The Compute Brief · July 2026 in review

China’s July AI Compute Push Centered on 1GW Data Centers, Kimi K3 and Domestic Supply Chains

July was dominated by large AI compute infrastructure plans, domestic-chip deployments, and heavy market attention on semiconductors, optical modules and power systems. Moonshot AI’s Kimi K3 became a major thread, driving cloud and chip adaptations, financing reports, and overseas debate over access to Chinese open models.

• Jul 21-22: Zhipu and Z.ai were reported to be advancing large domestic-chip data centers, including partial operations at a 1GW facility using Chinese-made chips and reports that Zhipu had launched or completed a 1GW-level domestic computing power center.

• Jul 21-22: Chinese semiconductor and chip-equipment stocks rallied sharply, with the ChiNext Index up more than 6% and STAR 50 up nearly 8% on Jul 21; chip ETFs including Fullgoal STAR Market Chip ETF saw strong gains and inflows.

• Jul 21-24: Local and corporate compute projects expanded, including Wuxi and Sugon’s agreement to jointly build the National Supercomputing Center in Wuxi, Pingzhi Information becoming first-ranked candidate for Chengdu Data Group’s 300 million yuan Tianfu AI Data Center Phase IV procurement, and ITC signing a memorandum for a 1GW data center near Shanghai.

• Jul 24-25: Domestic memory became a major supply-chain focus after ByteDance reportedly signed a five-year deal with CXMT worth more than $7 billion, CXMT was said to be building Shanghai and Hefei plants to double output to more than 600,000 wafers a month, and Apple reportedly lobbied U.S. officials to use CXMT and YMTC memory chips.

• Jul 24-26: Power and energy policy moved closer to AI infrastructure, with Huawei saying transport capacity is a bigger AI compute bottleneck than single-chip process nodes, the National Energy Administration calling for stronger summer energy supply oversight, and MIIT launching a zero-carbon factory program that includes zero-carbon computing power facilities with at least 95% non-fossil energy consumption by 2030.

• Jul 25-28: Chinese open-source models became a U.S. policy issue, with nearly 200 Silicon Valley companies urging the White House not to cut off U.S. access to Chinese models such as Kimi K3 and Qwen 3.8, while NVIDIA’s Jensen Huang and Elon Musk publicly opposed restrictions on Chinese open-source models.

• Jul 26-29: Kimi K3 emerged as the month’s highest-profile model release: Moonshot AI open-sourced the 2.8 trillion-parameter model with native visual understanding, a 1 million-token context window, model weights, technical report, and training infrastructure technologies including MoonEP, FlashKDA and AgentEnv.

• Jul 27-30: Domestic and global platforms moved to support Kimi K3, with Alibaba Cloud saying its Zhenwu M890 supernode supported Kimi K3 from day zero, Huawei saying Ascend completed day-zero training and inference adaptation, Moore Threads adding support on its MTT S5000 and MUSA stack, and Microsoft saying users can deploy Kimi K3 on Microsoft Foundry through Fireworks AI.

• Jul 27-31: Policy agencies pushed AI infrastructure construction, with SASAC calling for central SOEs to deepen the “AI+” campaign and accelerate computing-power infrastructure, Hangzhou drafting AI industry rules covering compute networks and power grids, and the NDRC saying China’s AI computing power at end-June was 2.8 times a year earlier.

• Jul 28-29: Large compute-service contracts and financing appeared, including Xingyun Technology raising a customer agreement to 3.0528 billion yuan, Yitian Intelligent’s Gansu Yisuan signing a 1.106436 billion yuan 60-month compute resource contract, and Entive Smart’s Gansu Yisuan planning up to 2 billion yuan of server and supporting equipment purchases.

• Jul 29-31: Major data-center and green-power projects advanced, including China State Construction winning a 2.59 billion yuan Shenzhen City AI Data Center project in Xinjiang, Gansu Energy receiving approval for a 1GW new-energy project tied to the Qingyang East Data West Computing Industrial Park, and DeepSeek reportedly planning a 1GW AI data center in Ulanqab with partial compute expected online by late 2027 or early 2028.

• Jul 29-31: Domestic semiconductor tools and chip projects progressed, with Yuwei Semiconductor shipping its Raptor-600 mask pattern defect inspection system covering 90nm to 28nm nodes, a state-backed company reportedly developing DUV lithography tools, Montage Technology starting trial production of a CXL 3.2 memory expansion controller, and Huawei open-sourcing its 505B-parameter openPangu-2.0-Pro model trained on Ascend NPUs.

• Jul 21-31: Optical communications remained a key AI compute supply-chain beneficiary, with Eoptolink and TFC Communication forecasting strong first-half growth, Lianxun Instruments guiding for 802% to 926% net profit growth, Eoptolink listing in Hong Kong after raising about HK$53.41 billion, and Mingpu Opto-Magnetic planning to raise up to 1.283 billion yuan for optical module, device, component and chip manufacturing projects.