The Compute Brief · Q3 2026 in review

China Scales AI Compute Infrastructure and Domestic Supply Chains in Q3 2026

China’s AI compute sector expanded rapidly in Q3, combining large data-center and cluster plans with stronger domestic chip, memory, packaging, cooling and networking capabilities. Model releases and rising inference demand also pushed cloud providers and governments toward lower-cost access, domestic hardware adaptation, enterprise agents and on-device AI.

• National compute capacity continued to rise: the NDRC said China’s AI computing power was 2.8 times higher at the end of June than a year earlier, while the National Data Administration put total AI computing power at 2.45 million PFLOPS FP16 by the end of July. By September, the national computing-power platform had achieved coordination across all 31 provincial-level regions.

• Large regional infrastructure projects moved forward, particularly in Inner Mongolia. Envision Galaxy Base began operations in August with a planned 2GW park capable of supporting up to 1 million AI accelerator chips and more than 1 million PFLOPS when fully built; the Ulanqab site also received direct renewable-power connections, while Hohhot and Ulanqab signed 13 green-computing projects totaling 136.1 billion yuan.

• China’s largest domestic-compute projects expanded in scale. A fully domestic 100,000-card AI supercluster entered operation at the Zhengzhou core node of the National Supercomputing Internet in August, while JD Cloud and Moore Threads announced plans for another 100,000-card domestic AI data-center cluster in September. Zhipu said it had deployed more than 50,000 domestic computing chips.

• Cloud and telecom operators disclosed growing AI capacity and investment. China Unicom said its AI computing capacity exceeded 45 EFLOPS, Baidu reported a 283% year-on-year increase in GPU cloud revenue within AI cloud infrastructure, and Alibaba Cloud reported 45% growth in external commercial revenue while saying AI-compute capital expenditure could be recouped within three years.

• Power efficiency and data-center engineering became recurring constraints. A CCTV-cited estimate put the daily electricity use of a 10,000-GPU cluster at no less than 200,000 kWh, while TrendForce expected liquid-cooling penetration in AI chips to rise from about 33% in 2025 to 53% in 2026. China’s 2026-2030 digital-green transition plan covered renewable power, liquid cooling and waste-heat recovery.

• Domestic accelerators gained visibility through deployments, public markets and procurement. Huawei Ascend chips supported day-zero training and inference for several models, Fanshi Intelligence reportedly ordered more than 1 billion yuan of Ascend 950 chips, and Enflame’s STAR Market debut followed an IPO priced at 142.18 yuan per share and a reported 170 billion yuan market capitalization. Innocn’s debut valued the GPU startup at 177.7 billion yuan.

• The domestic semiconductor supply chain broadened beyond accelerators. CXMT reportedly began risk trial production of HBM3E, ChangXin expanded LPDDR5X and LPDDR6 production, and domestic companies introduced DDR5 server memory. Advanced packaging also progressed, with Hubei Xingchen bringing China’s first 3D advanced-packaging mass-production line online after more than 7 billion yuan of combined investment.

• Supply constraints extended into networking and server components. Optical-module companies reported strong growth and fundraising, Mentech began small-batch shipments of 800G LPO optical modules, Wuxi Muxi raised funds for 400G and 800G smart NICs and GPU interconnect chips, and companies reported tight supply and rising prices for DRAM, HBM, NAND and AI-server MLCCs.

• Model releases increased pressure on compute access and inference economics. Moonshot AI open-sourced the 2.8-trillion-parameter Kimi K3 in July with a 1-million-token context window, while Alibaba expanded the Qwen3.8 family and said its open-source models had exceeded 3 billion cumulative global downloads. DeepSeek released V4-Pro and V4.1 Flash, lowered Flash-series API prices and introduced peak-valley pricing and off-peak weekend billing.

• Demand for model access exposed capacity limits and encouraged shared infrastructure. Kimi K3 reportedly saturated its cluster within 48 hours, prompting a suspension of new consumer memberships, while the Yangtze River Delta Token Operations Center launched unified API access to more than 100 models. DeepSeek-V4-Flash ranked first globally on OpenRouter for the week with 7.22 trillion tokens, according to OpenRouter.

• Model developers increasingly optimized software for domestic and heterogeneous systems. Huawei said more than 40 models had been pretrained on Ascend, China Telecom released a fully domestically trained Xing4.0 model, and DeepSeek open-sourced TileLang, compute libraries and distributed communication libraries for Ascend while working toward a 128-card Ascend 950 supernode. Zhipu’s GLM team described an inference-optimization agent operating on a cluster with more than 100,000 domestic chips.

• Enterprise agents and lower-cost application access became important deployment channels. ByteDance expanded Doubao and Feishu enterprise products, Tencent’s WorkBuddy integration drove traffic that required inference-cluster expansion, Alibaba launched Agent Teams, and Moonshot deployed a financial solution at institutions including ICBC and CSC Financial. Beijing also launched nearly 2 billion yuan of token-based computing-power loans, while MIIT introduced a program aimed at giving more than 10,000 AI-oriented SMEs pooled access to lower-cost compute.

• On-device AI broadened the market beyond data centers. Nubia launched a phone with system-level Doubao assistance, Xiaomi introduced the Xuanjie O3 SoC and MiMo on-device model in the 18 Fold, and AI-PC makers including Houmo AI, Guoke Yidao, ACEMAGIC and GMKtec targeted local execution of models ranging up to 320 billion parameters.