AnswersPublished28 sources
China’s GPU Four Tigers Are Splitting by Workload, Customers and Software
China’s GPU four tigers are no longer interchangeable NVIDIA alternatives: Moore Threads is the broadest full function GPU platform, Biren is focused on high end data centre GPGPU, Enflame is Tencent led, and MetaX is... The decisive contest is software, not just silicon: customers must be able to migrate models, ac...
Published byEdited with GPT-5.6 LunaImages generated with GPT Image 1.5
Research answer
Key takeaways
- China’s GPU four tigers are no longer interchangeable NVIDIA alternatives: Moore Threads is the broadest full function GPU platform, Biren is focused on high end data centre GPGPU, Enflame is Tencent led, and MetaX is...
- The decisive contest is software, not just silicon: customers must be able to migrate models, access mature kernels and operate distributed clusters without NVIDIA’s CUDA ecosystem.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: How are China’s four leading domestic GPU makers—Moore Threads, Biren Technology, Enflame, and MetaX—entering a differentiation phase after. Article summary: China’s GPU “four tigers” are no longer pursuing one undifferentiated “domestic NVIDIA” story. After listings or IPO milestones, they are separating by workload, customer access, and software strategy: Moore Threads is t. Topic tags: general, news, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts wi
China’s leading domestic GPU makers are entering a more differentiated phase. Public-market milestones have made their commercial performance easier to compare, but they have not produced four equivalent “domestic NVIDIA” businesses. Instead, the companies are separating by workload, customer access, software architecture and the speed at which revenue can become sustainable profit.
The four companies at a glance
| Company | Main position | Commercial signal | Principal question |
|---|---|---|---|
| Moore Threads | Broad, full-function GPU platform spanning AI and graphics | Revenue reached about RMB1.51 billion in 2025, while first-half 2026 revenue rose 147.42% year on year to RMB1.736 billion. |
Can its broad product range and MUSA software stack translate into repeatable, profitable cluster shipments? |
| Biren Technology | High-performance data-centre GPGPU for training and inference | FY2025 revenue rose 207.2% to RMB1.0346 billion, with a 53.8% gross margin. |
Can it scale production and software support for large AI deployments? |
| Enflame | Cloud AI accelerator and full-stack systems supplier | Revenue grew to RMB990 million in 2025, but Tencent accounted for about 74.9% of it. |
Can it diversify beyond its powerful but concentrated anchor customer? |
| MetaX | Data-centre GPU and computing-platform business | 2025 revenue more than doubled to about RMB1.6 billion, and first-half 2026 net profit reached RMB612 million. |
Is its early profitability repeatable in a capital-intensive chip cycle? |
Moore Threads: the broadest platform bet
Moore Threads has the widest strategic addressable market among the four. Its product and software approach covers AI computing alongside graphics and visual workloads, rather than concentrating only on a specialized AI accelerator niche. The company’s MUSA architecture is intended to support products across computing and graphics use cases. 38
That breadth gives Moore Threads a distinctive potential role in rendering, video, simulation and other workloads where AI and visual computing overlap. It also creates a heavier execution burden: a broad platform must support several customer types while building enough software depth for data-centre deployment.
The financial trajectory shows rapid commercialization but continuing investment pressure. Moore Threads reported approximately RMB1.51 billion in 2025 revenue against a net loss of about RMB1.0 billion, according to a review of its annual-report figures. 15 In the first half of 2026, revenue reached RMB1.736 billion, up 147.42% year on year, while attributable net loss narrowed sharply.
32
33
Likely role: enterprise and cloud AI inference, general-purpose computing clusters and the quartet’s clearest graphics opportunity. Large-scale training remains an ambition that depends heavily on software and system execution.
Biren: the clearest high-end GPGPU contender
Biren is the most explicitly data-centre-focused of the four. It sells GPGPU-based intelligent-computing solutions for AI training and inference across cloud and edge environments. Its BIRENSUPA platform connects the company’s hardware with AI applications, frameworks, libraries, programming tools and cluster-management capabilities. 1
12
Its reported commercial growth is substantial. FY2025 revenue rose 207.2% to RMB1.0346 billion, with gross profit of RMB557 million and a 53.8% gross margin. The company attributed the increase primarily to volume production and shipments of its BR10X series. 4
6
Biren also projected first-half 2026 revenue of RMB1.15 billion to RMB1.30 billion, compared with approximately RMB58.9 million in the same period of 2025. Its projected adjusted loss was RMB290 million to RMB360 million. These are company forecasts rather than audited full-period results, so they should be treated accordingly. 3
The next test is whether Biren can convert strong shipment growth into dependable, large-scale deployments. Its planned BR20X successor was expected to launch commercially in 2026, with the company describing additional support for lower-precision AI workloads and a super-node design. 2
Likely role: large-scale AI training and high-performance inference, particularly where customers prioritize domestic supply and can absorb the cost of migrating software.
Enflame: a Tencent-anchored AI accelerator strategy
Enflame is more purpose-built around cloud AI acceleration than around a conventional all-purpose GPU and graphics business. Its product portfolio spans AI chips, accelerator cards and modules, intelligent-computing systems and clusters, as well as programming software. 46
49
The company’s commercial position is unusually dependent on Tencent. Enflame’s revenue rose from RMB301 million in 2023 to RMB722 million in 2024 and RMB990 million in 2025, while adjusted net losses narrowed from RMB1.567 billion to RMB1.197 billion over the same period. 50 However, Tencent contributed approximately 74.9% of 2025 revenue.
48
That relationship provides a substantial customer and deployment base, but it also defines Enflame’s central business risk. The company must show that its software and hardware can attract customers beyond Tencent, rather than relying on one anchor ecosystem for most demand.
Enflame’s software strategy is also different from a direct attempt to recreate a universal GPU platform. Available reporting describes its proprietary software ecosystem, including TopsRider, as part of a vertically integrated alternative to NVIDIA’s CUDA-based model. 54 Whether that approach can build an independent developer and customer base remains less established than the revenue relationship with Tencent.
Likely role: Tencent-linked and cloud-based AI training and inference. Graphics and rendering are not its natural center of gravity.
MetaX: the strongest early profitability signal
MetaX focuses on high-performance GPU and computing platforms for data-centre AI rather than consumer graphics. Its 2025 revenue rose 121.3% to about RMB1.6 billion, although the company still reported an adjusted net loss of approximately RMB830 million for that year. 24
The most striking change came in the first half of 2026. MetaX reported RMB612 million in net profit, reversing a RMB186 million loss in the comparable period, while first-half revenue rose 44.7% to RMB1.32 billion. 18
That makes MetaX the quartet’s clearest profitability outlier in the available figures. It is an important commercial signal, but it should not be treated as proof that MetaX has permanently solved the economics of GPU development. Semiconductor businesses can show volatile results as shipments, product mix, subsidies, customer acceptance and development costs change.
MetaX’s strategic opportunity is to make migration from existing AI software practical and production deployments reliable. Its strongest positioning is therefore likely to come from execution: shipping usable data-centre systems, supporting customers after deployment and improving the economics of inference.
Likely role: enterprise and cloud AI training and inference, with profitability and deployment efficiency as its differentiators.
Software—not “CUDA compatibility”—is the real battleground
None of these companies owns NVIDIA’s CUDA ecosystem. The relevant comparison is not whether a vendor can claim compatibility in the abstract, but whether customers can move real workloads with acceptable engineering risk.
That requires support for common frameworks, mature operators and libraries, stable drivers, distributed-training tools, debugging, performance tuning and dependable production inference. The supplied evidence does not support treating any of the four as being at software parity with NVIDIA.
The companies are taking different routes:
- Moore Threads is using MUSA to unify AI and graphics across a broader GPU platform. Reporting on the company describes MUSA as designed with CUDA compatibility in mind, while its developer ecosystem has reportedly surpassed 800,000 developers.
38
32
- Biren is presenting BIRENSUPA as an integrated hardware-and-software platform for training, inference and large GPGPU clusters.
12
- Enflame is pursuing a more vertically integrated AI-accelerator stack through its own software ecosystem rather than focusing on a general-purpose graphics platform.
49
54
- MetaX has the opportunity to compete through practical migration, reliable deployment and lower operating friction, although the available sources provide less evidence for declaring software leadership.
For buyers, the key metric is therefore switching cost. A chip that is theoretically competitive but difficult to port, tune or operate may lose to an inferior chip with better tools and support.
NVIDIA and Huawei: why a four-way replacement race is misleading
Against NVIDIA, the domestic vendors can compete where supply assurance, local support, policy alignment and workload-specific integration matter more than absolute frontier performance. NVIDIA’s established CUDA ecosystem remains a major advantage, and the available evidence does not establish comparable software maturity for any of the four.
Huawei also matters in the domestic competitive picture, but the supplied sources do not provide sufficiently comparable figures on Huawei’s market share, performance or customer mix to rank it precisely against these startups. The safer conclusion is that China’s market will not divide neatly among four equal replacements for NVIDIA. Huawei, NVIDIA where available, the four companies and other domestic suppliers will compete across different procurement and workload contexts.
What the market-share evidence can—and cannot—show
Precise, comparable market-share percentages for Moore Threads, Biren, Enflame and MetaX are not established by the available evidence. Companies report different measures, including chips, cards, systems, clusters and revenue, while private deployments are difficult to verify.
The wider field is also larger than these four. Cambricon, for example, reported RMB6.497 billion in 2025 revenue and RMB2.059 billion in net profit, showing that the domestic competitive landscape includes established listed suppliers beyond the so-called four tigers. 22
Forecasts that China’s AI-accelerator market could exceed RMB1 trillion by 2028, with domestic solutions approaching 90%, should therefore be treated as scenarios rather than current market-share data. The supplied sources do not establish the methodology behind those forecasts.
If domestic adoption does reach that level, the most plausible outcome is workload segmentation rather than four equal NVIDIA substitutes:
- Biren, alongside major domestic systems suppliers, is likely to face the most direct challenge in large-scale training.
- MetaX, Moore Threads and Enflame can compete for different mixes of enterprise and cloud inference.
- Moore Threads has the clearest dedicated opportunity in graphics and visual computing.
- Enflame must prove that its Tencent-led demand can broaden into a durable external customer base.
- MetaX must show that its early profitability can persist.
The strategic question has changed. It is no longer which company will become “China’s NVIDIA.” It is which vendor can combine dependable supply, a focused customer wedge and a software stack strong enough to reduce the cost and operational risk of leaving CUDA.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "China’s GPU Four Tigers Are Splitting by Workload, Customers and Software"?
China’s GPU four tigers are no longer interchangeable NVIDIA alternatives: Moore Threads is the broadest full function GPU platform, Biren is focused on high end data centre GPGPU, Enflame is Tencent led, and MetaX is...
What are the key points to validate first?
China’s GPU four tigers are no longer interchangeable NVIDIA alternatives: Moore Threads is the broadest full function GPU platform, Biren is focused on high end data centre GPGPU, Enflame is Tencent led, and MetaX is... The decisive contest is software, not just silicon: customers must be able to migrate models, access mature kernels and operate distributed clusters without NVIDIA’s CUDA ecosystem.