Moxi Shares' revenue in 2025 increased by 1.2 times year-over-year, net loss narrowed significantly to 790 million yuan, and GPU sales exceeded 55,000 units cumulatively | Financial Report Insights

AI Questions · GPU sales exceed 55,000 units, how to promote the construction of domestic computing power ecosystem?

Against the backdrop of accelerated domestic intelligent computing center development and rising demand for domestic computing power, Muxi Co., Ltd. achieved an annual revenue of 1.644 billion yuan, a year-on-year increase of 121.26%, with revenue continuing to grow rapidly.

Simultaneous improvement in losses. According to the financial report released on Thursday, the company’s net loss attributable to parent in 2025 was 789 million yuan, a 43.97% reduction from the previous year; net loss after non-recurring gains and losses was 830 million yuan, a 20.52% decrease.

Business progress: The company emphasizes it has become one of the few GPU suppliers in China to achieve large-scale commercial application of “thousand-core clusters.” Cumulative GPU sales exceed 55,000 units. It is also advancing the implementation of more industry scenarios and larger-scale clusters around the CUDA-compatible MXMACA software stack and self-developed MetaXLink high-speed interconnection capabilities.

Revenue: 1.644 billion yuan, doubling, with the second quarter contributing the most

In 2025, Muxi Co. achieved revenue of 1.644 billion yuan, a 121.26% increase year-on-year. Looking at quarterly distribution, the company’s income was concentrated in the second quarter:

• Q1: 320 million yuan
• Q2: 595 million yuan
• Q3: 321 million yuan
• Q4: 408 million yuan

The company’s main business revolves around full-stack GPU chips and computing platforms. Revenue mainly comes from GPU cards and servers, all-in-one machines/workstations with integrated cards, and intelligent computing clusters. It also provides software stacks compatible with mainstream CUDA ecosystems (MXMACA) to meet customer computing needs through a “hardware + software + system” approach.

Profit and Loss Statement: Significant reduction in losses, but quarterly fluctuations are large, and profit quality still needs observation

The full-year net loss attributable to parent was 789 million yuan, and net loss after non-recurring gains and losses was 830 million yuan, both narrower than the previous year. Roughly calculating net loss margin based on revenue, it remains around -48%, indicating that scale expansion has not yet fully covered R&D and industrialization costs.

Notably, profits fluctuate significantly between quarters:

• Q2: Net profit attributable to parent was +46.62 million yuan, and after non-recurring items, still +16.60 million yuan;
• Q4: There was a net loss attributable to parent of -444 million yuan, and a net loss after non-recurring items of -437 million yuan.

During the phase when AI chip companies are delivering products, confirming R&D investments, and climbing the industrialization curve, quarterly fluctuations are not uncommon. However, this also signals to investors that profit recovery is not a linear process. Future monitoring should focus on gross profit margins, expense ratios, customer acceptance, and payment collection rhythms (full details in the annual report).

Products and Commercialization: Three product lines covering training, inference, and graphics rendering, following a “solution” approach

Muxi Co. is still in the typical “high investment, long cycle” AI chip R&D phase. The annual report shows R&D expenditure accounts for 62.49% of revenue. The parent company’s accumulated unrecouped losses amount to 1.549 billion yuan, and no dividends are planned for 2025; operating cash flow is a net outflow of 1.26 billion yuan, which, although improved from last year, still exerts cash pressure.

The company’s product lines are more clearly categorized by scenario:

• Xiyun C series (C500/C600): training and inference integration, general computing, AI for Science;
• Xisi N series (N100/N260/N300): from traditional AI inference to generative AI inference;
• Xicai G series (G100): graphics rendering in development and layout.

In terms of business model, the company adopts a fabless model (outsourcing wafer manufacturing, packaging, testing, and board processing). Sales are conducted through direct sales and distributors in parallel. Compared to single-card sales, the annual report emphasizes delivering servers, all-in-one machines/workstations, and even intelligent computing clusters composed of multiple servers and network storage, reflecting a strategy of “engineering delivery + ecosystem adaptation” in the domestic GPU industry.

The company states its products have been deployed in over ten intelligent computing clusters, covering national AI public computing platforms, operator intelligent computing platforms, and commercial AI centers, with regions including Beijing, Shanghai, Hangzhou, Changsha, and Hong Kong. It is also promoting scenario applications in education and research, finance, transportation, energy, healthcare, entertainment, as well as in “embodied intelligence” and “low-altitude economy.”

Muxi’s industry trend outlook in the annual report focuses on three main lines:

1. From “computing power sovereignty” to “computing power equality”: Represented by algorithm optimization and open-source diffusion brought by DeepSeek, short-term training efficiency improvements may impact single training demands, but in the long term, under the “Jevons paradox,” AI application explosions could still lead to overall growth in computing power demand; additionally, the demand structure is shifting from “heavy pre-training” to a more balanced “pre-training + post-training + inference.”

2. Interconnection determines effective computing power under distributed parallelism: ScaleUp/ScaleOut parallelism is becoming mainstream, and multi-GPU interconnection bandwidth directly affects cluster efficiency. Upgrading high-speed interconnection technology will continue.

3. Software ecosystem determines ease of use and deployment speed: In the era of large models, high-frequency updates of operators mean that how quickly chips can adapt to frameworks, models, and toolchains is crucial for scene richness.

Market space data: The company cites third-party research indicating that in 2024, China’s AI acceleration chip market size is approximately 142.537 billion yuan, a 98.49% increase year-on-year, expected to reach 239.8 billion yuan in 2025; IDC data shows that in 2024, China’s accelerated computing server market size is 22.1 billion USD, a 134% increase YoY, with projections exceeding 100 billion USD by 2029.

Not meeting cash dividend conditions

The company notes that it “was not profitable at the time of listing and has not yet achieved profitability.” For 2025, it plans not to distribute cash dividends, issue bonus shares, or convert shares, due to the parent company’s accumulated unrecouped losses, with distributable profits being negative.

Shareholder structure: As of the end of the reporting period, there were 31,302 ordinary shareholders. Among the top ten shareholders, founder Chen Weiliang and his concerted action platforms (Shanghai Jiaomai, Shanghai Xiji) hold prominent stakes; Ge Weidong and his concerted action platform (Shanghai Hunt Investment) are also among the top ten, along with several equity and guiding funds. The audit opinion from Lixin is a standard unqualified opinion.