Moore Threads' quarterly non-GAAP net profit turns positive for the first time, as domestic GPUs enter the delivery competition stage

Moore Threads has delivered its first quarterly financial report showing positive non-recurring net profit.

On the evening of October 28, Moore Threads released its 2026 third-quarter report. In that quarter, it achieved operating revenue of 1.552 billion yuan, up 1773.92% year on year and 55.4% quarter on quarter; net profit attributable to shareholders was 49.2604 million yuan, and non-recurring net profit attributable to shareholders was 20.0724 million yuan, both turning from losses to profits year on year.

Revenue hit a new single-quarter high, and growth is beginning to show up in profit. For a GPU company that still needs to keep investing in research and development, turning non-recurring profit positive is a noteworthy change. It means that after excluding relevant non-recurring gains and losses, Moore Threads still achieved a profit in the quarter. Of course, the third-quarter non-recurring net margin was about 1.29%, so the profit cushion is not yet thick.

Extending the timeline to the first three quarters, Moore Threads' cumulative revenue was 3.289 billion yuan, net profit attributable to shareholders was 37.6973 million yuan, and non-recurring net profit attributable to shareholders was still a loss of about 131 million yuan.

This makes Moore Threads a sample for observing the commercialization of domestic GPUs. As sales scale expands, the market will pay further attention to how revenue is converted into profit and how much capital cost growth requires.

Profitability moved forward by a year, and gross margin ranks in the first tier of domestic GPUs

Moore Threads previously said it expected to achieve profitability in 2027, and this financial report significantly moves that profitability timetable forward. It said revenue growth was due to demand from the artificial intelligence industry for computing power and full-function GPUs, as well as the accelerated deployment of intelligent computing products represented by MTT S5000.

It disclosed that customers include internet and cloud service providers, token service providers, AI innovation companies, operators and intelligent computing centers. In September, Moore Threads and JD Cloud announced a deep cooperation to build a 100,000-card full-function GPU cluster for large-model training, inference and embodied intelligence. On the operator side, China Mobile disclosed that the Shunyi node of Mobile Cloud's Beijing Token Center has deployed the S5000. In October, the first phase of the Wuxi Huishan domestic intelligent computing center is running at full capacity. The PD-separated heterogeneous inference solution based on the S5000 has already carried real token traffic from leading domestic model vendors.

Generally speaking, once a computing power chip enters a customer's production environment, subsequent expansion must take into account existing software adaptation, operations and maintenance experience and migration costs, and existing customers have the opportunity to bring continued procurement.

In the first three quarters, Moore Threads' R&D investment was 1.340 billion yuan, up 55.72% year on year, already exceeding the full-year 2025 figure of 1.305 billion yuan; third-quarter R&D investment was 571 million yuan, up 87.89% year on year. The company said R&D progress is accelerating on projects including products based on the new-generation full-function GPU "Huagang" architecture.

What requires more patient observation than profit is cash flow back into the business. In the first three quarters, Moore Threads' net cash flow from operating activities was -3.309 billion yuan. The company explained that to respond to growing demand and secure its supply chain, it expanded production scale and carried out strategic stockpiling, which led to increased procurement spending.

As of the end of September, the company's inventory was approximately 3.363 billion yuan, compared with 1.332 billion yuan at the end of last year; prepayments rose from approximately 1.782 billion yuan at the end of last year to 2.599 billion yuan; accounts receivable rose from approximately 434 million yuan to 994 million yuan.

In terms of overall gross margin, Moore Threads is at a relatively high level among domestic AI chip manufacturers. In the first half of 2026, its overall gross margin was 56.95%, close to MetaX's approximately 57.2% and Cambricon's approximately 55.3%; Biren Technology, Enflame Technology and Iluvatar CoreX were approximately 42.7%, 30.4% and 17.2%, respectively. Business structures differ among the companies, and overall gross margin cannot be directly equated with GPU product gross margin.

The fact presented by Moore Threads' financial report is that while revenue expanded, quarterly non-GAAP profit crossed the break-even line, but the cumulative non-GAAP loss in the first three quarters, the relatively large net operating cash outflow, and continuously growing R&D investment show that the company is still in a stage where expansion and profitability are being reconciled with each other.

Competition in domestic computing power chips enters the delivery stage

Moore Threads is not the only domestic GPU company bringing good news. On the evening of October 23, Cambricon disclosed its third-quarter report: revenue in the first three quarters was 9.526 billion yuan, up 106.76% year on year, and net profit attributable to shareholders was 3.715 billion yuan. Two days earlier, Hygon Information released its third-quarter report: revenue in the first three quarters was 14.955 billion yuan, up 57.58% year on year, and net profit attributable to shareholders was 2.847 billion yuan. MetaX's earnings forecast gave a range of 2.541 billion to 3.106 billion yuan for the first three quarters, turning profitable year on year.

This is a change in the industry's prosperity, mainly occurring at the level of demand structure. Over the past two years, domestic computing power buyers were concentrated among a few large model companies and intelligent computing centers in various regions. Now the share of inference is rising rapidly, and token service providers have been added to customer lists. Their business is directly tied to Token call volume; the larger the call volume, the more rigid the demand for GPUs.

Domestic substitution in inference scenarios is progressing faster than in training. Some products from Cambricon and Kunlunxin can already replace H20-class small cards, while large-parameter training still mainly relies on overseas cards, and domestic cards are mostly used for Agent inference and platform scheduling.

HBM and foundry capacity remain important factors constraining growth, and manufacturers are all increasing inventories preventively, with the direct trigger being upstream supply shortages and price increases.

For example, TSMC has decided to raise wafer prices by 3% to 6% depending on the process node; Samsung has raised foundry order prices, with advanced process nodes seeing increases of up to 15%; mainland foundries such as SMIC, Hua Hong, and Nexchip have also raised prices one after another. On the memory side, TrendForce expects general-purpose DRAM prices to rise 10% to 15% quarter-on-quarter in the fourth quarter of 2026, with the overall market remaining in short supply.

The gap is mainly concentrated in HBM. Global HBM capacity has long been concentrated among three companies—SK Hynix, Samsung Electronics, and Micron—and this is precisely where China's domestic supply chain is weakest.

CXMT's DRAM monthly capacity is expected to reach about 350,000 wafers by the end of 2026, with a target of 500,000 wafers in 2028, but on HBM it is on a different timeline: trial production of HBM3 began in early September 2026, with small-batch mass production only around the end of the year, targeting monthly capacity of about 60,000 wafers, while HBM4 is still in the pre-research stage.

CXMT lags leading manufacturers by about three to four years in HBM technology. A survey by Bloomberg Intelligence estimates that in 2026, China's demand for domestically produced AI chips will be about 4.2 million units, with supply of about 2.6 million units; the gap of 1.6 million units is mainly constrained by HBM allocation rather than chip design or logic process capacity.

Industry consensus is that the supply tightness will last at least 24 months, and 2026 to 2027 will remain a seller's market. A buyer-side source said that the technical capabilities of each company's latest-generation products are converging, with differences falling in adaptation, delivery, and commercial aspects—whether they can sustain scaled delivery is the real dividing line. (This article was first published on TMTPost APP, Text | TechPulse)