First financial reports from China's four little GPU dragons: after revenue doubles, profitability and the CUDA ecosystem become the real thresholds

On September 11, 2026, Enflame Technology was listed on the STAR Market at an issue price of 142.18 yuan and opened at 410 yuan. With that, all four companies known as the "four little dragons of domestic GPUs"—Moore Threads, MetaX, Biren Technology, and Enflame Technology—had entered the public market. The capital market's initial pricing was extremely generous: Moore Threads closed up 425.46% on its first trading day, and MetaX closed up 692.95%; the former's market value topped 300 billion yuan at the open, while the latter's market value briefly reached 400 billion yuan seven months after listing.

What has really changed is the yardstick for evaluation. From October 28 to 30, Moore Threads, MetaX, and Enflame successively delivered their first quarterly reports since listing, while Biren still disclosed under the Hong Kong interim reporting basis. The three companies' year-on-year revenue growth rates were 319.15%, 128.45%, and 331.74%, respectively. Both Moore Threads and MetaX moved into positive non-recurring profit, while Enflame remained in loss, though its single-quarter loss narrowed 56.12% year on year.

However, the four little dragons' revenue doubling and profitability appearing at the same time does not mean domestic computing power is failing; rather, the outside world has changed the question it asks: in the past, it asked whether chips could be built and whether tape-out succeeded; now it asks whether they can be sold, whether a single order makes money, and whether the money earned can stay on the books.

Divergence: the same boom, four different sets of accounts

The four companies' listings span nine months, and their disclosure bases are inconsistent: three have already provided third-quarter reports, while Biren has only Hong Kong interim data. But putting the disclosed data side by side, the divergence in routes and performance is already clear enough.

Moore Threads has pulled its scale advantage to a new position. Revenue in the first three quarters was 3.289 billion yuan, up 319.15% year on year, exceeding MetaX's 2.824 billion yuan and Enflame's 2.332 billion yuan. Looking only at the third quarter, revenue of 1.552 billion yuan rose 55.4% quarter on quarter, approaching the sum of the first two quarters in a single quarter. It follows a full-function GPU route, covering both graphics rendering and AI computing, with its self-developed MUSA software stack, and markets itself through the MTT S5000 intelligent computing card and the Kuae cluster; in September it signed a 100,000-card cluster cooperation with JD Cloud, planned for completion in 2027.

The turning point in the income statement occurred in the third quarter. Net profit attributable to shareholders in the quarter was 49 million yuan, and non-recurring net profit attributable to shareholders was 20 million yuan. Looking back at the first two quarters, the first quarter was profitable at 29 million yuan, the second quarter lost 41 million yuan, and by the third quarter both measures turned positive together. Taken across the three quarters, the picture is different: cumulative net profit attributable to shareholders was 38 million yuan, while non-recurring profit was a loss of 131 million yuan; the difference came from 168 million yuan in non-recurring gains and losses, of which changes in the fair value of financial assets and disposals contributed 89 million yuan.

R&D investment in the first three quarters was 1.340 billion yuan, up 55.72% year on year, already exceeding the full-year 2025 figure of 1.305 billion yuan, but the expense ratio fell from 109.7% in the same period last year to 40.75%—revenue grew faster than investment, and that is the direct reason single-quarter non-recurring profit could turn positive.

MetaX's third-quarter report was delivered on the evening of October 30, with actual figures landing in the middle of the forecast range given on October 15: first-three-quarter revenue of 2.824 billion yuan, up 128.45% year on year; net profit attributable to shareholders of 639 million yuan and non-recurring net profit attributable to shareholders of 259 million yuan, both turning profitable from the same period last year. Third-quarter single-quarter revenue was 1.500 billion yuan, up 367.12% year on year, exceeding the first half's 1.324 billion yuan in a single quarter.

The contrast appears between two lines of the same income statement. Third-quarter non-recurring net profit attributable to shareholders was 308 million yuan, nearly five times the second quarter's 54 million yuan; in the same quarter, net profit attributable to shareholders was only 27 million yuan, down 96.27% quarter on quarter. The difference was about -281 million yuan in non-recurring gains and losses—the investment account that contributed 887 million yuan in fair value gains to MetaX in the first half turned negative in the third quarter. The company itself noted at the time that this gain was not sustainable, and the statements have now cashed in half of it.

What truly supports the main business is products and delivery. The Xiyun C600 entered mass production in May, equipped with 144GB HBM3e, using an all-domestic supply chain from wafer to packaging; the Xijing S600 supernode released in July supports high-speed interconnection within the rack and can scale horizontally to 10,000-card clusters. The gross margin in the first three quarters was 54.99%, down only 0.77 percentage points year on year, and the third-quarter single-quarter margin was 53.03%; R&D investment was 828 million yuan, up 19.12% year on year, and the expense ratio fell steadily as revenue expanded.

The position bought by scale has not yet translated into a valuation premium. Seven months after listing, MetaX's market value briefly topped 400 billion yuan, and by the November 2 close it was back at 186.6 billion yuan. The capital market is still willing to pay for domestic computing power, but it no longer pays a premium for "being able to build it."

Biren Technology has the greatest revenue elasticity. Revenue in the first half of 2026 was 1.236 billion yuan, up about 1997.6% year on year, exceeding the full-year 2025 figure of 1.035 billion yuan; the loss for the period narrowed to 377 million yuan, a 76.4% reduction year on year, and the gross margin rose to 42.7%.

Biren Technology's approach leans toward the system level. The "Guangyue" optical interconnect optical switching GPU supernode jointly released with Shanghai Instrumentation, Xizhi Technology, and ZTE has deployed thousands of cards, improving large-model training performance by more than 30%; its self-developed BLink 2.0 interconnect protocol supports expansion to 1,024 cards, and R&D spending in the first half was 804 million yuan. It is the only one of the four without third-quarter data, but the interim basis is enough to show its approach.

Enflame took the most extreme path of all. It is not CUDA-compatible; instead it chose a self-developed DSA route and the GCU instruction set, diverging from the mainstream GPGPU approach, and over eight years it iterated through four generations of architecture and five chips, all of which succeeded on the first tape-out. The price was cumulative R&D investment of 3.676 billion yuan from 2023 to 2025, equivalent to 1.8 times cumulative revenue over the same period.

Enflame, which only listed on September 11, delivered its first financial report on October 30. Revenue for the first three quarters was 2.332 billion yuan, up 331.74% year on year, almost exactly at the lower bound of its earlier forecast range of 2.3 billion to 3 billion yuan and far from the 455.36% growth rate corresponding to the upper bound; net loss attributable to shareholders was 754 million yuan, narrowing by 15.07% year on year. Third-quarter revenue alone was 1.212 billion yuan, up 395.34% year on year, with a loss of 122 million yuan, a 56.12% year-on-year reduction and a 35.31% quarter-on-quarter reduction; the single-quarter gross margin was 33.80%, up 2.35 percentage points year on year.

The returns are still pinned on Tencent. Tianyancha data shows that Tencent is both the largest external shareholder and the largest customer, contributing 830 million yuan in revenue in 2025, or 83.79% of revenue. What is really worth watching is two balance-sheet line items: contract liabilities of 6.385 billion yuan at the end of September, up 1,745% from the end of the previous year and equivalent to 2.7 times first-three-quarter revenue; and prepayments of 7.709 billion yuan over the same period, money paid in advance to lock in wafers and memory. Net cash outflow from operating activities in the first three quarters was 2.017 billion yuan, with all the money tied up at these two ends.

The commonalities are clear, the divergences are equally clear, and the split in routes is more fundamental. CUDA-compatible GPGPU versus self-developed independent DSA: the former has low migration cost and starts quickly, while the latter is not beholden to others, but the compiler and operator libraries have to be built from scratch. The ecosystem barrier is being gnawed down inch by inch through engineering effort, and every inch costs money.

Dislocation: hardware can catch up, but the CUDA wall cannot be bypassed

The domestic differentiation is only the first layer. What really determines the ceiling for the four companies is the distance between them and Nvidia, and which parts of that distance can be closed and which cannot.

The global coordinate system is set by Nvidia. Fiscal 2026 revenue was $215.9 billion, up 65% year on year, of which data center business was $193.7 billion; second-quarter fiscal 2027 revenue alone was $96.221 billion, with a gross margin of 75.0%. Nvidia's cumulative R&D investment exceeds $76.7 billion, and about 31,000 of its 42,000 employees work in R&D.

This is not a contest over a single chip, but a system reinforced over twenty years, tens of billions of dollars, and millions of developers. The hardware gap is actually narrowing — the 144GB HBM3e specification of MetaX's C600 already exceeds the H20 that Nvidia sells to China, and Biren's supernode achieved more than 30% end-to-end performance improvement in measured training of mixture-of-experts models.

What is truly hard to catch up with is the ecosystem. The operator libraries, debugging tools, developer habits, and existing code that CUDA has accumulated over twenty years themselves constitute migration costs; the gap in R&D spending is even more concrete: Nvidia's fiscal 2026 R&D expense was $18.497 billion, equivalent to about 131 billion yuan.

What is special about this round of competition is that it is not a purely market-based head-on clash. Export controls keep Nvidia's most advanced Blackwell out of China; although H200 has been approved for export, it contributes less than 1% of data center revenue, and Nvidia's CFO explicitly stated in the second-quarter earnings report that forward guidance assumes no computing revenue from Chinese data centers.

Jensen Huang's account is that this policy forced Nvidia to give up a $50 billion Chinese market, and the share of revenue from China has fallen to about 9%, whereas three years ago China accounted for at least one-fifth of its data center revenue. IDC data shows that in the first half of 2026 Nvidia's share of China's data center AI chips fell to 34%, with shipments of 678,000 units, down 44% year on year; Huawei followed at 28%, a gap that was still 31 percentage points a year earlier.

The window is real, but it is not a permanent dividend. What is clear-headed about the domestic vendors is that they did not set their main battlefield on benchmarking against the world's strongest pretraining cards, but rather on domestic thousand-card clusters, industry-private large models, inference compute, and Xinchuang scenarios, winning orders through localized services, customized adaptation, and delivery certainty.

One structural change is more convincing. In the first half of 2026, total AI accelerator shipments in China were 2.008 million units, of which GPU shipments were 991,000 units, down 26% year on year, while non-GPU accelerator shipments were 1.016 million units, up nearly 80%, surpassing GPUs in volume for the first time; Huawei, T-Head, Kunlunxin, and Cambricon took 95% of the non-GPU segment. The center of demand has shifted to inference, which happens to fall exactly where domestic vendors can best play to their strengths.

Shifting landscape: the days of scrambling for cards are over; next it is about who gets their money back fastest

Dislocation can buy time, but it cannot buy the endgame. What determines how far the four little dragons can go is how long domestic compute demand can stay strong, and in what manner this boom ultimately ends.

Demand remains strong in the short term. Data from the CAICT shows that as of the end of June 2026, the scale of intelligent computing power in use nationwide reached 2,185 EFLOPS, up 177.3% year on year; China Insights Consultancy forecasts that China's AI accelerator card market will grow from 216.477 billion yuan in 2024 to 1,107.646 billion yuan in 2028, of which inference will be 808.582 billion yuan, accounting for more than 70%.

Tencent plans to raise the share of domestic chips in its procurement to above 65%, and Goldman Sachs judges that domestic AI chips are expected to reach a deployment share of 70% to 85% in servers. This is where Enflame's confidence comes from in daring to give a growth guidance of up to 455.36% for the first three quarters.

Yet the supply side is simultaneously flashing red. Advanced processes, HBM, substrates and high-speed interconnect remain tight; SK Hynix has explicitly stated that its 2026 HBM capacity is fully sold out, and TrendForce data shows that the quarter-on-quarter increase in DRAM contract prices rose from 58% to 63% in the second quarter. At the end of the first half of 2026, Biren Technology had 1.534 billion yuan in inventories and service-related prepayments, up 230% from the end of 2025—all of it paying for this.

What is more troublesome is utilization. Under CAICT's methodology, the actual utilization rate of domestic intelligent computing resources in 2025 was only 36.8%; for every 100 yuan invested in building computing power, the resources corresponding to about 70 yuan sit idle. Zhang Jin of Tencent Cloud gave an even lower figure at the MWC26 Smart Data Center Summit: the average utilization rate of intelligent computing centers nationwide is only 20% to 30%, and some enterprise-level ones are below 10%. Minmetals Securities calculates that the ratio of AI capital expenditure to corresponding revenue for five leading cloud vendors has risen from 59.2% in 2023 to 128.3% in the first half of 2026.

The industry is thus shifting from extensive expansion—scrambling for chips and building clusters—to a refined stage that prioritizes computing utilization and ROI. This is a double-edged sword for the four little dragons: inefficient computing power is eliminated, while clusters that can actually run become more valuable; but buyers are beginning to charge by token consumption and settle by outcome-based delivery, and chip vendors can no longer simply sell the cards and be done with it.

The competitive factors therefore shift from single-card performance to a comprehensive contest of full-stack capability, cluster delivery, ecosystem migration costs and yield control. On October 15, the day StepFun open-sourced Step 5 Preview, Ascend, Moore Threads, MetaX, Biren and Hygon DCU all completed adaptation; Day 0 adaptation has become widespread among leading vendors, meaning the differentiation brought by model adaptation is narrowing.

The risk boundaries also become clear at this point. The first is application rollout falling short of expectations: Gartner warns that by the end of 2027, more than 40% of Agentic AI projects will be canceled due to high costs and unclear value. The second is that improved algorithm efficiency reduces computing consumption—the same intelligent output can be produced with fewer cards, which hurts those selling the cards.

The third is changes in overseas supply policy: once controls are relaxed, the time window will immediately narrow; domestic price involution and the pace of ecosystem adaptation determine who can turn book revenue into sustainable profit. Morgan Stanley estimates that domestic cloud vendors' ROIC is in the 13% to 20% range, with a cash payback period of about 3 years.

Back to the yardstick at the beginning. The four little dragons have already completed the hardest step: making chips, achieving successful tape-out, getting them into data centers, and putting revenue into financial reports; the remaining step is harder and more valuable—getting customers, in the next procurement cycle, to keep placing orders because the products are good to use, not because they have no choice. The window is still there, but windows never wait for anyone.

This article comes from the WeChat public account "Deep Dive atom" (ID: deepatom), author: Song Que, published by 36Kr with authorization.