On October 23, Cambricon Technologies reported third-quarter net profit of 1.40 billion yuan (about $210 million), up 147.7%. Revenue doubled to 3.53 billion yuan, but analysts had expected about 6.26 billion, so sales came in roughly 44% short. The profit jump comes from the company's own filing. The cause of the miss, and whether these chips are training Chinese AI models or only running them, are less clear. This is not investment advice.
What the Cambricon Q3 Filing Shows
Revenue rose 104.5% from a year earlier. Net profit was 1.40 billion yuan, up from 566.6 million. For nine months, revenue was 9.53 billion yuan and net profit 3.71 billion. Cambricon credited market growth and the spread of AI applications, and says higher revenue drove the profit. It kept its target of at least 13.5 billion yuan in revenue for 2026. By our arithmetic, that needs about 3.97 billion in the fourth quarter, 12.5% more than the third.
Why Revenue Missed
Bloomberg ties the miss to manufacturing limits and memory shortages, not weak demand. Chinese chip designers lack enough high-bandwidth memory (HBM), made mainly by Samsung, SK Hynix and Micron, which US-led rules block from selling to Chinese buyers. Local foundries such as SMIC also cannot get ASML's most advanced machines. The supply explanation comes from Bloomberg and analysts, not from Cambricon's filing.
Morgan Stanley analysts add a timing point. They wrote that Cambricon delivered its MLU692 chip to a major cloud provider for third-quarter testing, with most of that revenue expected in the fourth quarter. They also said MLU580 shipments stayed constrained because higher HBM costs hurt unit economics. So part of the miss may be timing, not only supply.
What the Balance Sheet Shows
China AI Dispatch, reading Cambricon's filings, found inventory of 11.66 billion yuan at the end of September, up 3.42 billion in the quarter, almost as much as the quarter's revenue. Most of it is raw material or material out for processing, not finished chips. The company says it is stockpiling on purpose because it expects supplies to get scarcer and dearer.
That has a cost. Cambricon booked 501.6 million yuan of inventory write-downs for nine months. Profit is also not turning into cash. Nine-month operating cash flow was 228 million yuan against 3.71 billion of profit, and negative 84 million in the third quarter. Notes payable rose from 351 million to 4.97 billion yuan this year, while cash fell to 555 million.
Training or Inference?
Bloomberg-based reports say Cambricon's chips run AI models from labs including DeepSeek and Moonshot, as China tries to cut its reliance on Nvidia. Some summaries go further and say those labs are routing training onto the chips. The underlying reports describe chips that run models, and no company or lab statement that training runs on Cambricon was found.
For comparison, ChinaTalk reported that DeepSeek's V4 was still trained on Nvidia chips. When V4 launched, Cambricon announced day-one support for running it, which is inference. Morgan Stanley puts Cambricon's share of China's AI chip market at about 14% this year, second to Huawei. That is a market-share estimate, not proof that these chips can train frontier models.
What to Watch
Watch fourth-quarter revenue against the 3.97 billion needed for the full-year target, how fast MLU692 deliveries are recognized, whether HBM supply eases, and whether the inventory turns into sales or more write-downs. Shares closed at 932.20 yuan on Friday, about 42% below the June 30 close, according to China AI Dispatch. Whether domestic chips can carry frontier training is a prediction, not a finding.
