Chinese cloud buyers weighing Nvidia H200 imports gain another domestic training option, in a segment that brought Iluvatar CoreX only RMB262 million in the first half.
Iluvatar CoreX, a Shanghai designer of general-purpose graphics processors, says its Tiangai 300 chip has matched Nvidia’s H200 in measured performance, the first Chinese GPU to do so by its account. At a media open day in Shanghai on October 28, the company demonstrated the chip in public for the first time and cited a third-party test that, converted to a single card, put the Tiangai 300’s overall throughput at 1.05 times the H200’s, IT Home reported.
Nvidia’s H200 is a widely deployed data-center chip for training and running AI models, which makes it the reference point Chinese GPU makers now measure themselves against. A credible domestic equivalent would give Chinese cloud providers and model developers an alternative to imported accelerators.
Iluvatar CoreX did not name the testing party or publish the workload conditions behind the 1.05 figure, IT Home noted.
The company paired the chip with a rack-scale system and an inference scheme
Alongside the Tiangai 300, Iluvatar CoreX showcased its IXL144 supernode, a single rack holding 144 GPUs and 18 CPUs with about 10.1 terabytes of memory. The company estimates the system’s cost per million effective tokens at about 40% below mainstream international solutions, another figure based on its own calculation. The IXL144 made its debut at the China International Industry Fair in Shanghai earlier this month.
The company also said it had run China’s first heterogeneous deployment of attention–feed-forward disaggregation, a technique that splits the two main stages of model inference across different processors, cutting inference latency by 60%, according to Guandian. For edge computing, it said its Tongyang TY1200 module outperformed Nvidia’s Thor T5000 in testing and was the first domestic edge product to run GPU-accelerated physics simulation on MuJoCo-Warp, a tool used to train robots in simulation. Futu News, carrying a Gelonghui report, listed the same set of announcements.
Executives framed the strategy around cost and usability rather than peak specifications. At the industry fair, vice president Zou Xuan described a supernode as a systems-engineering effort in which the real contest is economic efficiency. Software vice president Shi Jiasheng said what matters is whether customers can actually use and afford such systems, the 21st Century Business Herald reported, carrying a STAR Market Daily article. The company says the IXL144 supports CUDA and other mainstream software ecosystems.
Training chips remain a small part of the business
The Tiangai line is Iluvatar CoreX’s training-oriented product family, and it has so far been the smaller business. Iluvatar CoreX’s training GPUs brought in RMB262 million in the first half, while its Zhikai inference chips generated RMB654 million. Training revenue grew 38% from a year earlier, while inference revenue rose 651.8%. Total first-half revenue rose 191.6% to about RMB946 million, and net profit of about RMB106 million included an unrealized gain of about RMB756 million on a stake in another chip company, according to a Guosen Securities note published by Gelonghui. The training line carried the higher gross margin, at 64.5% against 35.6% for inference chips.
The company listed on the Hong Kong exchange’s main board in January at HK$144.60 a share, raising about HK$3.7 billion, and is one of four Chinese GPU makers to go public since December 2025, alongside Moore Threads, MetaX and Biren, Yicai reported. In 2024 it held about 0.3% of China’s general-purpose GPU market. A July share placement raised about HK$7.07 billion more. Its shares closed at HK$73.60 on October 28, up 3.08%, about half the listing price.
Volume orders would be the test that matters
Large buyers are already diversifying. Industry sources told Yicai in a June report that ByteDance was discussing a purchase of at least 50,000 Iluvatar chips, mainly Zhikai inference GPUs, which would make it ByteDance’s third GPU supplier after Huawei and Cambricon; neither company commented. The Tiangai 300 was launched at the World Artificial Intelligence Conference in July, when the company said it was ready for large-scale application, but it has given no shipping date, price or named customer.
The next markers are a shipping date, a price and a named customer training models on the Tiangai 300. EastFrontier has tracked how Huawei’s advance is pressing Nvidia’s China lead; a training order for Iluvatar CoreX would add a second domestic supplier at that level.