--- title: "Cambricon books ¥3.7B profit as inventory bet balloons to ¥11.7B" date: 2026-11-04 category: Chips & Compute site: NeuroAI canonical: https://neuroai.site/a/na-chips-cambricon-profit-turn language: en --- # Cambricon books ¥3.7B profit as inventory bet balloons to ¥11.7B > China's leading domestic AI accelerator designer swung to a nine-month net profit of ¥3.7B, and its component stockpile now exceeds ¥11.7B — a bet on demand it still refuses to quantify in unit terms. A company that lost money for years just told the market it is now profitable by the billion. The surprise is not the profit itself — it is what the company is doing with the cash, and what it will not say out loud. Cambricon (寒武纪), the Shanghai-listed designer of the 思元 (Siyuan) line of AI accelerators, has become the clearest read on whether China's push for self-made compute is turning into real business. The latest quarterly numbers say yes. The silence around shipments says something too. ## Key takeaways - Nine-month 2026 revenue reached ¥9.53B (≈ US$1.33B / HK$10.4B), up 106.8% year on year, with net profit of ¥3.71B (≈ US$519M / HK$4.05B), up 131.5%. - The company is sitting on ¥11.66B (≈ US$1.63B) of inventory and ¥3.04B of prepayments, up sharply, as it locks in wafers, HBM and packaging capacity. - Management explicitly denied having guided any "over one million units shipped in 2027" figure — a reminder that unit volumes here are estimates, not disclosures. ## The numbers behind the turn For the third quarter alone, Cambricon reported revenue of ¥3.53B (≈ US$494M), up 104.5% from a year earlier, and net profit attributable to shareholders of ¥1.40B (≈ US$196M), up 147.7%. The first nine months together produced revenue of ¥9.53B and net profit of ¥3.71B, with a gross margin of 56.5% — a level that would be the envy of most chip designers anywhere. Research spending stayed high: ¥522.6M in the quarter (up 73.7%) and ¥1.23B for the nine months (up 45.4%). That is the right posture for a company whose moat depends on software as much as silicon. The 思元 family only matters if developers actually port their large model (大模型) workloads onto it. ## Reading the inventory pile The most telling line on the balance sheet is not profit. It is inventory. Cambricon's stockpile of components and work-in-progress climbed to ¥11.66B at the end of September, from ¥8.25B at mid-year — a jump of roughly 41% in a single quarter. Prepayments to suppliers reached ¥3.04B, and notes payable swelled to ¥4.97B. In plain terms: the company is paying upstream partners early and in size to secure wafers, high-bandwidth memory and advanced packaging. This is the classic signature of a fabless chip designer that expects demand to outrun supply. It is also a risk. The company disclosed it took about ¥502M of inventory impairment provisions in the first nine months, because some raw-material stock has aged. If orders slow, that pile cuts both ways. ## The question it won't answer At the October 26 earnings call, chairman and general manager Chen Tian石 (陈天石) fielded investor questions about how many chips the company might ship in 2027. His answer was notable for what it excluded: he said the company had never communicated any "over one million units in 2027" figure, and pointed investors back to official filings. That is the honest texture of this whole sector. Huawei, Cambricon, Biren (壁仞) and others operate in a market where third-party trackers publish unit estimates and investors repeat them, but the companies themselves disclose revenue and margins, not chip counts. Anyone quoting a precise shipment number for a Chinese AI accelerator is almost certainly quoting a forecast, not a fact. ## The competitive lane A year ago, the bear case on Cambricon was simple: it burned cash, leaned on a few customers, and faced entrenched foreign competition. The profit swing undercuts the first point. Customer concentration remains live — the company does not break out customer names, and much domestic accelerator demand flows through a handful of hyperscalers and state-linked compute projects. Peer Hygon (海光信息), which sells CPU and DCU parts, reported nine-month 2026 revenue of ¥14.96B (≈ US$2.09B) and net profit of ¥2.85B (≈ US$398M), a reminder that Cambricon is not the only domestic name scaling fast. The rise of a second profitable domestic accelerator vendor matters more than any single quarter, because it shows the demand pool is real and widening. ## Why the software still decides Cambricon's 思元 chips compete on a roadmap that trades raw per-chip performance for system-level availability. Like Huawei's Ascend, the long game is a home-grown software stack — compilers, operators and frameworks — that lets a large model (大模型) train and infer without Nvidia's CUDA. That stack is where the real moat, and the real risk, lives. Revenue proves the hardware is shipping; only developer adoption proves the platform will last. ## Honest limitations This article relies on Cambricon's official third-quarter report filed to the Shanghai Stock Exchange and contemporaneous coverage from Shanghai Securities News and Xinhua Finance. It does not include any verified unit-shipment figure for any 思元 product; none was disclosed. Inventory, prepayment and impairment numbers come from the company's balance sheet. RMB figures are converted at an approximate rate of 7.15 to the US dollar and 1.09 to the Hong Kong dollar for reader convenience only, and are rounded. Statements about 2027 output are the company's own denials of market rumours, not projections endorsed here. ## What readers can do now - Treat "X million chips shipped" headlines about Chinese AI accelerators as estimates; check whether the number is sourced to the company or to a tracker before citing it. - If you follow semiconductors, watch Cambricon's inventory-to-revenue ratio quarter over quarter — a falling ratio signals real sell-through, a rising one signals a bet that could sour. - For the wider domestic compute picture, read foundry and memory capacity data alongside single-company results rather than in isolation. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-chips-cambricon-profit-turn Free to quote with attribution and a link to the original.