NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

Cambricon Posts Its First Annual Profit — and a Test for China's AI Chip Bets

The once-loss-making AI chip (AI芯片) designer turned a 2025 net profit of about ¥2.06 billion as domestic substitution (国产替代) demand arrived.

2026-11-15 · 854 words · NeuroAI
Cambricon Posts Its First Annual Profit — and a Test for China's AI Chip Bets

For five years, Cambricon (寒武纪) was the poster child of China's expensive AI silicon ambitions: a company whose chips the market loved to talk about and whose income statement the market loved to worry about. That story changed on March 12, when the designer filed a 2025 annual report showing its first full-year profit since listing in 2020.

The numbers were not subtle. Revenue reached ¥6.497 billion (≈ US$906 million), up 453% from a year earlier. Net profit attributable to shareholders came to ¥2.059 billion (≈ US$287 million), against a loss of ¥452 million in 2024. The company even proposed a dividend — ¥15 per ten shares, plus a share conversion — the kind of gesture loss-making peers cannot make.

From a long losing streak to black ink

The turnaround is sharper in context. Cambricon lost money every year from 2022 through 2024: ¥1.256 billion, ¥848 million and ¥452 million respectively. Each year the losses narrowed, but the 2025 result is the first time the operating business itself cleared the line.

The momentum carried into 2026. In its first-quarter report, Cambricon posted revenue of ¥2.885 billion (≈ US$403 million), up 159.6% year on year, and net profit of ¥1.013 billion (≈ US$141 million), up 185%. Operating cash flow swung to a positive ¥834 million from a negative ¥1.399 billion a year earlier — a sign that growth is beginning to show up in cash, not just accruals.

Where the money came from

Cambricon is, for now, almost entirely a cloud-story company. Its cloud-product line accounted for more than 99% of 2025 revenue — about ¥6.476 billion, up 455%. Edge products and IP licensing were rounding errors by comparison. Its cloud chips, sold under the Siyuan (思元) brand, span both training and inference cards and are the workhorses behind the revenue surge. In other words, the profit is concentrated in data-center AI chips (AI芯片) bought by operators, financial firms and internet platforms for training and inference of large models (大模型).

That concentration is the point. Cambricon's rise tracks a single macro force: domestic substitution (国产替代). As Chinese buyers sought compute that did not depend on U.S. suppliers, a homegrown accelerator with a maturing software stack moved from pilot to deployment. Cambricon says its products passed the harsh-validation environments of major customers in telecom, finance and internet sectors during the year.

What the profit does — and does not — prove

A first profit is a milestone, but it is worth separating three claims the result supports from one it does not.

  • It proves demand. The 453% revenue jump is not a subsidy line item; it is shipments into real data centers.
  • It proves scale economics. Gross margin held at 55.15% even as volume surged, and R&D investment of ¥1.169 billion (≈ US$163 million) landed at a manageable 17.99% of revenue.
  • It does not yet prove technological leadership. Cambricon's edge today is being the available, supported option inside China's procurement walls — a position shaped by policy and supply, not by beating Nvidia on a global benchmark.

According to Li An, chief scientist at BrainNet (脑机网), a single profitable year confirms that the demand side of domestic substitution (国产替代) has arrived, but it does not yet confirm that the underlying technology leads.

A sector at the tipping point

Cambricon is not alone at the inflection. Fellow STAR Market (科创板) AI-chip names moved together in 2025: Hygon (海光信息), whose DCU line serves the same data-center training market, reported 2025 revenue of about ¥14.38 billion (≈ US$2.0 billion) and net profit near ¥2.54 billion (≈ US$354 million), up 32%. Moore Threads (摩尔线程) and Muxi (沐曦) stayed in the red but narrowed losses sharply as sales more than doubled. The pattern — one designer turning profitable while peers slash losses — suggests domestic substitution (国产替代) is shifting from subsidy-dependent R&D to real orders, not that any single firm has lapped Nvidia.

Honest limitations

  • Revenue quality carries caveats. The company's full-year operating cash flow was still negative in 2025, and filings flag customer concentration and a rising inventory build as risks worth watching.
  • The profit rests on a narrow base. With cloud products at over 99% of sales, any slowdown in domestic data-center spending or a single lost major customer would hit hard.
  • 2026 quarterlies look strong, but a first-quarter profit of ¥1.013 billion annualizes to a level that assumes the ramp continues uninterrupted — which it may not.
  • We have not independently verified end-customer identities or the exact share of revenue tied to government or state-owned procurement, both of which would matter for durability.
  • Software maturity against Nvidia's CUDA remains unproven in open, third-party benchmarks.

What readers can do now

  • Watch the next two quarterly reports for operating cash flow turning sustainably positive, not just net profit — that separates real traction from accounting timing.
  • Track whether Cambricon's inference software platform keeps pace with new large models (大模型) through fast "day-zero" adaptation, since ecosystem stickiness is the real moat.
  • Treat "first profit" as a demand signal, not a quality verdict: compare gross margin, R&D intensity and cash conversion against peers before reading it as a winner.

Related coverage

Cite this article

NeuroAI (脑机网), "Cambricon Posts Its First Annual Profit — and a Test for China's AI Chip Bets," 15 November 2026, https://neuroai.site/a/na-chips-cambricon-first-profit

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Chips & Compute” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories