--- title: "Cambricon turns first profit as China's AI chip demand explodes" date: 2026-10-27 category: Chips & Compute site: NeuroAI canonical: https://neuroai.site/a/na-cambricon-first-profit-ai-chip language: en --- # Cambricon turns first profit as China's AI chip demand explodes > The Beijing accelerator maker posted its first-ever annual profit in 2025 on 453% revenue growth, a milestone for China's domestic computing-power push. For six years the nickname followed it like a shadow: "China's little NVIDIA." On March 12, 2026, Cambricon (寒武纪) finally shook it off — not by beating NVIDIA, but by proving a domestic AI-chip company can actually make money. The numbers came from its own annual report, and they were not small. ## Key takeaways - Cambricon reported 2025 revenue of RMB 6.50 billion (≈ US$915 million / HK$7.1 billion), up 453% year on year — its first full-year profit since listing. - Net profit attributable to shareholders reached RMB 2.06 billion (≈ US$290 million / HK$2.2 billion), swinging from a RMB 452 million loss in 2024. - Cloud-line accelerators supplied 99.7% of revenue; just five customers accounted for 88.7% of sales, a concentration risk the company itself flags. - The maker produced 127,700 cloud chips in 2025, up 410%, and sold 117,400, up 202% — volume, not price, drove the turnaround. ## The quarter that changed the story Cambricon's 2025 wasn't a slow burn; it was a step function. Revenue of RMB 6.497 billion (the company reports RMB 6.50 billion) came almost entirely from its cloud product line — RMB 6.476 billion, or 99.69% of the total. That concentration tells you exactly who is buying: large domestic cloud and compute operators rebuilding training and inference capacity on home-grown silicon rather than imported GPUs (图形处理器). The cost side did the rest. Revenue quintupled while R&D spend rose only 9% to RMB 1.169 billion (17.99% of revenue), meaning earlier engineering bets finally hit commercial scale. The company closed 2025 with 887 R&D staff — over 80% of its workforce — and 1,734 authorized patents. ## A concentrated, fragile boom The bull case and the warning label sit on the same page of the annual report. Cambricon's top five customers delivered 88.66% of sales (RMB 5.76 billion), and four of those five were new in 2025. Translation: a handful of large buyers executed rapid, massive deployments of Cambricon's Siyuan (思元) architecture as a substitute for restricted foreign hardware. Supply chain risk mirrors the demand risk. The top five suppliers made up 75.23% of procurement, and a single primary supplier absorbed RMB 4.198 billion — 55.34% of all purchases. For an advanced chip built on SMIC's (中芯国际) N+2 7nm process and dependent on scarce HBM (高带宽内存), that single-source exposure is structural, not incidental. ## What the chips actually are Cambricon's commercial workhorse is the Siyuan 590, which the company says matched roughly 80% of NVIDIA's A100 on relevant workloads; an updated Siyuan 690 is on the roadmap. These are inference-and-training accelerators for data centers, not consumer parts. The 2025 surge reflects domestic enterprises absorbing capacity that, two years earlier, would have gone to imported accelerators now blocked or rationed under US export controls. The financial market noticed. On March 12, 2026, Cambricon shares closed at RMB 1,099, valuing the company at roughly RMB 463 billion — second only to Hygon (海光) among STAR Market semiconductor peers. It removed the special "U" (unprofitable) marker on March 16 and declared its first dividend: RMB 15 per 10 shares (about RMB 632 million) alongside a 4.9-for-10 stock split. ## The export-control tailwind Cambricon's breakout is inseparable from US curbs. When the most advanced NVIDIA accelerators were barred from the Chinese market, large buyers did not simply wait — they qualified domestic parts. That forced adoption is why 2025 volumes look like a step function rather than a glide path. The risk is that demand is policy-driven; if restrictions ease, some buyers may drift back to imported GPUs (图形处理器). For now, though, the substitution is real and large, and it is rewiring procurement habits that took a decade to form around CUDA. ## Why this matters beyond the ticker Cambricon's profit is a data point in a larger question: can China's semiconductor independence move from subsidy-dependent promise to self-sustaining business? Rivals Moore Threads and MetaX grew revenue sharply in 2025 but had not yet reached profitability, making Cambricon the first of the "little NVIDIA" cohort to cross the line. If others follow, the domestic accelerator supply base stops being a policy project and starts being an industry. ## Honest limitations The 88.7% customer concentration means a single cancelled order would dent results hard. The 55% supplier concentration exposes Cambricon to foundry (代工) and packaging bottlenecks it does not control — notably HBM (高带宽内存) supply still dominated globally by SK Hynix and Samsung. Yield on SMIC's largest dies for these parts has been estimated near 20% by some analysts, capping margin upside. And "first profit" follows years of losses; profitability has not yet been tested across a full demand cycle. All shipment and yield figures here are drawn from the company's filings or analyst estimates, not independent teardown. ## What readers can do now - Watch Cambricon's 2026 customer concentration — if the top-five share falls, it signals broadening, healthier demand rather than one-off deployments. - Track whether Moore Threads and MetaX also reach profitability; one winner is an outlier, several is a trend. - Monitor HBM (高带宽内存) availability in China, since memory supply — not chip design — is the binding constraint on how many Siyuan parts actually ship. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-cambricon-first-profit-ai-chip Free to quote with attribution and a link to the original.