On October 24, Cambricon (688256.SH) released its third-quarter 2026 report: revenue of 9.526 billion yuan for the first three quarters, up 106.76% year-on-year; net profit attributable to shareholders of 3.715 billion yuan, up 131.49%; and net profit excluding non-recurring items of 3.523 billion yuan, up 148.30%. At the earnings briefing on the afternoon of October 26, investors pressed not on the numbers but on inventory. Chairman and General Manager Chen Tianshi also denied market rumors that the company would ship more than one million chips next year: "The company has never communicated anything of the sort."

Revenue Mix Has Shifted: Almost Entirely Cloud Products

Cambricon's portfolio spans cloud and edge AI chips and accelerator cards, AI servers, processor IP and foundational software. One easily overlooked change in the first three quarters: Chen made clear at the briefing that revenue came mainly from the cloud product line, and that "the revenue mix does not include the intelligent computing cluster systems business."

The cloud product line serves commercial customers of all kinds; the intelligent computing cluster systems business sells to local governments and state-linked enterprises building municipal AI data centers. In other words, the pattern of a single large project carrying a quarter did not repeat this time. The quarterly pace is also picking up: third-quarter revenue reached 3.531 billion yuan, up 104.47% year-on-year and 13.50% from the previous quarter, ahead of the 7.84% sequential growth posted in the second quarter.

Margins Held Up. Cash Flow Did Not

Gross margin for the first three quarters was 56.53%, essentially flat against 56.71% in 2024 and 55.15% in 2025. Net profit excluding non-recurring items grew faster (148.30%) than attributable net profit (131.49%), and non-recurring gains amounted to just 191 million yuan, or about 5.15% of attributable net profit. The earnings quality is genuine.

The other side of the story sits on the balance sheet. As of the end of September, inventory stood at 11.665 billion yuan, roughly 52.15% of total assets — up 3.417 billion yuan, or about 41.43%, from the end of the second quarter, and up 6.721 billion yuan from the end of 2025. Prepayments reached 3.044 billion yuan, an increase of 2.300 billion yuan since the start of the year. Together the two total 14.709 billion yuan. Over the same period, notes payable climbed from 351 million yuan to 4.966 billion yuan. Cash flow thinned as a result: net cash from operating activities was 228 million yuan for the first three quarters, and negative 84 million yuan in the third quarter alone.

The company booked roughly 502 million yuan in inventory write-downs over the first three quarters. CFO Ye Haoyin said this stemmed mainly from strategic stockpiling of raw materials in earlier periods, with provisions rising as the inventory aged.

Taken together, the numbers look like a "lock in supply first, deliver later" strategy: using prepayments and notes payable to reserve upstream capacity in advance, at the cost of cash tied up in materials and write-down risk pulled forward. Chen said procurement plans are "set according to customer order demand and forecasts of future market demand," but the composition of inventory, the share covered by orders, and the expected drawdown period were all left unquantified.

From Selling Chips to Building Systems: In-House Switch Chips Enter the Fundraising Plan

What determines the ceiling for the next phase is not this quarter's profit, but whether the company moves from selling chips to selling systems.

At the briefing, Chen said the fundraising project "chip platform for large models" now includes switch chip development. Under the plan, the associated interconnect protocols and switch chips will be paired with the company's own AI chips and accelerator cards to build multi-card computing power system equipment. The next-generation AI processor microarchitecture and instruction set will also be optimized specifically for natural language processing, video and image generation, and training and inference on vertical-domain large models.

Beyond single-card performance, cluster interconnect is a shared weak spot for domestic AI chips: stitching multiple cards into a usable computing power domain via high-speed interconnect is the practical way around limits on process node and on-card memory. Cambricon has until now supplied chips and cards; developing its own switch chips means stepping up into the system layer.

On the software side, the company has completed support for leading domestic open-source models including GLM, DeepSeek, Qwen, Kimi and MiniMax, and is advancing adaptation and optimization for DeepSeek, Qwen and Hunyuan, among others. Triton-MLU bridges the Triton programming model with the backend of Cambricon's MLU architecture. What is missing is a timeline — asked about the development stage of the switch chips, tape-out and mass production dates, and progress on super-node deployment, Chen replied that investors should "refer to the periodic reports for disclosure."