← China Tech Daily
About Editorial Standards Corrections Privacy Policy
Subscribe →
China Tech · Daily chinatechdaily.org
OCTOBER 28, 2026
Your morning brief on Chinese technology. Published by the AI Policy Institute
Bottom Line Up Front

ByteDance, Alibaba and Tencent have combined order demand for more than 3 million of Nvidia's RTX PRO 5500 workstation graphics cards, worth more than $35 billion at list price, which would exceed Nvidia's expected global output of the card in 2027. Beijing's municipal government issued an energy plan for the 2026-2030 period that calls for every new computing center in the city to use 100% green electricity. Huawei unveiled what it calls the first commercial cell-tower gear for upper 6GHz airwaves, which it says lets carriers add the next-generation band to existing towers. Infinigence AI's agents held 88 first places out of 235 on SOL-ExecBench, an Nvidia Research benchmark for optimizing the GPU code that runs AI models, as of Monday.

ITOP STORIES
 
ByteDance, Alibaba and Tencent seek more than 3 million Nvidia workstation GPUs
ByteDance, Alibaba and Tencent have combined order demand for more than 3 million of Nvidia's RTX PRO 5500 workstation graphics cards, worth more than $35 billion at list price, per Reuters. That alone would exceed Nvidia's expected global output of the card, no more than 3 million units in 2027, so how many the companies receive depends on how many Nvidia can make. The card, from Nvidia's Blackwell generation, is less powerful than the company's data-center chips and would be used mainly for inferenceInferenceRunning a trained AI model to answer a query, as distinct from training it. Inference is where most ongoing compute and energy demand now sits, and it is what chip export rules increasingly target alongside training., running trained AI models for users. TMTPost notes that exports of Nvidia's H200 data-center chip to China must meet specific license conditions and that the more powerful B300 cannot be sold there under current rules.
Read at TMTPost ↗
Beijing energy plan calls for new computing centers to run entirely on green electricity
Beijing's municipal government issued an energy plan for the 2026-2030 five-year period that calls for every new computing center in the city to use 100% green electricity. Existing data centers are to bring power usage effectiveness (PUE), total facility electricity divided by what the computers draw, below 1.35. A PUE of 1.35 means no more than 35% extra for cooling and other overhead. The plan also calls for energy-saving retrofits of computing facilities, telecom base stations and server rooms.
Read at Yicaistate media ↗
Huawei claims first commercial radio for upper 6GHz band eyed for next-generation networks
Huawei unveiled VitaAAU, a combined radio and antenna for cell sites that it calls the world's first commercial equipment for the Upper 6GHz band, spectrum operators are weighing for next-generation networks. Huawei says the unit matches the coverage of C-band gear, the mid-band equipment that carries most 5G traffic, so operators could add the new band on sites they already use. Its redesigned antenna array covers about 30% more area than the industry's 1,024-element designs, the company says. Commercial deployment has begun in markets including the United Arab Emirates, by Huawei's account.
Read at Leiphone ↗
Chinese firm Infinigence's AI agents win most first places on Nvidia GPU-code benchmark
Infinigence AI's agent-written code held 88 of the 235 first places on SOL-ExecBench, an Nvidia Research benchmark for optimizing GPU code, as of Monday, more than twice the next team's count. Kernels are the small programs that run each math operation of an AI model on a graphics chip, and faster ones get more work out of the same hardware. The benchmark draws its problems from real frontier models and scores each entry against a ceiling set by the chip's physical limits. Rivals among hundreds of teams include data platform Databricks and coding-tool maker Cursor. Infinigence says its agents test their code in containers cut off from the network so they cannot tamper with scoring scripts.
Read at QbitAI ↗
IIPOLICY & REGULATION
 
Revised Chinese standards guide for connected cars is out for public comment
A revised national guide to the standards planned for connected vehicles, cars that exchange data with networks, roads and each other, was presented at the 2026 World Intelligent Connected Vehicles Conference after China's industry ministry posted it for public comment Oct. 19. China's standards, some of them mandatory, set what connected cars sold there must do, so the guide shapes the requirements carmakers and their suppliers will have to meet. Sun Hang, chief engineer of the China Automotive Standardization Research Institute, said 90 national and industry standards for connected vehicles have been published and 62 more are planned. Mandatory national standards on vehicle cybersecurity and software updates took effect this year.
Read at Gasgoo ↗
China's industry ministry proposes 34 top-tier smart factories, with Xiaomi's EV plant listed first
The Ministry of Industry and Information Technology (MIIT), China's industrial policy body, published for public review 34 projects it plans to cultivate as "leading-level" smart factories. The list names the plants the ministry holds up as models of automated, data-driven manufacturing, adding to the 15 it has cultivated at that level before. Leading-level is the top tier of the ministry's four-level rating. Xiaomi's car unit heads the list with what it calls an AI driven super factory for high-end electric and plug-in hybrid cars. FAW-Volkswagen ranks seventh with lines that build fuel, electric and hybrid models together, and Seres ranks 29th.
Read at Mydrivers ↗
IIIAI & FOUNDATION MODELS
 
AI developer DeepSeek's code lists placeholders for an unreleased V5 model
Developers found three previously unseen model identifiers, deepseek-v5, deepseek-v5-open and deepseek-v5-next, in a test file in DeepSeek's official deepseek-harness code repository on GitHub. The entries are the first sign of a V5 model in DeepSeek's own code. They set up two ways to move users onto it, one forcing a client software upgrade and one leaving older clients working. IT Home, which reported the find, cautions that the entries are placeholders no one can call and do not mean a release is near. DeepSeek released its V4 series in preview on April 24 and V4.1-Flash, which it says uses a new architecture, in September.
Read at IT Home ↗
Ant Group open-sources base version of its Ling AI model under permissive license
Ant Group's Bailing team today open-sourced Ling-3.1-flash-base, the pretrained model behind the Ling-3.1-flash release CTD reported in the October 16th edition. The base weights carry an MIT license, a permissive license that lets any developer in China or abroad retrain the 560-billion-parameter model and build it into commercial products. Ant also released Ling-3.1-flash-1019, an updated version of the Ling-3.1-flash model. On Ant's own tests, the base model beat DeepSeek-V4-Pro-Base and MiMo-V2.5-Pro-Base on 32 of 47 benchmarks. A vision variant, Ling-3.1-flash-VL, already runs inside Ant's LingGuang assistant app and will be released publicly later.
Read at InfoQ China ↗
IVCHIPS & SEMICONDUCTORS
 
Huawei consumer chief credits industry-wide effort for Kirin phone chip's 2023 return
He Gang, Huawei's consumer business chief, said at a launch today that the company lost access to advanced global inputs after 2020 for reasons he said were widely known. He said years of work across the whole industry chain brought its Kirin chip back in the Mate 60 phones in 2023. Kirin chips are Huawei's own phone processors. He also called this year's Kirin 9050 series the world's first "logic-folding" chip, the chip-stacking technique CTD reported October 13th.
Read at IT Home ↗

Thanks for reading. See you tomorrow.— Daniel

Comments or corrections? hello@theaipi.org

 
Editor-in-Chief:Daniel Colson
Managing Editor:Min Goodman-Cheng
China Policy Fellow:Tyler Liggett
Lead Software Engineer:Christopher Käck
Research Manager:Philip Wieczorek
Operations Specialist:Joanne Chua
Research Assistants:Claude Code + Codex
 

View in browser · LinkedIn

AI Policy Institute · 700 Pennsylvania Avenue SE, Washington, DC 20003

China TechDaily
About Editorial Standards Corrections Privacy Policy
© 2026 AI POLICY INSTITUTE · WASHINGTON D.C.