China’s Four AI-Chip Challengers: Revenue, Technology and Profitability in 2026

China’s domestic AI-chip industry has moved from policy narrative to measurable commercial traction. In the first half of 2026, Moore Threads, MetaX, Biren and Enflame all reported substantial revenue or revenue growth, helped by demand for locally controlled computing capacity amid U.S. export restrictions and Chinese localization policies.

The more important conclusion is less dramatic: sales momentum has arrived before dependable operating profitability. MetaX’s headline profit was investment-gain-driven, Moore Threads was close to break-even but still loss-making on an adjusted basis, and Biren and Enflame remained in the red. The four companies are also pursuing two different technology strategies: CUDA-compatible general-purpose GPUs on one side, and Enflame’s domain-specific, full-stack approach on the other.

First-half 2026: Moore Threads leads sales, Biren leads growth

The reported figures point to four companies at different stages of commercialization:

• Moore Threads: Revenue reached RMB 1.736 billion, up 147.42% year over year and above its RMB 1.506 billion revenue for all of 2025. Its attributable net loss narrowed to RMB 11.56 million, but adjusted net profit remained negative at roughly RMB 151 million. The company is therefore closest to operating break-even, not yet proven sustainably profitable. 51821

• MetaX: Revenue was RMB 1.324 billion, up 44.67%. The company reported attributable net profit of RMB 612 million, but RMB 887 million of fair-value gains on financial investments accounted for more than the reported profit. Adjusted net profit was still a loss of approximately RMB 49 million, making the distinction between accounting profit and core GPU profitability essential. 1718

• Biren: Revenue reached RMB 1.236 billion, a 1,997.6% increase from a very low prior-year base. It nevertheless reported a loss of approximately RMB 377 million, with an adjusted loss of about RMB 337 million. The numbers show rapid scaling, but not a completed transition to profitable operations. 181922

• Enflame: First-half revenue reached RMB 1.120 billion—reported in Chinese financial coverage as RMB 11.20 hundred million—while its prospectus projected an adjusted first-half loss of RMB 582 million to RMB 614 million. Enflame’s revenue growth is significant, but its losses remain much larger than those of the other three companies in absolute terms. 21922

These figures are not perfectly like-for-like: companies may recognize revenue from chips, boards, systems, clusters and related services differently. Even so, the ranking is informative. Moore Threads had the largest reported first-half sales, Biren had the fastest percentage growth, and MetaX’s reported profit requires the greatest adjustment before it can be interpreted as evidence about the chip business itself. 18

MetaX’s reported profit is not the same as operating profitability

MetaX illustrates why headline net income can give an incomplete picture of an AI-chip company.

Its RMB 612 million attributable profit looks like a sharp turnaround. But the reported result included RMB 887 million in fair-value gains from financial investments. Once non-recurring items are excluded, MetaX recorded an adjusted loss of roughly RMB 49 million. In practical terms, the company sold enough products to generate meaningful revenue, but the available figures do not show that its GPU operations were already producing durable profits. 1718

The distinction matters across the sector. Semiconductor companies spend heavily before and during commercialization, and one-off investment gains, government support or other non-operating items can temporarily change the appearance of the income statement. Core margins, recurring customer orders and cash generation provide a more useful test of whether the business model is working.

Two technology strategies: CUDA compatibility versus a specialized stack

The four companies are often grouped together as China’s leading domestic GPU challengers, but they are not making the same architectural bet.

MetaX, Moore Threads and Biren: make migration easier

MetaX, Moore Threads and Biren are associated with the general-purpose GPU route and emphasize compatibility with the CUDA software ecosystem. The goal is to reduce the cost of moving existing AI workloads away from Nvidia hardware: developers can preserve more of their familiar programming model, frameworks and operators rather than rebuilding an entire software stack. 11

Moore Threads’ MUSA platform is described as highly compatible with CUDA and is intended to lower migration friction for developers. That approach gives a company a broader potential workload base than a narrowly specialized accelerator, although compatibility is not the same as identical performance, reliability or ecosystem maturity. 11

Enflame: a domain-specific, full-stack alternative

Enflame is the outlier. Rather than closely reproducing a CUDA-style general-purpose GPU, it pursues a domain-specific architecture (DSA) focused on AI training and inference. Its product strategy combines Suisi processors, accelerator cards and modules, computing clusters and the TopsRider software platform. 12

TopsRider is intended to provide the software layer needed to make Enflame hardware usable in production, including drivers, compilers and operator libraries. Enflame’s proposition is therefore not simply a chip specification; it is an integrated hardware-and-software platform for cloud and data-center workloads.

The trade-off is straightforward:

• CUDA-oriented GPUs can reduce switching costs and support a wider range of workloads.
• A specialized DSA stack may offer tighter hardware-software optimization and greater control over the platform.
• The DSA route also asks customers to adopt another toolchain, making software quality, model portability and long-term support especially important. 1112

No single architecture has won this contest yet. The commercial test is whether each company can turn its software strategy into repeat deployments, stable cluster performance and lower total operating cost for customers.

Enflame’s September 2 IPO is a financing and commercialization test

Enflame planned to open subscriptions on September 2, 2026, for a STAR Market offering targeting approximately RMB 6 billion—about $892 million according to Reuters—with the planned stock code 688801. The proceeds are intended to support later-generation AI chips and hardware-software co-innovation. 136

The listing gives Enflame capital to continue its chip roadmap and expand commercialization. It also raises a central strategic question: can a company whose platform was closely tied to Tencent broaden its customer base?

Tencent contributed 74.9% of Enflame’s RMB 990 million 2025 revenue according to one report. Another source gives a higher figure of 83.79%, indicating that the exact percentage may depend on reporting definitions or the period analyzed. Either way, the available evidence shows substantial customer concentration. 1012

That concentration is not automatically a weakness—an anchor customer can help validate products and finance early scale—but it makes independent-market demand one of Enflame’s most important milestones after listing.

China’s market is already large enough to support localization

China shipped roughly 4 million AI-accelerator cards for servers in 2025. Domestic GPU and AI-chip suppliers collectively shipped about 1.65 million, or approximately 41% of the market, while Nvidia remained the largest individual supplier with about 2.2 million cards and a roughly 55% share.

The figures show two things at once. First, domestic suppliers have achieved substantial adoption rather than merely announcing prototypes. Second, the domestic market is not yet equivalent to a fully displaced Nvidia market: foreign suppliers still held a large position in the 2025 shipment data, and the domestic share was collective rather than attributable solely to these four companies.

U.S. export controls and Chinese procurement preferences have accelerated the shift. Restrictions have reduced access to Nvidia’s most advanced products, while localization policies have encouraged Chinese cloud providers and AI companies to validate domestic alternatives. Forecasts cited in industry coverage have suggested that domestic accelerators could eventually supply as much as 90% of China’s market, but that is a forecast—not a verified outcome—and Huawei and Cambricon are identified as likely major beneficiaries.

Deployment, however, involves more than replacing one card with another. Customers must address model porting, compiler and operator support, interconnects, cluster-level performance, supply reliability and production stability. Domestic shipment share is an important milestone, but it does not by itself establish parity with Nvidia across the full AI-computing stack.

The 2028 outlook is positive, but a precise forecast is not established here

The available evidence supports a strong multi-year growth direction for China’s AI-accelerator market. Export controls, state-backed localization, data-center investment and the spread of AI-model deployment all support demand for domestic hardware. 17

But the provided evidence does not establish a reliable, specific market-size forecast for 2028. It would be misleading to assign a precise 2028 figure on that basis. The better-supported conclusion is directional: the market is likely to keep expanding, while the distribution of that growth among Huawei, Cambricon, the four companies discussed here and other domestic suppliers remains uncertain.

Why rapid revenue growth has not solved the profitability problem

AI-chip companies face unusually high fixed costs. R&D includes chip architecture, repeated tape-outs, software engineering, compilers, libraries, systems validation and customer support. Moore Threads’ 2025 R&D spending was reported at 87% of revenue, illustrating how quickly development costs can absorb sales during the scale-up phase. 9

A successful product therefore needs more than competitive silicon. It needs:

1. Useful performance on real customer workloads rather than only peak specifications.
2. A mature software ecosystem that supports frameworks, operators, compilers and debugging.
3. Reliable supply and system integration at the board, server and cluster levels.
4. Enough recurring orders to spread R&D and manufacturing-related fixed costs.
5. A broad customer base that reduces dependence on a single buyer or procurement cycle.

The first-half results show progress on demand, but not yet a durable moat. Moore Threads appears closest to operating break-even, while MetaX’s reported profit was primarily investment-gain-driven. Biren and Enflame remained loss-making despite rapid revenue expansion. 18192122

Bottom line

China’s four leading domestic AI-chip challengers have reached a meaningful commercialization stage. Moore Threads is the current revenue leader, Biren is scaling fastest from a small base, MetaX has strong reported earnings but weak adjusted profitability, and Enflame is combining a major IPO with the most differentiated non-CUDA architecture of the group.

The strategic opportunity is real: domestic suppliers now account for about 41% of China’s AI-accelerator server market, and export controls have created powerful incentives to adopt local alternatives. But the industrial race is not decided by revenue growth alone. The winners will be the companies that convert policy-driven demand into repeat orders, software adoption, reliable clusters and recurring operating profit.