Shanghai-based Iluvatar CoreX has carried out demonstrations that it says show its Tiangai 300 GPU matching or surpassing Nvidia’s H200 chip in real-world AI inference workloads, as Chinese companies seek to build up the country’s AI infrastructure in spite of limited access to high-end manufacturing technology.

In the live demonstration for mainland Chinese media outlets on 29 October, the company showed the Tiangai 300 outperforming the H200 in complex multimodal inference tasks, such as text-to-image and text-to-video generation using frontier models such as Z.ai’s GLM-5.3, local media outlet Xinzhixun reported.

In one text-to-image test, the Tiangai 300 reportedly took 0.64 seconds longer to complete the task, but achieved a 100 percent completion rate with a throughput of 0.101 images per second, while the H200 achieved a 87.5 percent success rate at 0.086 images per second.

Performance gap

Earlier tests by state-backed Shanghai INESA in July found the Tiangai 300 reaching 1.05 times the H200’s throughput under “typical deployment conditions”, making it the first Chinese general-purpose GPU to match the H200 in empirical testing.

A separate report by Wallstreetcn on 30 October said the Tiangai 300 beat the H200 in mixture-of-experts throughput by 15 percent, while reducing first-token latency by up to 30 percent in long-text processing.

The Iluvatar chip achieved 1.33 percent the effective throughput of Nvidia’s earlier H100 chip in a training task involving Alibaba Group’s Qwen3-32B model, the report said.

Trade tensions

The H200, launched in late 2023, is the most advanced Nvidia chip approved for export to China, and has become a flashpoint for political tensions this year, with some in the US wary of allowing China access to the relatively advanced technology, while China has limited H200 imports and pushed growth for domestic rivals.

Earlier in October, Iluvatar showed its first supernode system, the Tiangai 300 IXL144, with 144 GPUs in a single cabinet, breaking an industry record, at the 2026 China International Industry Fair (CIIF) in Shanghai, saying the system was designed for large-scale model training and inference.