state-affiliated]