The Four Rising Stars of Domestic Computing Power: After Their Revenue Doubled, the Real Industry Threshold Has Surfaced.
On September 11, 2026, Enflame Technology was listed on the Sci-Tech Innovation Board at an issue price of 142.18 yuan, with an opening price of 410 yuan. At this point, all four companies known as the "Four Little Dragons of Domestic GPUs" — Moore Threads, Muxi Co., Ltd., Biren Technology, and Enflame Technology — have entered the public market. The capital market initially offered extremely generous valuations: Moore Threads closed up 425.46% on its first trading day, Muxi Co., Ltd. closed up 692.95%, the former's opening market value exceeded 300 billion yuan, and the latter's market value once hit 4000 billion yuan seven months after listing.
What has truly changed is the evaluation benchmark. From October 28 to 30, Moore Threads, Muxi, and Enflame successively released their first three-quarter financial reports since listing, while Biren still disclosed its data in accordance with the Hong Kong Stock Exchange's interim reporting standards. The year-on-year revenue growth rates of the three companies reached 319.15%, 128.45%, and 331.74% respectively. Both Moore Threads and Muxi have turned positive in non-recurring profit and loss, while Enflame is still operating at a loss, but its single-quarter loss has narrowed by 56.12% year-on-year.
However, the simultaneous doubling of revenue and profitability improvement among the four dragons does not mean that domestic computing power is losing competitiveness. Instead, the outside world has changed the way of raising questions: In the past, people asked whether the chip could be manufactured and whether the tape-out was successful; now they ask whether the product can be sold well, whether a single order is profitable, and whether the earned revenue can be retained on the books.
Divergence: The same industry boom, four different sets of financial accounts
The four companies were listed across a nine-month span, and their disclosure calibers are not consistent: three have released three-quarter reports, while Biren only has interim data from the Hong Kong Stock Exchange. But when the disclosed data are put together, the divergence in technical routes and business performance has become clear enough.
Moore Threads has pushed its scale advantage to a new level. Its total revenue in the first three quarters reached 3.289 billion yuan, a year-on-year increase of 319.15%. This figure exceeds Muxi's 2.824 billion yuan and also surpasses Enflame's 2.332 billion yuan. Looking at the third quarter alone, its revenue of 1.552 billion yuan increased by 55.4% quarter-on-quarter, which is close to the sum of revenue in the first two quarters. It adopts the full-function GPU route covering both graphics rendering and AI computing, independently develops the MUSA software stack, and expands the market relying on the MTT S5000 intelligent computing card and Kuane cluster. It also signed a 100,000-card cluster cooperation agreement with JD Cloud in September, which is scheduled to be completed in 2027.
The turning point of the profit statement occurred in the third quarter. The attributable net profit for the current quarter was 49 million yuan, and the attributable net profit after non-recurring gains and losses was 20 million yuan. Looking back at the first two quarters, it recorded a profit of 29 million yuan in the first quarter and a loss of 41 million yuan in the second quarter, and both indicators turned positive together in the third quarter. Combining the three quarters presents a different picture: the cumulative attributable net profit was 38 million yuan, and the cumulative loss after non-recurring gains and losses was 131 million yuan. The difference comes from 168 million yuan of non-recurring gains and losses, of which changes in fair value and disposal of financial assets contributed 89 million yuan.
The R&D investment in the first three quarters was 1.34 billion yuan, a year-on-year increase of 55.72%, which has exceeded the full-year R&D investment of 1.305 billion yuan in 2025. However, the expense ratio dropped from 109.7% in the same period of the previous year to 40.75% — revenue grows faster than investment, which is the direct reason why the single-quarter non-recurring profit and loss turned positive.
Muxi Co., Ltd. released its three-quarter report on the evening of October 30, and the actual figures fell in the middle of the forecast range announced on October 15: the revenue in the first three quarters was 2.824 billion yuan, a year-on-year increase of 128.45%, the attributable net profit was 639 million yuan, and the attributable net profit after non-recurring gains and losses was 259 million yuan. Both indicators reversed losses compared with the same period of the previous year. The single-quarter revenue in the third quarter was 1.5 billion yuan, a year-on-year increase of 367.12%, which exceeded the 1.324 billion yuan revenue in the first half of the year in just one quarter.
The contrast appears between two lines on the same profit statement. The attributable net profit after non-recurring gains and losses in the third quarter was 308 million yuan, nearly five times that of 54 million yuan in the second quarter; however, the attributable net profit in the same quarter was only 27 million yuan, down 96.27% quarter-on-quarter. The difference is approximately -281 million yuan of non-recurring gains and losses — the investment account that contributed 887 million yuan of fair value change income to Muxi in the first half of the year turned negative in the third quarter. The company originally noted that this income was not sustainable, and the financial report now has fulfilled half of this expectation.
What really supports the main business is products and delivery capacity. The Xiyun C600 entered mass production in May, equipped with 144GB HBM3e, and adopts a fully domestic supply chain from wafer to packaging. The Xijing S600 super node released in July supports high-speed interconnection inside the cabinet and can be scaled horizontally to a 10,000-card cluster. The gross profit margin in the first three quarters was 54.99%, down only 0.77 percentage points year-on-year, and the gross profit margin in the third quarter alone was 53.03%. The R&D investment reached 828 million yuan, a year-on-year increase of 19.12%, and the expense ratio continued to decline as revenue expanded.
The position brought by scale has not yet been converted into a valuation premium. Seven months after its listing, Muxi's market value once stood at 4000 billion yuan, and it returned to 186.6 billion yuan at the close of trading on November 2. The capital market is still willing to pay for domestic computing power, but it no longer pays a premium for the simple fact that "the chip can be manufactured".
Biren Technology has the highest revenue elasticity. In the first half of 2026, its revenue reached 1.236 billion yuan, a year-on-year increase of approximately 1997.6%, exceeding the full-year revenue of 1.035 billion yuan in 2025. The loss during the period narrowed to 377 million yuan, a year-on-year loss reduction of 76.4%, and the gross profit margin increased to 42.7%.
Biren Technology's strategy leans towards the system level. The "GuangYue" optical interconnection and optical switching GPU super node jointly released with Shanghai Yidian, Lightelligence, and ZTE has deployed thousands of cards, and the large model training performance has been improved by more than 30%. Its self-developed BLink 2.0 interconnection protocol supports a maximum of 1024-card expansion, and the R&D expenditure in the first half of the year was 804 million yuan. It is the only one among the four companies that does not have third-quarter data, and the interim caliber is enough to show its strategic orientation.
Enflame Technology has taken the most extreme path. It is not compatible with CUDA, and chooses the self-developed DSA route and GCU instruction set, which is different from the mainstream GPGPU route. It has iterated four generations of architectures and five chips in eight years, and all tape-outs are successful at the first attempt. The cost is that the cumulative R&D investment from 2023 to 2025 reached 3.676 billion yuan, equivalent to 1.8 times the cumulative revenue in the same period.
Enflame, which was just listed on September 11, released its first financial report on October 30. The revenue in the first three quarters was 2.332 billion yuan, a year-on-year increase of 331.74%, which was almost close to the lower limit of the previously announced forecast range of 2.3 billion to 3 billion yuan, far from the 455.36% growth rate corresponding to the upper limit of the range. The attributable net loss was 754 million yuan, narrowing by 15.07% year-on-year. The single-quarter revenue in the third quarter was 1.212 billion yuan, a year-on-year increase of 395.34%, with a loss of 122 million yuan, a year-on-year loss reduction of 56.12% and a quarter-on-quarter loss reduction of 35.31%. The single-quarter gross profit margin was 33.80%, up 2.35 percentage points year-on-year.
The return still relies heavily on Tencent. According to data from Tianyancha, Tencent is not only the largest external shareholder but also the largest customer, contributing 830 million yuan in revenue in 2025, accounting for 83.79% of total revenue. What is really worth paying attention to are two balance sheet items: the contract liabilities at the end of September were 6.385 billion yuan, an increase of 1745% over the end of last year, equivalent to 2.7 times the revenue in the first three quarters. The prepaid items in the same period were 7.709 billion yuan, which is the funds paid in advance to lock in wafers and storage. The net cash outflow from operating activities in the first three quarters was 2.017 billion yuan, and all the funds were occupied in these two aspects.
The commonalities are clear, the divergences are equally clear, and the divergence in technical routes is more fundamental. GPGPU compatible with CUDA and self-developed independent DSA: the former has low migration cost and fast start, while the latter is not subject to external restrictions, but the compiler and operator library need to be developed from scratch. The ecological barriers are being broken down inch by inch through massive engineering work, and every inch of progress requires investment.
Misalignment: Hardware can catch up, but the CUDA barrier cannot be bypassed
The domestic divergence is only the first layer. What truly determines the ceiling of the four companies is the distance between them and NVIDIA, and which parts of this distance can be caught up and which cannot.
The global coordinate system is set by NVIDIA. Its revenue for fiscal year 2026 reached 215.9 billion US dollars, a year-on-year increase of 65%, of which the data center business contributed 193.7 billion US dollars. In the second quarter of fiscal year 2027, its single-quarter revenue reached 96.221 billion US dollars, with a gross profit margin of 75.0%. NVIDIA's cumulative R&D investment has exceeded 76.7 billion US dollars, and about 31,000 of its 42,000 employees are engaged in R&D.
This is not a competition of a single chip, but a system that has been repeatedly reinforced by 20 years, hundreds of billions of dollars, and millions of developers. The gap at the hardware level is actually narrowing: the 144GB HBM3e specification of Muxi C600 has exceeded NVIDIA's H20 sold to China, and Biren's super node has achieved more than 30% end-to-end performance improvement in the actual measurement of mixture-of-experts model training.
What is really hard to catch up with is the ecosystem. The operator libraries, debugging tools, developer habits, and existing code accumulated by CUDA over 20 years constitute the migration cost itself. The difference in the magnitude of R&D investment is more realistic: NVIDIA's R&D expenses for fiscal year 2026 are 18.497 billion US dollars, equivalent to about 1310 billion yuan.
The special feature of this round of game is that it is not a pure market-based head-on confrontation. Export controls have blocked NVIDIA's most advanced Blackwell from entering China. Although H200 is allowed to be exported, the data center revenue it contributes accounts for less than 1%. NVIDIA's CFO clearly stated in the second-quarter financial report that the forward guidance does not assume any computing revenue from China's data centers.
Jen-Hsun Huang said that this set of policies forces NVIDIA to give up the Chinese market worth 500 billion US dollars. The proportion of revenue from China has dropped to about 9%, while 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 in China's data center AI chip market dropped to 34%, with shipments of 678,000 units, a year-on-year decrease of 44%. Huawei followed closely with a 28% share, and the gap was 31 percentage points a year ago.
The window of opportunity is real, but it is not a permanent dividend. The clear awareness of domestic manufacturers is that they do not set their main battlefield on competing with the world's most powerful pre-training cards, but focus on domestic thousand-card clusters, industry private large models, inference computing power, and information technology application innovation scenarios, and win orders through localized services, customized adaptation, and delivery certainty.
A structural change is more convincing. In the first half of 2026, the total shipment of AI accelerators in China reached 2.008 million units, of which GPU shipments were 991,000 units, down 26% year-on-year, and non-GPU accelerator shipments were 1.016 million units, an increase of nearly 80%, exceeding GPU shipments for the first time in quantity. Huawei, T-Head, Kunlunxin, and Cambricon took 95% of the non-GPU market segment. The demand focus has shifted to inference, which just falls on the position where domestic manufacturers can give full play to their advantages.
Changing situation: The days of rushing to buy chips are over, and the next step is to compete for the speed of return on investment
Misalignment can buy time, but it cannot bring the final outcome. What determines how far the four dragons can go is how long the domestic computing power demand will remain strong, and how this round of prosperity will eventually end.
Demand remains strong in the short term. According to data from the China Academy of Information and Communications Technology, by the end of June 2026, the scale of intelligent computing power in use across the country reached 2185 EFLOPS, a year-on-year increase of 177.3%. CIC Consulting predicts that the scale of China's AI accelerator card market will increase from 216.477 billion yuan in 2024 to 1.107646 trillion yuan in 2028, of which inference will account for 808.582 billion yuan, accounting for more than 70%.
Tencent plans to increase the proportion of domestic chip procurement to more than 65%, and Goldman Sachs judges that the deployment proportion of domestic AI chips in servers is expected to reach 70% to 85%. Enflame dares to give a maximum growth rate guidance of 455.36% for the first three quarters, and this is where its confidence lies.
However, red lights are also lighting up on the supply side. Advanced process, HBM, substrate, and high-speed interconnection are still in short supply. SK Hynix has clearly stated that its 2026 HBM production capacity has been fully sold out. TrendForce data shows that the sequential increase of DR