중국 국내 AI 칩 회사들이 수년간의 정부 보조금과 R&D 투자를 거쳐 2026년 상반기에 수익성을 달성했으며, 공급망 현지화에서 중요한 이정표를 기록했다.

Recent half-year reports from seven Chinese AI chip companies—Hygon, Cambricon, Moore Threads, Moxin, Biren Technology, Tianshuo Intelligent Chip, and Enflame (which included updated interim data in its IPO prospectus)—reveal a collective revenue of approximately 21.46 billion yuan for the first half of 2026, nearly double the same period last year.

Revenue growth is nearly universal: Cambricon up 108.13%, Moore Threads up 147.42%, Tianshuo up 191.6%, Enflame up approximately 279%, and Biren up 1997.6% on a low base. Domestic AI chip companies that once relied on financing to support R&D are collectively crossing the threshold from sample chips and testing into volume production.

**Who Is Actually Making Money?**

The answer to "who is most profitable" cannot be determined simply by ranking net profit. Hygon has the highest revenue; Cambricon has the highest net profit attributable to shareholders; Moxin and Tianshuo have returned to profitability on their statements despite still-negative core business margins; Moore Threads is just one step away from breakeven; Biren and Enflame are growing revenue rapidly but remain in tens of millions in losses.

Four clear trends emerge:

**First, revenue growth is collective and widespread, with multiple companies doubling.** Cambricon's revenue increased 108.13%, Moore Threads 147.42%, Tianshuo 191.6%, Enflame roughly 279%, and Biren an astounding 1997.6% on a low base. Hygon and Moxin, though not doubling, still achieved 66.52% and 44.67% growth—well above most mature chip companies. More significantly, several companies' half-year revenue already exceeds their full-year 2025 results. Moore Threads' 17.36 billion yuan in H1 revenue surpassed its 2025 full-year 15.06 billion yuan; Biren's 12.36 billion yuan also exceeded last year's total; Enflame's 11.20 billion yuan in half-year revenue surpassed last year's full-year 9.90 billion yuan. Chinese AI chipmakers are stepping up in scale, signaling a shift from small-batch testing and project validation toward larger-scale product and cluster deliveries.

**Second, profitable and cash-burning companies are beginning to diverge.** While revenue grows together, profitability gains are not synchronized. Hygon and Cambricon have entered large-scale, core-business profitability stages. Moore Threads and Moxin occupy the next tier, both touching the breakeven line on core business. Moore Threads' net loss narrowed to 11.56 million yuan; Moxin's non-GAAP net loss for the half-year was 49 million yuan. Biren posted 527 million yuan in gross profit in H1 while investing 804 million yuan in R&D; Enflame's net loss of 630 million yuan reflects its simultaneous expansion of products and customers.

**Third, gross margins retain resilience overall; differences largely reflect product mix.** Chinese AI chips have not universally resorted to low prices to grab market share after ramping. Cambricon's H1 combined gross margin was 55.25%, rising further to 56.10% in Q2; Moxin's gross margin of approximately 57.22% saw a slight 1.1 percentage-point year-over-year increase; Moore Threads' gross margin stands at 56.95%. All three companies maintain margins above 55%, indicating that Chinese chips retain pricing power in a supply-constrained market. Notably, Tianshuo's overall gross margin dropped from approximately 50.1% in the prior year to 17.2%.

**Fourth, R&D investments are entering payoff phase, with differences in conversion speed.** High R&D spending is standard for the chip industry. Among the GPU companies still climbing: Moore Threads, Moxin, Biren, Tianshuo, and Enflame invested 769 million, 525 million, 804 million, 559 million, and 640 million yuan respectively in H1 R&D—a combined 3.297 billion yuan, or roughly 52% of these five companies' combined half-year revenue. By contrast, Cambricon's H1 R&D expenses of approximately 703 million yuan are comparable to Moore Threads and Biren. Cambricon's profit performance reflects that its existing products have entered large-scale delivery and profit-realization phases.

Expense ratios better reveal each company's development stage. Moore Threads and Moxin's R&D expense ratios are 44.3% and 39.65% respectively; Biren, Tianshuo, and Enflame are at roughly 65%, 59.1%, and 57.1%. High R&D spending is merely the price of entry; what truly determines financial results is whether investments produce products that can actually be manufactured at scale, whether those products enter large clusters, and whether they generate repeat orders. Competition among Chinese AI chipmakers is shifting from "will you invest?" to "can you deliver?"

**Supply Shortages: Orders Stretch Into Next Year**

According to CCTV, domestic computing power chips are experiencing a surge in orders, with delivery schedules for some products already extending a year ahead. Domestic demand for computing power exceeds supply by tenfold. Multiple computing service providers report that orders have maintained rapid growth for two consecutive years, with some companies seeing customer counts grow roughly tenfold over the past year. This demand propagates up the supply chain, creating tightness in chips, servers, storage, and packaging.

Behind the supply shortage lies Chinese chips increasingly handling real, high-volume services. In late August, Zhipu disclosed deployment details of GLM-5.3-Flash, the first native multimodal model in the GLM-5 series: all online traffic is carried by 100,000 units of Chinese chips. Before the model's formal launch, Zhipu tested it anonymously on overseas platforms OpenRouter and OpenCode, accumulating 62 trillion token calls, with all relevant request traffic powered by domestic chips. Following underlying architectural modifications and inference service optimization, the cluster's hardware efficiency and per-token cost have reached parity with mainstream NVIDIA GPUs. "This proves that Chinese chips can fully support large-scale, efficient, and economical inference demands for cutting-edge models," Zhipu stated in its announcement.

Such cases validate not just peak computing power but also concurrent scheduling, inter-chip communication, fault recovery, model adaptation, and cost control. Customer procurement standards have accordingly expanded from single-card performance to system-wide availability. Chinese chips are transitioning from "can light up" to "can serve reliably," while supply-demand dynamics acquire new anchors.

Demand structure is shifting in parallel. Training clusters remain the source of large-scale purchases, but inference—driven by search, recommendations, intelligent customer service, content generation, and enterprise AI agents—is unleashing steady, high-frequency computing demand. Tianshuo's revenue structure already reflects this curve. In H1 2026, the company's inference series revenue reached 654 million yuan, up 651.8% year-over-year, surpassing its training series' 262 million yuan and approaching double its full-year 2025 inference revenue. The company attributes this to both higher unit volumes and pricing of premium product versions. For Chinese chip companies, the inference market opens a second entry point beyond training.

**Commercialization Shifts From Selling Cards to Delivering Systems**

The half-year reports show that Chinese AI chips have formed a clear commercial path: cloud products provide standardized revenue, cluster delivery amplifies per-order value, industry ecosystem expands customer reach, and software adaptation plus ongoing services determine repurchase depth.

Cloud products are the starting point for revenue scale. Cambricon's cloud product line generated 5.994 billion yuan in H1, representing 99.98% of total revenue, with training and inference chips concentrated on large-model infrastructure. Hygon DCUs have achieved compatibility with over 400 mainstream large models, leveraging CPU-DCU synergy to penetrate cloud service providers and intelligent computing centers.

Moore Threads has extended platformization further into clusters. The company's H1 cloud product revenue reached approximately 1.69 billion yuan, 97.49% of total revenue, with MTT S5000 volume production and Kuaer intelligent computing cluster delivery as growth drivers. The 660 million yuan Kuaer cluster sales contract disclosed in March was fully delivered and revenue-recognized in H1; Kuaer clusters are now deployed in Beijing, Wuxi, Hangzhou, and elsewhere. On September 9, JD Cloud announced partnerships with Moore Threads and others to build a 10,000-card domestic cluster and plan a 100,000-card fully functional GPU cluster—domestic GPU system delivery continues to scale upward from ten-thousand-card projects.

Moxin proposed a "1+6+X" industry ecosystem strategy, centered on its autonomous GPU computing platform and MXMACA software stack, covering six sectors—finance, healthcare, energy, education and research, transportation, and entertainment—while extending into embodied AI and low-altitude economy scenarios. The fully domestic manufacturing process Xiyun C600 achieved volume production in May; the Xijing S600 super node targeting training, inference, and intelligent computing center construction is progressing in parallel, with Moxin GPUs commercialized in thousand-card-scale clusters.

Biren has pursued cloud training paired with high-speed interconnection. Its R100 series targets intelligent computing centers, operators, and large-model customers; H1 revenue of 12.36 billion yuan was driven by training product ramp, lifting gross margin to 42.7%. In partnership with Shanghai Yidian, Xizhi Technology, and ZTE, it launched the commercial "Light Leap" optical interconnection super node, now deployed at thousand-card scale, and unveiled next-generation optical interconnection super node plans in July—product form has expanded from training chips to large-scale interconnection systems.

Enflame has centered on training and inference chips plus cluster solutions serving internet customers like Tencent. Its products have expanded from small-scale pilots to multiple scenarios within Tencent. Tencent serves both as a major shareholder and Enflame's largest customer; revenue from Tencent via direct sales and related models accounted for 83.79% in 2025, while some non-affiliated tier-one internet customers have completed hardware and model testing and begun small-volume orders.

**Conclusion**

The current market strength has provided Chinese AI chips a collective window for volume ramp. When supply is tight, customers primarily need availability; as capacity recovers and chip prices and compute service pricing are renegotiated, customers will shift comparison dimensions toward utilization, power efficiency, software upgrades, and payment cycles.

China's Ministry of Industry and Information Technology has issued the Information Communication Industry's "15th Five-Year Plan," proposing enhanced computing infrastructure development levels and deepened cross-regional computing coordination capabilities. By 2030, China's intelligent computing scale is projected to grow more than fivefold from 2025 levels, with orderly deployment of ten-thousand-card, hundred-thousand-card, and larger intelligent computing clusters, demand-driven inference infrastructure deployment aligned with scenarios, and accelerated domestic computing chip adaptation.

Chinese AI chips are transitioning from "selling one card" to "delivering a system, serving an industry, and operating an ecosystem." Public listings and fundraising simply buy time. The next ranking will not be determined by who has the fastest revenue growth alone. Companies that weather computing cycles will be those that can replicate single projects into clusters and sediment clusters into stable industry workloads.