China's domestic AI chip sector published its third-quarter numbers, and the shape of the data is more informative than any single figure.
For the first nine months of 2026, Hygon reported revenue of 14.955 billion yuan, up 57.58%, with net profit attributable to shareholders of 2.847 billion yuan, up 45.12%. Cambricon reported revenue of 9.526 billion yuan, up 106.76%, and net profit of 3.715 billion yuan, up 131.49%. Moore Threads and MetaX posted revenue of 3.289 billion and 2.824 billion yuan, growing 319.15% and 128.45% respectively, both swinging to profit.
Across all four, profit grew faster than revenue and revenue grew faster than the market. That combination usually means improving utilization and a richer product mix, not just more units.
Two things changed on the demand side
The first is supply constraint. As access to high-end accelerators became harder, domestic parts moved from the "backup" column to the primary selection in procurement.
The second matters more for anyone who has actually tried to port a model: the software gap closed. The historical objection to domestic chips was never peak FLOPS, it was that models would not run: missing operators, incompatible frameworks, thin debugging tools. This year, when DeepSeek and Zhipu released new generations, Cambricon and MetaX announced same-day adaptation. The industry calls it Day0. Compressing a months-long integration into days changes procurement decisions more than any benchmark score.
Capital and supply chain
The third phase of the National Integrated Circuit Industry Investment Fund registered 34.4 billion yuan in capital and, by incomplete counts, has invested in more than ten STAR Market companies covering foundry capacity, EDA tooling and packaging equipment. On the supply side, mature-node fabrication, packaging and test, etch and deposition equipment, standard materials and server manufacturing can now be delivered domestically. AI chips have a closed loop from design to foundry to end customer.
Some analysts expect domestic AI accelerators to be supply-constrained by 2028 as demand spreads from government IT into finance, energy and edge inference.
The model-side flywheel
Chip value ultimately shows up as token volume. Zhipu's GLM-5.3 Flash passed 20 trillion tokens per day in September, and the company raised its year-end ARR guidance from $2.4 billion to $3 billion. Higher volume is what allows unit compute cost to fall and what justifies fab capacity expansions.
Cheaper models lead to more applications, which raises token volume, which raises chip orders, which grows cluster size, which lowers unit cost. Whether that loop keeps turning depends on utilization, not on announcements.
Four notes for buyers
Look at effective utilization, not peak throughput. A thousand cards that train stably for weeks beats a higher nameplate number that idles.
Require a report on your target model, not a generic benchmark: is it adapted, what precision loss, how many code changes.
Validate on 8 or 32 cards before committing to a cluster. Two weeks of real workload beats a hundred pages of whitepaper.
Price the exit. Operator coverage, framework release cadence and community activity determine whether you can still use this stack in three years.
China's AI Chip Makers Grew 58% to 319% This Year. Here Is What Changed
← Previous: Alibaba's Qwen4-Max Preview: 1M Context, 393K Output, No Spec Sheet
Next: →Embodied AI's Funding Week: UniX AI's $70M and DeepSeek's Two Bets →