A chip designer spends years burning cash to tape out silicon that almost nobody will buy on day one.
Then a foreign rival is forced out of the market by export rules, and suddenly every domestic buyer is knocking.
That is the opening Cambricon just stepped into — and its first annual profit says the timing was right.
Key takeaways
- Cambricon reported its first-ever annual profit for 2025, announced 12 March 2026.
- Net income: about 2.1 billion yuan (≈ US$296M / HK$2.3B), versus a 452-million-yuan loss (≈ US$64M / HK$497M) in 2024.
- Revenue: 6.5 billion yuan (≈ US$915M / HK$7.1B), up from 1.2 billion a year earlier.
- 2026 plan: more than triple 2025 output to roughly 500,000 AI accelerators, including up to 300,000 Siyuan 590 and 690 units.
- The hard constraint: production relies on SMIC's N+2 (~7nm) line, with reported yields around 20% on the largest dies.
From loss-making hard-tech to "China's little Nvidia"
Cambricon, founded in 2016 by Chen Tianshi (陈天石), has long been cast as China's answer to Nvidia in the accelerator market. For most of its life it lost money. The 2025 result flips that: a net profit of about 2.1 billion yuan against a 452-million-yuan loss the year before.
Revenue tells the demand story. Full-year revenue reached 6.5 billion yuan, more than five times the 1.2 billion of 2024. The swing came as Chinese AI developers, pushed by Beijing to cut reliance on foreign accelerators and constrained by US export controls, shifted procurement toward domestic silicon.
The inference (推理) angle
Cambricon's strength has been in inference (推理) rather than training. Inference is the quieter, larger market: every search ranking, ad placement, recommendation and customer-service bot runs on it continuously. That is exactly the workload internet giants need at scale, and it rewards low latency and low power more than peak training throughput.
This is why a first profit matters beyond the headline. It suggests domestic inference chips are moving from "approved substitute" to "good enough to buy on merit" for at least some large buyers.
The 2026 bet: half a million chips
Cambricon aims to deliver around 500,000 AI accelerators in 2026 — more than triple its 2025 output — including as many as 300,000 of its most advanced Siyuan 590 and 690 parts. If it lands even close, it would mark one of the largest domestic accelerator ramp-ups attempted in China.
Bloomberg has reported the target; Morgan Stanley estimates China's AI chip market could reach US$67 billion by 2030, with more than three-quarters supplied by domestic firms such as Cambricon, Alibaba and Huawei.
The constraints nobody can ignore
Two hard limits sit between the plan and the product.
Foundry capacity. Cambricon does not make its own silicon. The Siyuan 590 and 690 depend on SMIC's N+2 process, China's most advanced stable node, built on DUV rather than EUV lithography. Reports put yields on the largest dies around 20% — four of five chips off the wafer fall short. Even with allocated capacity, effective output is set by what survives manufacturing.
Customer concentration. Cambricon's largest customer is ByteDance, which also designs its own chips. When a buyer is building its own silicon, today's orders are not a guarantee of tomorrow's.
Ecosystem is the longer fight
A chip is only as useful as the software around it. Nvidia's moat is CUDA — fifteen years of libraries and developer habit. Cambricon is building its own stack, and large buyers porting workloads is the real vote of confidence. Profit shows the hardware cleared the bar; the next test is whether developers stay.
What one profitable vendor means for the field
Cambricon is not alone. Moore Threads listed on the STAR Market in 2026, pricing shares at 114.28 yuan (≈ US$16 / HK$126), per Bloomberg-sourced reporting, and peers such as Iluvatar CoreX have crossed shipment thresholds. But Cambricon is the first of the pure-play AI chip designers to post an annual profit.
That matters because profitability — not just funding — is what lets a company keep taping out new generations when the easy money cools. It positions Cambricon as the bellwether: if domestic inference (推理) demand is real and durable, its income statement is the cleanest evidence yet.
What readers can do now
- Read Cambricon's result as a sector signal, not a single stock story. A first profit means domestic inference chips cleared a commercial bar — relevant to every company betting on China-built AI stacks.
- Track shipments, not slogans. The 500,000-unit 2026 target is the number to watch each quarter; capacity and yield will decide the real figure.
- If you build on Chinese AI infrastructure, map your inference dependency. Which workloads run on domestic chips today, and what is your fallback if a single supplier's allocation slips?
Honest limitations
Financial figures are from Cambricon's 2025 annual results announced 12 March 2026, as reported by Bloomberg, Yicai and other financial outlets; net income is cited as "about 2.1 billion yuan" (some filings state RMB 2.059 billion) and revenue as 6.5 billion yuan. The 500,000-unit 2026 target and 300,000 Siyuan figure come from Bloomberg reporting and are company guidance, not guaranteed output. Yield and SMIC N+2 details are from industry reporting (Tom's Hardware / TrendForce citing Bloomberg) and are estimates. The Moore Threads IPO price is from Bloomberg-sourced coverage. I hold no position in any company mentioned. Not investment advice.