For most of its listed life, Cambricon (寒武纪) was the poster child of a painful truth in China's semiconductor story: designing an AI chip is hard, but selling enough of them to turn a profit is harder. That changed in 2025. The Beijing-based designer of artificial intelligence (人工智能) accelerators disclosed its first annual profit since going public in 2020 — and not by a thin margin.
What the numbers actually say
Cambricon's 2025 annual report shows revenue of RMB 6.497 billion (≈ US$915 million / HK$7.1 billion), up 453% from a year earlier. Net profit attributable to shareholders reached RMB 2.059 billion (≈ US$290 million), versus a loss of RMB 452 million in 2024. Gross margin held at 55.15%, and the cloud (云端) product line — the AI accelerator cards that go into data centers — accounted for 99.98% of revenue. Research and development spending was RMB 1.169 billion (≈ US$165 million), about 18% of sales, with a research team of 887 people, over 80% holding master's degrees or above.
These are company-disclosed, audited figures from the official annual report, widely carried by Chinese financial press including Securities Times (证券时报). They are results, not a broker forecast.
Why one profitable year matters
Cambricon is not the only Chinese firm building AI accelerators (AI 加速卡). Huawei's Ascend (昇腾), Hygon's DCU (深算), Biren, and Moore Threads are all in the same race. What makes Cambricon's result worth attention is what it implies about demand. For years, "domestic substitution" (国产替代) of Nvidia (英伟达) was a policy slogan more than a purchasing reality. Export controls on high-end accelerators changed that calculus: Chinese cloud providers, state-owned telecom operators, banks, and internet firms suddenly needed alternatives they could actually buy and deploy.
Cambricon's chips — its 思元 (MLU) series — have moved from demonstration projects into scaled deployments across operators, finance, and internet sectors. The company says its software stack and cluster toolchain matured enough that customers validated the products in demanding production environments. In plain terms: the chips worked well enough that buyers kept buying.
From pilot to production
The deeper shift is in where domestic chips first win. Frontier model training still favors the most advanced hardware available, but inference (推理) — serving a model to millions of users — is where Chinese accelerators are gaining a real foothold first. Inference is less demanding than cutting-edge training (训练) and far more price-sensitive, which gives a cost-competitive domestic card a natural wedge. Cambricon's near-total reliance on its cloud line shows it is riding exactly that wave.
For ordinary readers, an AI accelerator (AI 加速卡) is the engine room of every model they use — the hardware that turns a prompt into an answer. When that engine is built and sold domestically at scale, it changes who captures the value of the AI boom and how resilient China's services are to sudden supply shocks. A profitable Cambricon is therefore less a single-company story than a marker that the domestic compute layer is becoming self-sustaining rather than subsidized.
Reading the trend
The broader signal is that China's AI compute supply chain is moving from "we have a substitute" to "the substitute is shipping at volume." According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the 2025–2026 wave shows domestic accelerators clearing the deployment threshold where cost, stability, and ecosystem support finally line up for mainstream buyers.
That does not mean Nvidia has been replaced. The most advanced training clusters in China still lean on imported hardware where it is available, and Cambricon's revenue is concentrated in a small number of large customers. But a profitable year is a new kind of proof.
Honest limitations
- Figures are company-disclosed and drawn from Cambricon's 2025 annual report and Chinese financial-media coverage; independent third-party audits of unit shipments and end-use deployment are not publicly available.
- This article does not assess Cambricon's product performance against Nvidia or Huawei on any benchmark; such comparisons require standardized, independently run tests.
- We have not verified the geographic or sector breakdown of final customers beyond what the company disclosed.
- No investment advice is intended; outcomes depend on factors outside the scope of this piece.
What readers can do now
- If you build or buy AI infrastructure in China, treat domestic accelerators as a live option worth benchmarking against your own workloads, not a future possibility.
- For technical teams, examine software compatibility first: Cambricon's ecosystem now tracks open-source frameworks, so a pilot migration of an inference workload is lower-risk than a full training swap.
- Globally, watch the "profitable domestic accelerator" milestone as a leading indicator of how fast China's compute self-reliance is becoming commercial, not just political.
- Rely on primary filings (annual reports, exchange disclosures) rather than secondary commentary when forming views on any single company.