Moore Threads: After turning profitable in a single quarter, the cash consumption remains to be balanced
Recently, Moore Threads released its third-quarter financial report. In the third quarter, its revenue reached 1.552 billion yuan, up 1773.92% year on year and 55.4% quarter on quarter, hitting a new all-time high for a single quarter; its net profit attributable to shareholders stood at 49.26 million yuan, and the non-recurring profit and loss excluded net profit attributable to shareholders reached 20.07 million yuan — marking the first time that the non-recurring profit and loss excluded profit turned positive in a single quarter, which was earlier than the company's own expectation at the time of listing that "profitability would be achieved as early as 2027".
The market has been waiting for this figure for a long time. From the opening price of 650 yuan on the first day of listing in December 2025 with a market value exceeding 300 billion yuan, to the market value falling below 200 billion yuan in September 2026, the narrative of China's "first domestic GPU stock" has experienced a complete cycle of sentiment fluctuations in less than a year. The arrival of single-quarter profitability has been interpreted by many as "the inflection point has arrived".
However, if you check the cash flow section of the third-quarter report, the picture will become much more rational. The net cash outflow from operating activities in the first three quarters was 3.309 billion yuan. As of the end of Q3, prepaid expenses stood at 2.599 billion yuan and inventory at 3.363 billion yuan, totaling nearly 6 billion yuan, which is 3.4 times the revenue of the first half of the year. The company is using cash to lock in production capacity and delivery rhythm. The 20 million yuan non-recurring profit and loss excluded profit in a single quarter, in the face of cash consumption of this magnitude, is more like an accounting signal rather than a confirmation of a sustainable business model.
Moore Threads has indeed crossed a key threshold: its revenue scale has begun to cover high R&D investment. But the road after crossing the threshold is far longer than the threshold itself.
Quality of the Turnaround from Loss to Profit
The non-recurring profit and loss excluded net profit in Q3 was 20.07 million yuan, corresponding to a revenue of 1.552 billion yuan, with a non-recurring profit and loss excluded net profit margin of about 1.3%. In the same quarter, nearly 60% of the 49.26 million yuan net profit attributable to shareholders still came from non-recurring gains and losses such as changes in the fair value of financial assets. In other words, the real profitability thickness of its main business is much thinner than what the net profit attributable to shareholders figure presents.
What is more worth tracking is the trend of gross profit margin. The comprehensive gross profit margin in the first three quarters was 53.71%, down 8.16 percentage points year on year; the comprehensive gross profit margin in the first half of the year was 56.95%, which was already at a relatively high level among domestic GPU manufacturers, but the overall trend is downward. The reason for the declining gross profit margin is not difficult to infer: cloud intelligent computing products account for 97.5% of revenue, and the Kua'e intelligent computing cluster is the core driving force for revenue growth. The business model of cluster delivery naturally has the attribute of "project-based operation" — the amount of a single contract is large, the delivery cycle is concentrated, and the proportion of hardware BOM is high, leading to a gross profit margin lower than that of standardized chip products. Moore Threads is shifting from "selling graphics cards" to "selling clusters", so its revenue scale rises rapidly, but the gross profit margin left for itself per yuan of revenue is thinning.
This is not a problem in itself. Cluster delivery is the most realistic commercialization path for domestic GPU manufacturers in the current market. What customers need is operable computing power, not a bare graphics card that requires them to build the system on their own. The MFU of Kua'e cluster in Dense large model training reaches 60%, that in MoE reaches 40%, and the linear training scaling efficiency is 95%. These indicators prove that the product is technically reliable at the engineering level. The 660 million yuan cluster sales contract was recognized as revenue in the first half of the year, and the 100,000-card full-function GPU cluster cooperation with JD Cloud was also launched in September. There is no doubt that commercialization is accelerating.
However, between "project-based cluster delivery" and "sustainable chip platform revenue", there is a problem of customer structure and repurchase logic. Cluster orders come from internet and cloud service providers, operators, intelligent computing centers, and government and enterprise customers. In the procurement decisions of these customers, the weight of the localization requirement and the weight of the product performance itself determine the source structure of Moore Threads' current growth rate. If the orders are mainly driven by the window of "domestic products must be used", then after this window is filled, the next round of growth must come from the relative advantage of products in open competition. The downward trend of gross profit margin just indicates that this transition has not been completed: in a performance-driven market, a competitive product does not need to win customers through cluster bundling.
Cash Advance Payment
The two figures of 2.599 billion yuan in prepaid expenses and 3.363 billion yuan in inventory together form the most transparent window to understand the current operation rhythm of Moore Threads.
The company's explanation at the performance briefing is straightforward: in the context of tight supply chain, prepaid expenses help the company lock in scarce production capacity and resources, and inventory helps the company deliver orders in a timely and smooth manner. Against the background of the entity list and restricted advanced process OEM, this logic is valid. The production capacity ceiling of domestic GPU companies does not depend on demand, but on how many available wafers and packaging resources they can obtain. Converting cash into prepaid expenses and inventory in advance is a way to exchange financial redundancy for delivery certainty.
But the cost is also real. For the 3.3 billion yuan inventory, if the delivery rhythm of corresponding orders is delayed, it will turn into an impairment risk on the balance sheet; for the 2.6 billion yuan prepaid expenses, if the output efficiency of the upstream supply chain is lower than expected, the cash turnover cycle will be prolonged. The net cash outflow from operating activities of 3.3 billion yuan in the first three quarters has exceeded the total revenue of 3.29 billion yuan in the same period. For a company with rapidly expanding revenue, the negative operating cash flow that is larger than revenue means that for every yuan of business it does, more cash is paid in than is received. This state is normal in the early stage of large-scale delivery, but it cannot last indefinitely. When the revenue climbs from 3.3 billion yuan to the 12.6 billion yuan (2027) predicted by Shanxi Securities, the occupation of working capital will expand simultaneously with a larger absolute value. Whether the cash and financial assets on the books in Q3 are sufficient to support this expansion speed is a more urgent problem than single-quarter profit.
Gap in Ecosystem Construction
Moore Threads' investment in the software ecosystem is a variable that distinguishes it from other domestic GPU companies.
The MUSA architecture's CUDA compatibility strategy has become concrete in 2026. The torchada adapter allows PyTorch CUDA projects to run on Moore Threads GPUs with zero code modification by only adding one line of "import torchada". The runtime automatically translates CUDA API calls into equivalent MUSA calls, and the runtime overhead of frequent operations is less than 200 nanoseconds. The MUSIFY tool supports one-click conversion of most CUDA codes. The MUSA SDK covers 761 APIs, matching CUDA 12.8. According to official disclosed data, Automusify has realized 100% automatic migration of the top 100 artificial intelligence and scientific computing acceleration repositories, the number of MUSA software stack developers has exceeded 800,000, and the SGLang inference framework has completed the mainline integration of MUSA code.
These figures point to the right direction. Migration cost is the hardest part of the CUDA ecosystem barrier. What torchada and MUSIFY are doing is reducing "migration" from an engineering project that requires rewriting code to a configuration-level operation. For small and medium-sized development teams and academic research scenarios, this low-friction compatibility strategy is the most feasible path for Moore Threads to build its first batch of native users outside NVIDIA's ecosystem.
However, there is a gap that needs to be honestly marked between "able to run" and "running well". Torchada solves the call translation at the API level, which makes the code run without errors, but does not guarantee that the execution efficiency of operators on MUSA is equivalent to that on CUDA. The communication primitives in large-scale distributed training, the numerical stability of mixed-precision computing, and the performance optimization of specific operators are the deep-water areas that API translation cannot cover. 800,000 developers is a meaningful base, but how many of them are doing production-level training or inference on MUSA rather than test migration is the real scale of ecosystem stickiness. The mainline integration of SGLang is a good signal, but it is the adaptation of an open-source framework to hardware, not the hardware's right to define the framework. The barrier of CUDA does not lie in the number of APIs, but in the engineering inertia, tuning experience and troubleshooting knowledge base accumulated by millions of developers on it in the past 15 years. The 800,000 MUSA developers are facing this mountain, not having already climbed over it.
Gap in Industry Echelon
IDC's data shows Moore Threads' position in the domestic AI acceleration card market. In the first half of 2026, Huawei Ascend accounted for 28% of the market share, NVIDIA's share dropped to 34%, the second echelon composed of T-Head, Cambricon, Kunlunxin and Haiguang accounted for 34.6% in total, and the third echelon of six companies including Moore Threads, Mthreads and Iluvatar accounted for 13.3% in total. In the narrower statistical caliber of domestic AI chips, Huawei alone accounts for 52.1% of the shipments.
The meaning of this position is clear. Moore Threads is advancing rapidly in commercialization. Its revenue of 1.736 billion yuan in the first half of the year ranks first among the four listed domestic GPU companies, and its single-quarter turnaround from loss to profit in Q3 is also one of the earliest to cross this threshold among the "four little dragons". However, the leading position in revenue scale and the ranking in market share are two different coordinate systems. Relying on the Da Vinci architecture and full-stack self-development, Huawei Ascend has built a system-level barrier in the government, enterprise and operator markets; Cambricon enters the inference side with price and energy efficiency advantages; Mthreads focuses on reducing migration costs. Moore Threads' differentiation lies in "full-function GPU" — covering AI training and inference, graphics rendering and scientific computing at the same time. This positioning is scarce among domestic manufacturers, but the cost is that it has to face a well-established opponent in every battlefield.
The length of the window period depends on two variables: to what extent the proportion of localized procurement can continue to rise, and to what extent Huawei Ascend's production capacity can cover the demand of the low-end and mid-end markets. The first variable has reached a fairly high level in 2026, and the answer to the second variable is becoming unfavorable to the third echelon — the faster Ascend's production capacity ramps up, the narrower the space left for Moore Threads to grow in the gap of "having to buy domestic products but unable to get Ascend products". The turnaround to profit in Q3 proves that Moore Threads has the ability to expand revenue in the current window, but it has not yet proved what it can rely on to stay in the game after this window closes.
Valuation Outlook
On October 30, Moore Threads closed at 322.65 yuan, with a total market value of 151.655 billion yuan, down more than 65% from the intraday high of 442.2 billion yuan on the first day of listing. Shanxi Securities predicts that its revenue in 2026 will be 5.39 billion yuan with a net profit attributable to shareholders of 150 million yuan, and its revenue in 2027 will be 12.612 billion yuan with a net profit attributable to shareholders of 1.243 billion yuan. Even based on the predicted profit of 2027, the price-earnings ratio corresponding to the current market value exceeds 120 times; based on the predicted profit of 2026, the price-earnings ratio exceeds 1000 times.
This valuation does not need to be defined as "expensive" or "cheap", it only needs to be accurately described: the market is not buying the 150 million yuan profit in 2026, nor even the 1.2 billion yuan profit in 2027 completely. It is buying a probability — that Moore Threads' revenue growth slope can continue to maintain at a level that makes the current market value reasonable after 2028, and what supports this slope is that Kua'e cluster orders shift from project-based operation to repurchase, the MUSA ecosystem shifts from "able to run CUDA" to "native development by developers", and the gross profit margin stops declining under the scale effect of cluster delivery.
The single-quarter non-recurring profit and loss excluded turnaround from loss to profit is a real milestone. It proves that the revenue scale has been large enough to cover the current fixed costs of R&D and operation. However, the coexistence of the three figures — negative 3.3 billion yuan operating cash flow, nearly 6 billion yuan in inventory and prepaid expenses, and an 8.16 percentage points year-on-year drop in gross profit margin in the first three quarters, indicates that this milestone is standing on a very narrow ridge line. Moore Threads does not need to choose between being a "technology company" and a "commercial company". What it needs to prove is that the delivery capacity obtained through cash investment can eventually be transformed into a revenue structure that does not rely on the localization procurement window. Before that, the "future" priced by the 151.6 billion yuan market value is far more distant than the 20 million yuan non-recurring profit and loss excluded profit in Q3.
This article is from the WeChat official account "Tech Business", authored by Tech Business, and authorized for release by 36Kr.