China’s AI Compute Push Broadens Across Chips, Clusters and Models

September saw China expand domestic AI compute capacity through large cluster plans, data-center investment, financing tools and coordinated national infrastructure. Domestic chips and packaging gained market momentum, while model providers pushed lower-cost inference, agent deployments, on-device systems and broader enterprise access.

• Sep 1-5: Domestic compute procurement and financing accelerated. Fanshi Intelligence reportedly ordered more than 1 billion yuan of Huawei Ascend 950 chips, Beijing launched nearly 2 billion yuan of token-based computing-power loans, and MIIT introduced a program targeting more than 10,000 AI-oriented SMEs with pooled access to lower-cost compute.

• Sep 3-11: Domestic accelerator companies moved deeper into public markets. Enflame priced its STAR Market IPO at 142.18 yuan per share and reached a reported 170 billion yuan market capitalization on debut, while Innocn's debut valued the GPU startup at 177.7 billion yuan. Moore Threads also featured in plans for a 100,000-card domestic cluster with JD Cloud.

• Sep 4-19: Policy and infrastructure planning emphasized scale, coordination and efficiency. China issued a 2026-2030 digital-green transition plan covering renewable power, liquid cooling and waste-heat recovery; the national computing-power platform achieved coordination across all 31 provincial-level regions; and the Ulanqab Vision Galaxy Base began operations with capacity for 1 million chips and direct renewable-power connections.

• Sep 8-10: DeepSeek lowered Flash-series API prices and released V4.1 Flash, a 552-billion-parameter MoE model with 8 billion active input parameters and 16 billion active output parameters. The model became available through the National Supercomputing Internet Center, while DeepSeek consolidated service modes into a unified Intelligent mode and said it would maintain V4 Pro API access beyond Sept. 14.

• Sep 9-11: Plans for very large domestic clusters and DeepSeek financing underscored continued infrastructure demand. JD Cloud and Moore Threads announced a planned 100,000-card domestic AI data-center cluster, and reports said DeepSeek hired CITIC Securities to prepare a STAR Market IPO intended to fund compute infrastructure, model development, chip research and talent incentives.

• Sep 10-17: Governments and operators expanded local compute adoption. Zhipu, Hangzhou and Shangcheng District launched a citywide coding-plan subsidy for model API use, while Shanghai said its state-owned companies were moving from supplying compute toward networking, core technology and real-world AI applications. Zhipu's GLM team also described an Infra Agent optimizing inference infrastructure on a cluster with more than 100,000 domestic chips.

• Sep 11-20: Procurement and certification shifted further toward domestic hardware. China's central-government laptop procurement awarded all 10 lots to platforms from Zhaoxin, Phytium, Loongson, Hygon and Kunpeng, with Intel and AMD absent. Loongson said its 9A1000 GPU and AI accelerator should go on sale in the first half of 2027, while domestic CPUs from Loongson and Huawei passed Security and Reliability Level III certification.

• Sep 15-20: Data-center construction drove demand for servers, power and memory. Xingyun Technology raised its planned 2026 server and parts procurement ceiling to 20 billion yuan, Lenovo ranked first worldwide in x86 server shipments in the second quarter, and Weichai Power reported more than 1,400 data-center diesel generators sold in the first half of 2026. ChangXin Memory Technologies also entered full mass production of a 24GB LPDDR5X product for flagship smartphones.

• Sep 19-30: Memory, packaging and networking investments expanded. ChangXin Technology announced 24.1 billion yuan for technology R&D and 10.8 billion yuan for a phase-two DRAM back-end testing base; Hubei Xingchen brought China's first 3D advanced-packaging mass-production line online after more than 7 billion yuan of combined project investment; and companies reported tight supply and rising prices for DRAM, HBM, NAND and AI-server MLCCs.

• Sep 20-30: Model developers increasingly adapted software to domestic accelerators and heterogeneous systems. China Telecom released a fully domestically trained Xing4.0 model, Huawei said more than 40 models had been pretrained on Ascend, and DeepSeek open-sourced TileLang, compute libraries and distributed communication libraries for Ascend while working toward a 128-card Ascend 950 supernode. A domestic GPU and neuromorphic-chip inference system also reported more than twice the cost efficiency of comparable domestic GPU clusters in DeepSeek V4 tests.

• Sep 16-30: On-device AI became a major hardware and software theme. Nubia launched the NaviX Ultra with system-level Doubao Phone Assistant, Xiaomi introduced its Xuanjie O3 SoC and MiMo on-device model in the 18 Fold, and local AI PCs from Houmo AI, Guoke Yidao, ACEMAGIC and GMKtec targeted local execution of models ranging up to 320 billion parameters.

• Sep 17-30: Enterprise and consumer agents broadened beyond coding. Moonshot deployed a financial solution at institutions including ICBC and CSC Financial, ByteDance expanded Doubao and Feishu enterprise investment, and Doubao added or tested travel, hotel, flight, train, ride-hailing and navigation functions. Meituan also opened API and web access to LongCat-2.5-Preview, a 1.6-trillion-parameter model for long-horizon and multimodal tasks.

• Sep 23-30: Cooling, interconnect and advanced substrates became central supply-chain priorities. AAC Technologies and Intel formed a liquid-cooling partnership for high-density servers and racks, Sichuan Run shifted fundraising toward liquid cooling, H3C launched a supernode supporting up to 40 domestic accelerator cards, Mentech began small-batch shipments of 800G LPO optical modules, and BOE delivered large glass-based package substrates for customer validation with server CPUs and GPUs.