--- title: "Cambricon (寒武纪) turns profitable: Chen Tianshi's (陈天石) seven-year bet on domestic AI chips pays off" date: 2026-08-29 category: Chips & Compute site: NeuroAI canonical: https://neuroai.site/a/na-cambricon-turns-profitable language: en --- # Cambricon (寒武纪) turns profitable: Chen Tianshi's (陈天石) seven-year bet on domestic AI chips pays off > Cambricon (寒武纪) posted its first profitable half-year: 5.085 billion yuan (about $706 million / HK$5.53 billion) in revenue, up 3,136% year on year, with 2.232 billion yuan (about $310 million / HK$2.43 billion) net profit. A recent earnings report has, by Chinese accounts, quieted an entire sector. In Kunshan, Jiangsu province, Cambricon (寒武纪) disclosed its 2026 interim results: revenue of 5.085 billion yuan (about $706 million / HK$5.53 billion), up 3,136% year on year, and net profit attributable to shareholders of 2.232 billion yuan (about $310 million / HK$2.43 billion)—the company's first profitable half-year. At one point its market capitalization touched 450 billion yuan (about $62.5 billion / HK$489 billion). For context, between 2023 and 2024 the domestic AI-chip track was stuck in "losses, halved valuations, layoffs." From the bottom to this high, Cambricon took seven years. ## Key takeaways - First-ever profitable half-year: **2.232 billion yuan** (about **$310 million / HK$2.43 billion**) net profit. - Revenue **5.085 billion yuan** (about **$706 million / HK$5.53 billion**), **+3,136%** YoY. - Single internet-giant orders of **500–1,000 million yuan** (about **$69–139 million / HK$543 million–1.09 billion**). - Siyuan 690 trains about **25% slower** than Nvidia H100 but costs **60% less** with half the lead time. - Monthly shipments stabilized at **80,000–120,000** chips. ## An "atypical" path to profit Looking back, Cambricon did not follow Nvidia's route of building high-end GPU alternatives. Its path was domestic substitution plus compliant markets—a "crack" strategy. In the first half of 2026, its large orders came from three customer types: - Internet giants quietly shifting part of their inference workloads from Nvidia's H-series to Cambricon's Siyuan (思元) 590/690, in single orders of 500–1,000 million yuan (about $69–139 million / HK$543 million–1.09 billion). - Central state-owned enterprises (央企国企) under "Xinchuang" (信创, the domestic-technology procurement mandate) in finance, telecom, and energy, where localization is effectively compulsory. - Industry AI firms placing small-batch orders in autonomous driving, biomedicine, and robotics. This approach sidesteps a head-on clash with Nvidia and instead plants roots in the compliant, industrial, and local markets where Nvidia is unwilling or unable to sell. The trade-off: Siyuan 690 is about 25% slower than Nvidia's H100 on large-model training. The upside is equally clear—60% cheaper than H100, half the delivery lead time, and fully compliant to sell. ## Domestic substitution enters its "second half" Behind the profit is a deeper signal: Chinese AI chips have formally entered their "second half." The first half (2018–2024) was "PPT chip-making"—loud but with little real shipment. The second half, from 2025, shows three traits at once: it can run large models (Siyuan 690 has completed training on several hundred-billion-parameter models), it can mass-produce (monthly shipments of 80,000–120,000, with monthly revenue rising from 600 million yuan a year earlier to 5 billion yuan in this half-year), and it is compliant (fully self-controlled, on the Xinchuang procurement list). That reset valuations across the board. Hygon Information's (海光信息) DCU, Loongson (龙芯中科), Enflame (燧原科技), Moore Threads (摩尔线程), and Biren (壁仞科技) all reported commercial progress, and the market re-identified the substitution theme. ## One standout, or many rivals? Cambricon's lead is not permanent. Its biggest risk is a thin product matrix: about 90% of revenue comes from the Siyuan 590/690 families, and the next-generation Siyuan 780 is still in research. If Nvidia's next products cut prices sharply or unlock more capacity, Cambricon faces heavy pressure. Customer concentration is the other worry: the top five clients make up 65% of revenue, so any single strategy change hits results. Meanwhile Hygon, Enflame, and Biren have each carved differentiated positions—Hygon's DCU leads Cambricon in finance, and Enflame won a major ByteDance inference order. ## What an ordinary observer can do The reporting offers three practical notes rather than stock tips: consider a basket via domestic AI-compute ETFs instead of a single name; watch domestic AI-server integrators like Inspur (浪潮), Sugon (中科曙光), and Unisplendour (紫光股份), whose order swings are smaller than pure-chip firms; and read quarterly reports—revenue, orders, and capacity utilization are the cleanest signals for the track. Domestic chips are not trying to beat Nvidia. They are trying to stand firm where Nvidia will not sell. Cambricon, with seven years, a 5-billion-yuan half, and 2-billion-yuan profit, has given the first complete answer. And that answer is a beginning, not an end. ## Honest limitations All figures come from Cambricon's (寒武纪) disclosed 2026 interim report as summarized by Chinese media; NeuroAI has not independently audited them. Revenue growth of +3,136% is measured against a low 2025 base and should not be read as a sustained run-rate. Performance gaps versus Nvidia H100 are vendor- or media-reported and benchmark-dependent. Customer-concentration and competitor claims are analytical observations, not verified by the companies named. This article is informational and not investment advice. ## Sources Based on reporting by Chinese state media and company disclosures; specifically Cambricon's (寒武纪) 2026 interim financial report and Chinese financial-media analysis, with references to Hygon (海光信息), Loongson (龙芯中科), Enflame (燧原科技), Moore Threads (摩尔线程), Biren (壁仞科技), Inspur (浪潮), Sugon (中科曙光), and Unisplendour (紫光股份). --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-cambricon-turns-profitable Free to quote with attribution and a link to the original.