Battle for the “China’s NVIDIA” Crown: Domestic GPUs Surge Toward 50% Market Share—Which of the Four Rising Champions Will Seize the Opportunity?

After the spotlight-grabbing IPO debuts, two of China’s best-known homegrown GPU names—Moore Threads (688795.SH) and MetaX (688802.SH)—are now feeling the full force of public-market volatility.

Over the past few trading sessions, both stocks have pulled back noticeably from their eye-popping early surges. As the initial excitement fades, investors and industry watchers are shifting from celebration to closer scrutiny.

In December 2025, China’s computing-chip sector truly entered what many are calling a “historic moment.” In less than two weeks, Moore Threads and MetaX listed on the A-share market. Soon after, Biren Technology and Iluvatar CoreX passed their IPO hearings in Hong Kong. With that, the much-discussed “Domestic GPU Four”—long viewed as unicorn-level contenders—have collectively stepped into a new phase where the capital markets demand real proof.

But this is more of a milestone than a finish line.

With Nvidia’s market value remaining among the world’s highest, and AI compute demand gradually shifting from large-model training toward inference and commercial deployment, going public doesn’t reduce competitive pressure. If anything, it accelerates differentiation. After listing and hearings, these companies are placed into a direct comparison framework: technology maturity, product stability, ecosystem completeness, and sustained delivery capability will all be tested repeatedly by the market.

Sullivan China consulting advisor Chi Yu noted that, in terms of industry lifecycle, domestic GPUs are still in an early, fast-growth stage. Even the front-runners, he said, remain meaningfully behind established global players like Nvidia.

As the halo of capital cools, there’s still no unified answer to the biggest question: which of the “Domestic GPU Four” can truly break away from the pack?

That uncertainty comes largely from how different their paths are.

Biren Technology is oriented toward high-compute, priority industries, and its OCS (optical circuit switch) approach has already entered real-world deployment. Iluvatar CoreX has reduced its earlier dependence on a small number of major clients, with customer concentration dropping quickly. Moore Threads is pursuing a full-function GPU strategy, aiming to cover everything from gaming graphics cards to AI compute. MetaX, meanwhile, has moved first on profitability—its narrative reads less like a high-concept disruptor and more like a disciplined, pragmatic domestic substitution player.

Even as the competitive pressure intensifies, the market opportunity is being reshaped—and expanded.

A Shanxi Securities research note argues that, with capital-market support, 2026 will be a period of rapid supply-side expansion for domestic GPUs. On the demand side, requirements are becoming clearer: telecom operators, “sovereign AI,” and AI+ initiatives led by central and state-owned enterprises. On the inference side, internet companies and large-model developers are also expected to accelerate domestic GPU adoption.

Frost & Sullivan data suggests that, as shipments from China’s general-purpose GPU companies grow faster than international competitors, domestic market share will keep rising. The share of domestically produced general-purpose GPU products increased from 8.3% in 2022 to 17.4% in 2024, and is projected to exceed 50% by 2029.

What sets the four apart?

Founded in 2015, Iluvatar CoreX is the oldest—and the most low-profile—among the four. Its prospectus states that it was the first in China to mass-produce a general-purpose GPU chip capable of both training and inference. Its 2021 TianGai Gen1 was described as China’s first mass-produced general-purpose GPU product.

Iluvatar’s defining trait is breadth and stability. Its core customers span cloud service providers, AI model developers, research institutions, and a wide mix of sectors including electronics, semiconductors, manufacturing, and consumer internet. From 2022 to 2024, its customer count rose from 22 to 181.

On the supply chain side, the company says it works actively with a diversified set of qualified suppliers, aiming—where feasible—to maintain at least one backup option for each key supplier category. For specialized components with high technical barriers and limited alternatives, it claims to have built particularly strong partnerships with leading vendors, supported by enhanced monitoring systems to ensure supply stability.

IDC data shows that in China’s 2024 GPU market, Nvidia shipped 1.9 million units, Huawei Ascend 640,000, Kunlunxin 69,000, Iluvatar CoreX 38,000, Cambricon 26,000, MetaX 24,000, and Enflame 13,000. Among the “Domestic GPU Four,” Iluvatar ranked first by shipments.

Biren Technology was founded in Shanghai in September 2019 by Zhang Wen, former president of SenseTime—making it the second-oldest in the group. Biren is often characterized by its emphasis on talent and advanced engineering execution.

Zhang Wen has said in interviews that China may have fewer than 500 truly top-tier GPU experts—highlighting the scarcity of high-end semiconductor talent—and that Biren’s density of senior talent is among the highest in the country.

Technically, Biren has applied advanced capabilities such as high-speed interconnects, high-bandwidth memory, and chiplet packaging into its products and advanced packaging. According to China Insights Consultancy, Biren is among the first Chinese GPGPU companies to use PCIe Gen5, CXL, high-performance DRAM, and dual-die chiplet designs in commercial products.

In 2024, Biren reported revenue of RMB 337 million from 14 customers. In terms of orders, it disclosed that as of December 15 it held five framework sales agreements worth RMB 1.241 billion, plus 24 sales contracts.

MetaX and Moore Threads were both founded in 2020, but they have taken notably different routes.

MetaX launched its first training-and-inference integrated GPU chip, the XiYun C500, in 2023, and then rolled out the XiYun C550 and XiYun C588. A version based on a domestic supply chain—the XiYun C600—has reportedly reached the “tape-out and bring-up” stage (a key R&D milestone). As of September 5, 2025, MetaX disclosed outstanding orders (excluding tax) totaling RMB 1.43 billion, largely for XiYun C500-series cards.

Commercially, MetaX has broken through the “persistent losses” pattern. From January to September 2025, progress in its integrated training-and-inference product line accelerated, shipments increased sharply, and revenue rose significantly year over year. As revenue ramped quickly, losses narrowed. In particular, in April–June 2025, MetaX reported revenue of RMB 595 million (up 237.87% year over year) and net profit attributable to shareholders of RMB 46.62 million—achieving single-quarter profitability.

Moore Threads, though the youngest, was the earliest among the four to reach an A-share listing. Its prospectus positions it—alongside Nvidia—as a representative of “full-function GPUs,” emphasizing completeness in functionality and computing precision. It is widely viewed as China’s only company to achieve mass production and mass commercialization of full-function GPUs, spanning chips, cards, clusters, and a full software ecosystem.

On December 20, at the inaugural MUSA Developer Conference, Moore Threads unveiled its new “Huagang” architecture. The company said it introduces a next-generation instruction set, lifts compute density by 50%, boosts energy efficiency by 10x, and supports full-precision computing from FP4 to FP64. Based on this architecture, it plans to launch an integrated AI training-and-inference card called “Huashan,” as well as a graphics rendering card called “Lushan.”

Challenging Nvidia’s edge

In raw single-card performance, Nvidia still holds the advantage. However, efficient interconnect at the “supernode cluster” level—and system-level optimization—could weaken Nvidia’s dominance in large-scale data centers and cloud providers.

Shanxi Securities argues that supernode servers are among the most important beneficiary segments of domestic GPU acceleration.

Iluvatar CoreX’s prospectus notes that the exponential growth of AI model parameters has made traditional single-GPU cards or small compute clusters insufficient for efficient training and inference. China’s AI infrastructure is rapidly shifting toward ultra-large-scale general-purpose GPU clusters to support trillion-parameter model training—pushing optimization toward massive parallelism, ultra-high bandwidth, and higher system-level stability.

Within supernode architecture, optical interconnect has emerged as a mainstream direction. Chi Yu explained that there are currently two main implementation paths: one approach uses optics for high-speed transmission while electronic systems handle intermediate switching and scheduling; the other aims for optical switching, building end-to-end connections at the optical layer to reduce reliance on intermediate electronic processing.

Among the “Domestic GPU Four,” Biren has chosen the optical-interconnect optical-switching route. On July 28, Shanghai Electric, Xizhi Technology, Biren Technology, and ZTE jointly launched LightSphere X, described as China’s first optical-interconnect optical-switching GPU supernode. The system is based on Xizhi’s distributed optical switching technology, combining silicon photonics with Biren’s proprietary high-compute general-purpose GPU liquid-cooled modules and a new carrier-board interconnect.

MetaX, at the World Artificial Intelligence Conference, showcased optical interconnect capabilities, the Yaolong 3D Mesh, the Shanghai Cube domestically built high-density liquid-cooled cabinet, and a high-density liquid-cooled compute POD. A collaboration between Xizhi and MetaX also presented an optical-interconnect electronic-switching supernode solution featuring low latency, high bandwidth, and low power consumption, with support for long-distance transmission.

Chi Yu believes optical-interconnect optical-switching remains in an early stage of industrial rollout and technological evolution. For now, solutions centered on electronic processing remain mainstream, while optical switching is seen as the next major direction. As compute systems continue scaling, the energy-efficiency and expansion limits of architectures that depend on intermediate processing will intensify—making a transition from optical-interconnect electronic-switching toward optical-interconnect optical-switching increasingly likely.

Compared with Biren and MetaX, Iluvatar has not publicly detailed a specific supernode product plan. Its prospectus says it has developed a comprehensive infrastructure solution for large AI deployments, enabling AI training or inference applications built in mainstream ecosystems to migrate quickly while maintaining consistent model accuracy.

Still, Iluvatar has already shown progress in cluster optimization. An article published via the Shanghai Municipal Commission of Economy and Informatization’s official account states that Wuwenchiong and Iluvatar collaborated to achieve a significant performance leap in MoE large-model training on a 150,000-card TianGai cluster—reaching a level comparable to leading international standards.

Moore Threads, meanwhile, showcased its MTT C256 supernode plan at the MUSA Developer Conference. The design aims to interconnect 256 GPUs with a single network layer, avoiding bandwidth loss and added latency commonly introduced by multi-layer (two or more) networks.

As AI compute competition shifts from “single-card performance” to “supernode clusters,” a clearer window of opportunity opens for China’s GPU challengers.

Chi Yu emphasized that domestic GPU companies are currently in a rapid development and exploration phase, where technical routes and application scenarios are not yet fully settled. In this environment, any company that can deliver meaningful breakthroughs in core technologies and validate a workable commercial model in real deployments can win market and capital recognition quickly—potentially jumping into the leading tier.

Huang Fusheng, vice president and chief economist of China Post Securities, added that in 2026 the market will pay closer attention to tangible product launches and real application rollouts in Chinese tech stocks. He expects sharp differentiation, noting that the “concept-driven broad rally” pricing logic seen in 2025 will be difficult to sustain.