Chinese AI semiconductor startup Illuvatar CoreX Semiconductor has unveiled a multi-year GPU architecture roadmap and an edge computing product lineup, accelerating its push into the AI market and competition over semiconductor self-reliance.

The Hong Kong-based South China Morning Post reported on Tuesday that the announcement comes as Chinese companies move to reduce reliance on overseas technology such as Nvidia and strengthen their own capabilities.

The company said its first architecture, Tianshu, outperformed Nvidia's Hopper platform last year. It said its follow-on Tianxuan architecture was designed to take on Blackwell. It said Tianji aims to surpass Blackwell in 2026, while Tianquan is planned to exceed Rubin in 2027 and then shift to a breakthrough architecture, the company explained. It aims to secure global GPU competitiveness through successive generational designs.

The company plans to launch multiple product lines based on four generations of architecture under names drawn from the traditional Chinese term for the Big Dipper. It highlighted efficiency improvements in Tianshu, including reduced duplication in memory access, higher effective bandwidth, lower power consumption and eased resource contention through dynamic task allocation. In tests using the DeepSeek V3 model, Tianshu posted an average performance advantage of about 20 percent over Hopper.

It also unveiled 4 edge-computing products in its Tongyang, or TY, series. The series offers performance in the range of 100 to 300 tera operations per second, and the company said TY1000 outperformed Nvidia's AGX Orin in a range of test scenarios. Its strategy is to secure competitiveness not only in AI training but also in real-time processing and edge computing.

Founded in 2015, Illuvatar CoreX released what it introduced as China's first general-purpose AI training GPU, its first-generation TG. It later unveiled second generation and TG third generation products, which are scheduled for mass production in 2026. Its inference-only ZK series was first released in 2022, expanding product lines for both training and inference at the same time.

The company completed a listing on the Hong Kong stock market on Jan. 8. Before the listing, Centurium Capital was a major external shareholder with about a 23 percent stake, and after the listing the shares were 182 Hong Kong dollars at Tuesday's close, with a market capitalisation of 46.3 billion Hong Kong dollars, or about $6.0 billion. It is seen as having relatively greater room for growth compared with major AI semiconductor companies in China.

According to its prospectus, revenue in 2022 to 2024 was 189 million yuan, 289 million yuan and 540 million yuan, respectively, posting a compound annual growth rate of 68.8 percent, but net loss widened from 554 million yuan to 892 million yuan. Revenue in the first half of 2025 was 324 million yuan and net loss was 609 million yuan. The company said the losses were due to increased research and development investment and costs to prepare new product launches.

The company also said it had shipped a cumulative total of more than 52,000 general-purpose GPUs as of June 2025. It said this shows it is gradually expanding market share through the R&D investment and expansion of its new product lines mentioned earlier.

Separately, Nvidia's effort to resume sales of its China H200 chips remains uncertain amid the need for approval from U.S. authorities and an ambiguous response from Chinese authorities, while Nvidia Chief Executive Jensen Huang is taking part in customer meetings and events during a visit to China. This shows competition in the global GPU market and China's push for AI semiconductor self-reliance unfolding at the same time.