The Compute Brief · August 2026 in review

China’s AI compute buildout accelerated across data centers, chips and model platforms in August

August saw China’s AI compute sector move from model launches into larger-scale infrastructure, with national capacity figures rising, Inner Mongolia and other hubs expanding, and cloud providers emphasizing AI capex returns. Domestic chips, memory, cooling, interconnects and token-based pricing all featured prominently as model usage and inference demand grew.

• Aug 1-4: DeepSeek widened access to V4-Flash after its API entered public beta, with xFusion and the National Supercomputing Internet adding access; OpenRouter later said DeepSeek-V4-Flash ranked first globally for the week with 7.22 trillion tokens, while DeepSeek reported and resolved an Aug. 4 API performance degradation during heavy traffic.

• Aug 2-5: Model access infrastructure expanded as the Yangtze River Delta Token Operations Center launched in Jiaxing with unified API access to more than 100 models, while Kimi K3 request volumes reportedly saturated its cluster within 48 hours and Kimi suspended new consumer membership subscriptions.

• Aug 3-15: Alibaba accelerated Qwen releases, launching Qwen3.8, opening API access, later opening weights for Qwen3.8-2.4T-A95B, open-sourcing Qwen3.8-27B, and reporting that the Qwen open-source model family surpassed 3 billion cumulative global downloads over six months.

• Aug 4-10: Large data-center plans and cloud capacity buildout remained central: Moonshot AI reportedly leased 20,000 Nvidia GPUs from Alibaba Cloud, RedNote was reported to be considering a 600MW Inner Mongolia campus, DeepSeek a 1GW campus in Ulanqab, and Alibaba Cloud reportedly planned to more than triple global modular data-center capacity after cutting large AIDC delivery cycles to 100 days.

• Aug 5-18: Major cloud and carrier operators disclosed AI compute scale and capex signals, including China Unicom saying its AI computing capacity exceeded 45 EFLOPS, Tencent saying GPU resources should become more ample from late 2026 to early 2027, and Baidu reporting GPU cloud revenue in its AI cloud infrastructure rose 283% year on year.

• Aug 8-9: Inner Mongolia’s Envision Galaxy Base began operations, with a planned 2GW park, 120,000 square meters of building area, support for up to 1 million AI accelerator chips and more than 1 million PFLOPS when fully built; separately, China’s first fully domestic 100,000-card AI supercluster entered operation at the Zhengzhou core node of the National Supercomputing Internet.

• Aug 12-17: Power and cooling constraints became a recurring theme: a CCTV-cited report said a 10,000-GPU cluster consumes at least 200,000 kWh per day, China’s computing-power electricity use was estimated to reach 800 billion kWh by 2030, and TrendForce said liquid-cooling penetration in AI chips is expected to rise from about 33% in 2025 to 53% in 2026.

• Aug 13-23: DeepSeek shifted monetization and tooling, announcing V4-Pro across app, web and API, releasing a developer preview of the Harness code-agent framework, introducing peak-valley API pricing effective Aug. 17, and then billing weekend API calls at off-peak rates all day from Aug. 23.

• Aug 14-21: Domestic chip and semiconductor updates broadened, with SiEngine saying its 7nm Tiangong 100 automotive-grade AI accelerator entered full mass production, SMIC reporting Q2 revenue of $3.006 billion and net profit of $479 million, Qianhe Yibang saying its four-layer 3D DRAM-stacked compute-in-memory chip taped out and powered on, and Zhongcheng Hualong releasing its HL200 inference chip and supernode cluster plan.

• Aug 17-31: Memory supply-chain progress accelerated: POWEV launched a 48GB DDR5-5600 RDIMM based on domestic DRAM dies, ChangXin Technology said its LPDDR6 reached 12,800 Mbps and 16 GB maximum capacity, Xiaomi said the Xiaomi 18 Fold would be the first device to ship with ChangXin LPDDR6 memory in September, and CXMT reportedly began risk trial production of HBM3E.

• Aug 18-26: Chinese model companies continued to adapt to domestic compute, with Zhipu deploying more than 50,000 domestic computing-power chips and later launching GLM-5.3-Flash, MiniMax adapting M3 and H3 to domestic chips and preparing a large domestic cluster, and Moore Threads providing Day-0 support for GLM-5.3-Flash training and inference.

• Aug 20-25: Alibaba emphasized AI cloud economics and self-developed chips, reporting 45% external commercial revenue growth for Alibaba Cloud in fiscal Q1 2027, saying AI computing-power capex could be recouped within three years, describing T-Head’s second-generation domestic chip as expected to begin tape-out and output in the second half, and saying it would rely more on self-developed chips to improve margins.

• Aug 22-29: Green and regional compute policy intensified, with Hohhot and Ulanqab signing 13 green computing power projects totaling 136.1 billion yuan, MIIT naming 17 regional nodes for the national computing power interconnection system, and the National Data Administration saying China’s total AI computing power reached 2.45 million PFLOPS FP16 by the end of July.

• Aug 24-31: Domestic AI hardware financing and listings advanced, including Enflame planning and then pricing a STAR Market IPO expected to raise 6.119 billion yuan, Wuxi Muxi raising several hundred million yuan for 400G/800G smart NICs and GPU interconnect chips, Hubei X-Chip raising nearly 5 billion yuan for 3D advanced packaging, and Longsys filing for a Hong Kong listing seeking up to HK$6.28 billion.

• Aug 25-31: Enterprise and agent products became a major deployment channel, with ByteDance launching Doubao for Work integrated with Feishu, Tencent Hunyuan saying Hy4 preview traffic surged after WorkBuddy integration and forced inference-cluster expansion, Alibaba Qwen launching Agent Teams in Qwen Creative with Wan 3.0, and MIIT saying it will increase procurement of large models, agents and token-based services.