AI Compute
Data centers · chips · modelsEvery story across China’s AI compute stack — infrastructure, supply chain, foundation models, and the companies behind them.
CXMT Forecasts 131-132.5 Billion Yuan Nine-Month Profit, Reversing Year-Earlier Loss
CXMT expects revenue of 265-270 billion yuan and attributable net profit of 131-132.5 billion yuan for the first three quarters of 2026, reversing a 5.28 billion yuan year-earlier loss as AI demand drives a DRAM upcycle. The company disclosed on the same day that its chairman has proposed a share buyback of 1-2 billion yuan.
OCP 2026: Vera Rubin Becomes the Industry's Reference Point
At OCP 2026, NVIDIA contributed its Vera Rubin rack and compute tray designs to the Open Compute Project through the open MGX architecture, releasing the MGX 1.2 specification with support for 250kW racks, blind-mate liquid cooling and a 5,000A busbar. The summit's agenda widened from server racks to data center templates, 800VDC power and grid resilience, with OEMs and the Taiwanese supply chain aligning around the platform.
Suanmiao Technology Raises Nearly 2 Billion Yuan in a Year for 3D-Stacked Inference Chips
Beijing AI chip startup Suanmiao Technology has raised nearly 2 billion yuan across four rounds in a year, backed by state investment platforms including E-Town Capital, to develop its 3D-stacked A4E inference chip. The chip taped out in June but remains some distance from mass production and large-scale customer procurement.
Zhongxingwei Launches Full-Stack Interconnect Chip Platform, Claims China's First PCIe 6.0 Switch
Chinese interconnect chip maker Zhongxingwei used a Wuxi launch event to show a full-scenario platform spanning general-purpose and AI computing interconnect, saying cumulative shipments top 800,000 chips and 12 products are in volume production. It also signed strategic agreements with seven partners including Kylin, UnionTech, Tencent Cloud, Inspur, H3C, ZTE and Lenovo.
Against the Price War: StepFun Raises Step 5 Preview API Pricing Ahead of Weight Release
StepFun doubled the cache-hit input price for its Step 5 Preview API to 0.70 yuan per million tokens from 00:30 Beijing time on October 14, 2026, leaving cache-miss input and output unchanged, one day before open-sourcing the model's weights. The move runs against three years of Chinese large-model API price cuts and, by analysts' reading, reflects tight compute supply and an attempt to set a price floor against customer self-hosting.
Axera Launches AX8860 Chip Series for Edge and On-Device Large Models
Axera used CIIF 2026 in Shanghai to launch its AX8860 high-compute series, a 6 nm edge AI chip rated at 256 TOPS with 128 GB of LPDDR5X memory and native FP4 support, aimed at running large models on factory floors and devices. The company also released the Yuanxi A256S module and showed inference cards and servers, positioning a continuous product range from chips to systems.
Lanxin Computing Lands Nearly 2 Billion Yuan Round as LX500 Server Chip Draws 200,000-Unit Orders
Shenzhen-based RISC-V chip designer Lanxin Computing has closed a Series B round of nearly 2 billion yuan, backed by returning shareholders alongside local state funds, corporate investors and venture firms. The company says its LX500 server chip has won orders for more than 200,000 units from customers including Lenovo, China Mobile and Tencent Cloud.
SenseTime Upgrades Yunyu Xingkong Model as Shanghai Deploys First Quantum City Government AI Agents
SenseTime's SenseCore and the Shanghai Municipal Bureau of Planning and Natural Resources have upgraded the Yunyu Xingkong large model, billed as China's first industry foundation model for planning and natural resources, strengthening spatial reasoning and task execution. The bureau also released its first batch of Quantum City government AI agents covering flood-control dispatch, spatial analysis and planning updates.
Non-GPU Accelerators Overtake GPUs in China's AI Chip Market, IDC Says
IDC data cited by Reuters puts NVIDIA at 34% of China's data center AI chip market in the first half, with shipments down 44% year-on-year while Huawei reached 572,000 units and a 28% share. Non-GPU accelerators grew nearly 80% to 1.016 million units, outshipping GPUs for the first time.
State, Corporate and VC Money Converge in Infinigence AI's 2.3 Billion Yuan Round
Tsinghua-affiliated Infinigence AI closed a 2.3 billion yuan Series C, taking total funding to 4.3 billion yuan and setting a record for China's AI-native infrastructure sector. The company owns no chips, models or data centers, instead scheduling heterogeneous compute across 16 chip types and third-party facilities.
Black Sesame Says Next-Gen Huashan A2000 Chip Family Is Now in Full Mass Production
Black Sesame Technologies' Huashan A1000 automotive-grade autonomous-driving chip has been named to the third round of Hubei's "Hubei Premium Products" list, with the company citing 58 TOPS at a typical 18W and dual ISO 26262 ASIL-B and AEC-Q100 Grade 2 certification. The company says its next-generation A2000 family, built for physical AI, has now fully entered mass production.
Iluvatar CoreX Brings Cloud-Edge-Device GPU Lineup to CIIF, Claims 40% Lower Token Cost
Iluvatar CoreX showed its Tiangai, Zhikai and Tongyang cloud-edge-device product matrix at the China International Industry Fair in Shanghai, debuting the Tiangai 300 IXL144 super node with 144 GPUs and 18 CPUs in a single rack. The company said it now serves more than 380 commercial customers across over 30 industries with more than 1,000 deployed projects.
China Mobile's Changping Node Adds 1 EFLOPS Ascend SuperPod, Backed by Nine-Chip Compatibility Layer
China Mobile says a 1,024-card Huawei Ascend 950DT SuperPod delivering 1 EFLOPS (FP8) has gone live at its Mobile Cloud Beijing node in Changping. The carrier also says Mobile Cloud has completed compatibility work for nine domestic AI chips and that its MoMA platform now hosts more than 300 models.
Biren's Next-Gen BR20X AI Chip Powers On as Company Plans HK$4.04 Billion Share Placement
Chinese GPU maker Biren Technology says its next-generation BR20X AI chip has powered on and entered post-tape-out validation, ahead of its earlier schedule. The company also announced a placement of 130 million new H shares at HK$31.08 to raise about HK$4.04 billion, with roughly 70% earmarked for supply chain procurement and commercialization.
Zhipu's Tang Jie Lays Out GLM's Three Scaling Directions on the Road to AGI
Speaking at the National University of Singapore, Zhipu founder and Tsinghua professor Tang Jie outlined three scaling directions for GLM — pre-training, post-training, and reasoning plus agents — and named recursive self-improvement and robot action as the two capability gaps still to close. He also highlighted long-term memory, rather than longer context windows, as the key to continual learning in the agent era.
Samsung Q3 Operating Profit Jumps Nearly Eightfold on AI Memory Demand
Samsung Electronics said preliminary third-quarter 2026 operating profit reached 107.4 trillion won, up nearly eightfold year-on-year and its first quarter above the 100 trillion won mark, driven by AI-related memory demand and HBM shipment growth. The same price surge is raising component costs for Samsung's own smartphone and consumer electronics units.
DeepSeek Nears 80 Billion Yuan Raise to Fund Ascend Chip Buildout
DeepSeek is close to completing a second funding round of at least 80 billion yuan ($12 billion) led by CATL and Tencent, with signed term sheets pointing toward nearly 100 billion yuan. Proceeds are earmarked for an Inner Mongolia data center slated to hold at least 160,000 Huawei Ascend accelerators, ahead of an IPO planned for early 2027.
Data Centre World Asia Fills Four Floors as Power and Cooling Dominate the Floor
Data Centre World Asia in Singapore drew more than 600 exhibitors across four floors of Sands Expo, with attendance up 25% year on year. On the floor, the conversation had shifted from site selection to power availability, cooling and delivery speed.
IDC and Inspur: China's Intelligent Computing Power to Hit 2,576.5 EFLOPS in 2026
A joint IDC and Inspur Information report projects China's intelligent computing power will reach 2,576.5 EFLOPS in 2026, up 87.9% year on year, and 10,524.3 EFLOPS by 2030. Beijing, Hangzhou, Shenzhen, Shanghai and Guangzhou lead the city rankings, which for the first time weigh agent deployment and inference infrastructure rather than data center counts alone.
Huawei Launches Mate 90 Series With Kirin 9050 Pro, Touting 28% Higher Transistor Density
Huawei launched its Mate 90 series in Beijing on October 1, with all five models running in-house Kirin chips across three tiers, topped by the Kirin 9050 Pro and its 3D "logic folding" architecture. Rising memory prices pushed starting prices up 1,000 to 1,500 yuan versus the previous generation, with sales opening the same day as the announcement.
DeepSeek and Huawei Release Joint Ascend Compute, Communication and Compiler Tools
DeepSeek has open-sourced a set of infrastructure components for Huawei's Ascend platform, including a TileLang compiler toolchain plus compute and communication libraries that mirror its earlier NVIDIA releases. Huawei separately published joint results on Ascend 950 super nodes, citing 375 GB/s dispatch and 347 GB/s combine bandwidth and offline inference figures for DeepSeek-V4.1-Flash.
Alibaba Unveils T-Head Zhenwu V900, Says Single Cluster Can Scale to 500,000 Cards
At the 2026 Apsara Conference, Alibaba said the global data center capacity it operates will top 20GW by 2032 and unveiled T-Head's Zhenwu V900 AI chip, which it says can scale to 500,000 cards in a single cluster. CEO Eddie Wu said a tight global AI data center supply chain is limiting how fast Alibaba Cloud can add compute.
Report on the Development of Generative AI Applications (2026)" Released: China's Generative AI Users Top 700 Million, Intelligent Computing Power Reaches 2,185 EFLOPS
A new CNNIC report says China's generative AI user base passed 700 million in the first half of 2026, while intelligent computing power reached 2,185 EFLOPS, up 177% year over year. The report also flags the launch of the country's first all-domestic 100,000-chip AI supercluster and 250.14 billion yuan of AI funding in H1.
Chinese Academy of Sciences Launches Frontier Science Breakthrough Program With First 10 Projects
The Chinese Academy of Sciences launched its Frontier Science Breakthrough Program in Beijing on September 29, with the first 10 projects focused on the "four extremes": new element synthesis, dark matter and dark energy, microscopic measurement of chemical bonds, and artificially synthesized cells. The program pairs shared major research facilities with AI platforms and pilots evaluation rules that de-emphasize paper counts and short-term output.
China's Generative AI Users Pass 700 Million as Intelligent Computing Hits 2,185 EFLOPS
A CNNIC report says China's generative AI users topped 700 million in the first half of 2026, with intelligent computing power reaching 2,185 EFLOPS, up 177% year over year. It also says the country's first fully domestic 100,000-accelerator AI supercluster is now running.
Chinese Banks and Insurers Roll Out Computing Power Financing at 2026 CIFTIS
At the 2026 China International Fair for Trade in Services, ICBC, Bank of China, Postal Savings Bank of China and PICC unveiled financing and insurance products built around AI computing power, from data center project loans to token purchase credit. The move reflects lenders adapting to AI firms whose value sits in technology, data and model usage rather than fixed assets.
Honor Partners With Alibaba on On-Device AI Agent Technology
Honor is working with Alibaba to advance on-device AI agent technology, according to a Xinhua report, with the collaboration tied to its upcoming Magic9 flagship smartphone. Article body was not provided, so details of the partnership remain unverified.
Hangzhou Expo Showcases Neutral-Atom Quantum Computer and AI Satellite Constellation
The fifth Global Digital Trade Expo in Hangzhou drew more than 2,000 exhibitors and over 50,000 trade visitors from 163 countries, with debuts spanning a full-stack neutral-atom quantum computer, AI computing satellites, an urban traffic model and industrial AI agents. Organizers framed the lineup as AI moving from demonstration toward deployment across governance, consumer devices and traditional industry.
Chinese Listed Firms Pour Billions Into AI Data Centers, Optical Interconnect and Chip Funds
A batch of September 24 evening filings from Shanghai- and Shenzhen-listed firms shows capital flowing into AI data centers, optical interconnect capacity, servers and semiconductor funds. Sinnet added up to 1.265 billion yuan to its Horinger AI data center, Hengtong Optic-Electric planned a 6.636 billion yuan placement, and China Life committed up to 4.5 billion yuan to an AI and chip equity fund.
WorldSkills Shanghai: Skilled Workers Say Adaptability Still Beats AI
At the 48th WorldSkills Competition in Shanghai, 1,385 competitors from 68 countries and regions are competing across 64 skill categories as AI reshapes vocational work. Competitors and delegation heads argue human adaptability, judgment and care work remain hard to automate, with lifelong learning the key skill going forward.
Five Editions In: How AI Moved From Exhibit to Infrastructure at China's Digital Trade Expo
China's only national-level digital trade fair, the Zhejiang-based Global Digital Trade Expo, has completed five editions, drawing more than 5,000 exhibitors and 87,000 professional buyers. AI now runs across the whole event, with agents, multimodal tools and industrial platforms moving from showcase items into production lines and smaller manufacturers.
Changsha UWB Chipmaker Chixin Closes A+ Round Led by Yanpu Precision
Changsha-based Chixin Semiconductor has closed an A+ round of more than 200 million yuan led by Yanpu Precision, with Quectel, Hunan Jinxia Shengshi and Changsha Lushan Investment also joining. The company will use the proceeds to push automotive-grade UWB chips into production and expand its physical AI spatial sensing work.
Qwen-Audio-3.1 Adds Five Models Spanning Audio Understanding, Generation and Interaction
Alibaba's Qwen team released the Qwen-Audio-3.1 series on September 23, adding five models covering speech recognition, synthesis, real-time interaction and audio creation. Prices were cut across the board, with ASR down as much as 95%.
T-Head Reveals Zhenwu V900 Chip and Panjiu Super-Node, Both Due in 2027
Alibaba's T-Head has unveiled the Zhenwu V900, a unified training-and-inference AI chip delivering triple the performance of the Zhenwu M890, with 216GB of memory, 1,200GB/s chip-to-chip interconnect, and FP8/FP4 support. Mass production is slated for the first quarter of 2027 alongside the new Panjiu super-node server.
Tencent Hunyuan Launches Hy Image3.5 Preview With Multi-Turn Creation
Tencent's Hunyuan team has launched the Hy Image3.5 preview model for professional-grade image generation and editing, supporting text-to-image, image-to-image, and multi-turn conversational creation. The model is already live across Yuanbao and several partner apps, with API access on Tencent Cloud's TokenHub.
Xiaomi Open-Sources MiMo-V2.6, Scaling Reinforcement Learning for Model Self-Improvement
Xiaomi has released and open-sourced its MiMo-V2.6 series, two full-modality models built around large-scale reinforcement learning to explore recursive self-improvement. The company says MiMo-V2.6-Pro ranks among today's stronger open-weight models while keeping prices unchanged.
Scinetics Closes Nearly 50 Million Yuan First Funding Round for AI4S Foundation Model
Shanghai-based AI-for-science startup Scinetics has raised nearly 50 million yuan in a first round led by Innoangel Sci-Tech Fund, with Yijing Capital, Unity Ventures and Linge Capital joining. The money will fund its AI4S foundation model, an automated experiment platform and core-team hiring.
Qwen Open-Sources Qwen-Image-2.1, a 7B Unified Image Generation and Editing Model
Qwen has open-sourced Qwen-Image-2.1, a 7B-parameter model that unifies text-to-image generation and image editing, with native RGBA transparent output, up to 10 reference images, and mask-based local editing. The release also adds KV cache reuse for inference efficiency and targets portrait and product consistency.
Huawei Upgrades Xinghe AI Data Center Network Solution at HUAWEI CONNECT 2026
At HUAWEI CONNECT 2026, Huawei announced a comprehensive upgrade to its Xinghe AI Data Center Network solution, positioning it as a high-efficiency production network for AI compute centers. No article body was provided, so specifics of the upgrade remain unconfirmed.
Xili Photonics Raises Pre-A Round to Push Thin-Film Lithium Niobate Chips Toward Mass Production
Hangzhou-based Xili Photonics has closed a Pre-A round led by Fortune Capital, reportedly in the hundreds of millions of yuan, to fund R&D and mass-production validation of thin-film lithium niobate photonic chips. The startup, founded in January 2025, is targeting optical interconnect for AI data centers with comb light sources, modulator chips and photonic engines.