Iluvatar CoreX said its Tiangai 300 GPU matches and even surpasses NVIDIA’s H200 in real-world AI workloads.
Iluvatar CoreX (Shanghai Tianshu Zhixin) is a Shanghai-based maker of general-purpose GPUs for AI training and inference. Founded in late 2015 by Li Yunpeng, it listed on the Hong Kong exchange in January 2026 (9903.HK). Earlier products include the TianGai-100 training series. The Tiangai 300, unveiled in July, is its current flagship. The company has government-linked partnerships and says its Tiangai series has reached 10,000-card-scale deployments.
The US government relaxed licensing rules earlier this year to allow H200 sales to China. NVIDIA, which launched the H200 in late 2023, said in August that it had begun shipping its first data-centre H200s there. Beijing has meanwhile discouraged big Chinese tech companies from buying foreign chips. This has accelerated the domestic alternatives covered in our earlier pieces on China’s semiconductor industry and Huawei’s Ascend.
Shanghai INESA, a state-backed body, ran benchmarks that found the Tiangai 300 reached 1.05 times the H200’s throughput under typical deployment conditions. These were prefill-decode disaggregated setups. A report on those tests called it the first general-purpose GPU made in China to reach H200 level in empirical testing.
What the Tiangai 300 Benchmark Results Show
Xinzhixun reported that Iluvatar gave a live demo to mainland Chinese media on October 28. In it, the company showed the Tiangai 300 beating the H200 on complex multimodal tasks that mix text, images and video. The models included Z.ai’s GLM-5.3. That is one of the Chinese open models mapped in our 2026 Chinese LLMs overview.
Its figures put the Tiangai 300 at 100% completion and 0.101 images per second in a text-to-image test. The H200 managed 87.5% completion and 0.086 images per second. The Tiangai 300 took 0.64 seconds longer to finish the task.
Results reported on October 29 show the Tiangai 300 ahead of the H200 by 15% in mixture-of-experts throughput. In long-text processing it cut first-token latency by up to 30%.
The Tiangai 300 also fine-tuned Alibaba’s Qwen3-32B model at 1.33 times the effective throughput of the H100. The H100 is NVIDIA’s previous-generation chip. That was a training test, not a comparison with the H200.
Iluvatar’s Tongyang TY1200 edge processor beat NVIDIA’s Thor T5000 by 20% in specific large language model tests only.
Which Chinese GPU Makers Are Challenging NVIDIA
Biren, MetaX, Enflame and Moore Threads are each building home-grown GPU substitutes alongside Iluvatar. The highest-profile alternative remains Huawei’s Ascend line. All of them work under two pressures. Beijing wants China to make its own chips, and US trade curbs are tightening.
How Far Has Iluvatar CoreX Come
In January Iluvatar unveiled a multi-year architecture roadmap that aims to overtake NVIDIA’s Rubin within two years. That is a target, not a result. The company has not said what the Tiangai 300 costs or how many units it can supply.
The accounts of Iluvatar’s demo do not say whether the benchmarks shown were run independently or with what settings. The 1.05 figure applies only under typical deployment conditions, and the 30% latency cut is a ceiling.
Author: Grace Sharp
See Also:
What Is the Apex Logistics Probe, and Why Could It Become a Test Case for US Chip Export Controls?
What Is Huawei Ascend? The CUDA-Free AI Stack, Explained
Frequently Asked Questions:
Shanghai INESA measured the Tiangai 300 at 1.05 times the H200’s throughput under typical prefill-decode disaggregated deployment conditions. Iluvatar CoreX also reported a 15% throughput lead in mixture-of-experts workloads.
Shanghai INESA, a state-backed body, ran the third-party benchmarks. Iluvatar CoreX also held a live demo for Chinese media on 28 October showing results on multimodal tasks.
Iluvatar CoreX is a Shanghai-based maker of general-purpose GPUs for AI training and inference. It listed on the Hong Kong exchange in January 2026 under the ticker 9903.HK.
Company figures put the Tiangai 300 ahead of the H200 by 15% in mixture-of-experts throughput on the GLM-5.3 model at 64K context.
