With strong expectations for the approaching inflection point of Agentic AI, domestic GPU enterprises have become the darlings of capital markets in 2026.

On January 8, 2026, the Hong Kong stock market welcomed its second GPU enterprise, Iluvatar CoreX (09903.HK). Since its listing, the company's share price has trended upward amidst fluctuations. As of the close on March 10, 2026, the company was quoted at HK$338.20, hitting a new record high and achieving a cumulative gain that has more than doubled.

Full-Stack Self-Research: Constructing a "Cloud + Edge + Terminal" Full-Scenario Computing Layout

Compared to the more widely known "Four Little Dragons" of domestic GPUs, Iluvatar CoreX was established earlier.

Public records show that Iluvatar CoreX was founded in Shanghai in December 2015. It is the earliest established general-purpose GPU (GPGPU) chip enterprise in China, primarily dedicated to creating AI computing solutions characterized by high performance, high versatility, and ease of migration.

Following nearly a decade of development, Iluvatar CoreX has formed a general-purpose GPU product matrix centered on the "Tiangai" (Training) series and the "Zhikun" (Inference) series. This covers diverse computing needs ranging from complex model training to edge inference, with wide applications in key industries such as cloud computing, financial services, intelligent manufacturing, and healthcare.

NewTimeSpace has learned that since the formal launch of its first general-purpose GPU product in 2021—the first-generation training series, Tiangai Gen 1—the company's chip R&D appears to have entered an accelerated phase.

According to a report by Frost & Sullivan, among Chinese chip design companies, Iluvatar CoreX is the first to achieve mass production of both inference and training GPGPU chips, and the first to achieve this using 7nm process technology.

In late January 2026, Iluvatar CoreX released its fourth-generation architecture roadmap, proposing a "high-quality computing" design goal centered on "efficiency, predictability, and sustainability" to build an AI++ computing system. Among these, the "Tianshu" architecture released in 2025 has already surpassed Nvidia’s Hopper architecture, supporting everything from high-precision scientific computing to AI-precision computing. When executing calculations related to the Attention mechanism, the actual effective utilization rate of the AI chips exceeds 90%.

The company plans for the "Tianquan" architecture in 2027 to surpass Nvidia’s Rubin architecture, aiming for a breakthrough in innovation. Additionally, the company released its edge computing product line, the "Tongyang" series. In the field of edge computing, the Tongyang TY1000 has demonstrated performance superior to Nvidia’s AGX Orin in several real-world AI scenario tests. The "Tongyang" series completes the final piece of Iluvatar CoreX’s "Cloud + Edge + Terminal" full-scenario computing layout.

Persistent Losses of Nearly 3 Billion RMB Amidst High R&D; GPU Revenue Exceeds 80%

Currently, the GPU industry requires massive investment, and like its peers, Iluvatar CoreX has yet to escape its loss-making state.

NewTimeSpace understands that between 2022 and 2024, Iluvatar CoreX achieved revenues of RMB 189 million, RMB 289 million (+52.91%), and RMB 540 million (+86.85%), respectively, representing a compound annual growth rate (CAGR) of 68.8%. Revenue for the first half of 2025 reached RMB 324 million, already accounting for 60% of the total 2024 revenue.

Although the company has maintained high-speed revenue growth, it recorded net losses of RMB 554 million, RMB 817 million, and RMB 892 million between 2022 and 2024. As of the first half of 2025, the company recorded a loss of RMB 609 million, with cumulative losses exceeding RMB 2.8 billion.

Regarding the reasons for these losses, Iluvatar CoreX stated that they are related to sustained high R&D investment, limited economies of scale, being in the early stages of commercialization, and increased operating expenses such as sales and administration.

Data shows that during the 2022–2024 period, the company's R&D expenses were RMB 457 million, RMB 616 million, and RMB 773 million, accounting for 241.1%, 213.1%, and 143.2% of total revenue for the respective periods. In the first half of 2025, R&D expenditure was RMB 451 million, representing 139.2% of total revenue.

Behind the rapid revenue growth, as of the first half of 2025, Iluvatar CoreX had cumulatively delivered over 52,000 GPGPU units, with products deployed more than 900 times across over 290 customers. Compared to just 22 customers in 2022, the customer base has achieved an order-of-magnitude leap.

Furthermore, it is worth noting that the revenue share from GPGPU products was 99.6% in 2022 and shifted to 85.3% in the first half of 2025, indicating a relatively singular product revenue structure. Additionally, revenue from AI computing solutions was RMB 166 million in 2024 but only RMB 42.64 million in H1 2025—a 28.7% year-on-year decline—highlighting the instability of project-based income.

Industry Enters Golden Window; Single-Day Surge Exceeds 30%

From the perspective of industry progression, Iluvatar CoreX stands at the intersection of demand and policy tailwinds.

Currently, the focus of AI development is accelerating from the "training phase" to the "application phase." Deloitte pointed out in its 2026 Technology, Media, and Telecommunications Industry Forecast that by 2026, inference computing will account for 66% of overall AI computing, surpassing training computing for the first time.

This means that inference has transformed from a technical secondary role to the core driving force for AI commercialization. This trend is highly beneficial for domestic GPU manufacturers. Since inference scenarios focus more on video memory size, bandwidth, and INT8/FP8 computing power, the demand for extreme raw performance is slightly reduced. In this regard, the gap between domestic GPUs and Nvidia has narrowed significantly.

NewTimeSpace has learned that the Ministry of Industry and Information Technology (MIIT), in its 2025-2026 Stable Growth Action Plan for the Electronic Information Manufacturing Industry, called for increased efforts in tackling key technologies like CPUs and hardware-software synergy.

Meanwhile, eight departments, including the MIIT, the Cyberspace Administration of China, and the National Development and Reform Commission, jointly issued the Implementation Opinions on the "AI + Manufacturing" Special Action, proposing that by 2027, the supply of China's key core AI technologies should be secure and reliable.

According to the Frost & Sullivan report, revenue in China’s GPGPU market reached RMB 154.6 billion in 2024, with a CAGR of 70.1% from 2022 to 2024. The market is expected to reach RMB 715.3 billion by 2029, with a CAGR of 29.5% from 2025 to 2029.

Moreover, the revenue growth of domestic GPGPU companies is significantly faster than that of foreign competitors. The localization rate of China’s GPGPU market continues to rise, increasing from 2% in 2022 to 3.6% in 2024, and is expected to reach 31% by 2029.

NewTimeSpace understands that municipalities such as Beijing and Shanghai have established local funds to prioritize support for domestic computing power and domestic GPU application demonstrations.

Driven by these multiple favorable factors, as of the close on March 10, 2026, Iluvatar CoreX was quoted at HK$338.20, with a single-day gain of 30.18%. Since its listing, it has risen by 133.89% from its offer price of HK$144.6, bringing its market capitalization to HK$86.01 billion.

NewTimeSpace Disclaimer: All content herein is the original work of NewTimeSpace. Any reproduction, reprinting, or use of this content in any other manner must clearly indicate the source as "NewTimeSpace". NewTimeSpace and its authorized third-party information providers strive to ensure the accuracy and reliability of the data, but do not guarantee the absolute correctness thereof. This content is for reference only and does not constitute any investment advice. All transaction risks shall be borne by the user.