China's open-weight AI model race has kept intensifying through the first half of 2026, with Tencent, Zhipu's GLM series, Alibaba's Qwen family and DeepSeek all shipping rapid iterations -- a category where Western frontier labs like OpenAI and Anthropic have stayed comparatively less active, generally preferring closed, API-only deployment for their most capable models.
The model race is backed by an equally aggressive push on domestic chip supply. China's self-reliance drive is drawing an estimated $98 billion in government and corporate spending this year, with Cambricon Technologies emerging as the clearest beneficiary: the chipmaker posted its first-ever annual profit as Beijing has encouraged domestic AI developers to shift off Nvidia hardware amid US export restrictions, and its market capitalization has risen to roughly $81 billion, making it mainland China's most valuable listed company.
Cambricon is targeting roughly 500,000 AI accelerator shipments in 2026, including its Siyuan 590 and 690 processors -- though researchers have flagged low manufacturing yields and limited high-bandwidth memory supply as real constraints that could keep the company from hitting that target at the pace Beijing wants.
“The model race is backed by an equally aggressive push on domestic chip supply.”
The combination of cheaper open-weight models and a domestic chip supply chain gives Chinese AI developers a structurally different cost base than Western labs still paying Nvidia's margins and running primarily closed models -- and every efficiency gain on the model side compounds pricing pressure on closed frontier-lab APIs globally, regardless of where the end customer sits.
For investors evaluating AI-model exposure, the practical read is that "good enough and much cheaper" open-weight competition is no longer a China-only phenomenon in effect -- it directly shapes the pricing ceiling Western closed-model labs can sustain, particularly for enterprise customers running high-volume inference workloads.
What to watch: whether Cambricon can hit its 500,000-unit 2026 shipment target given yield and memory constraints, and whether the next generation of Chinese open-weight models continues closing the capability gap with Western closed models across coding and reasoning benchmarks specifically.