Iluvatar CoreX, a Shanghai-based maker of general-purpose GPUs that is listed in Hong Kong, said on October 28 that its Tiangai 300 chip reaches the measured performance of Nvidia’s H200, citing a third-party test. According to Chinese-language reports of the event, the test put the chip’s normalized single-card throughput at 1.05 times the H200 in a typical deployment that separates the two main stages of AI inference. The tester, the models and most of the test conditions have not been published, and every figure reported so far comes from the company.

What Iluvatar showed on October 28

The claim was made at the company’s 2026 media open day in Shanghai, which IT Home and the finance outlet 21st Century Business Herald describe as the first public live demonstration of the Tiangai 300 running on hardware. Until then, public information about the chip had covered mainly its architecture and roadmap, according to IT Home. The company called the Tiangai 300 the first domestic general-purpose GPU whose measured performance reaches H200 level.

InfoQ China, which published a detailed account on October 29, says the 1.05 times figure comes from a third-party test under a typical deployment that splits the prefill and decode stages of inference. The test covered four mainstream models and three application scenarios, InfoQ reports, but the article does not name the tester. The South China Morning Post also covered the claim; Lens could not open that article, and the figures here come from the Chinese-language reports of the Shanghai event.

The numbers, and the conditions attached

Beyond the third-party result, Iluvatar gave results from its own testing, as reported by InfoQ and by 21st Century Business Herald. In attention computation at a 64,000-token context, the chip reached 1.05 times the H200. For mixture-of-experts throughput on the GLM-5.3 model at the same context length and equal user-experience requirements, it reached 1.15 times. First-token latency for agent workloads at contexts from 8,000 to 128,000 tokens was about 10% to 30% lower. For multimodal generation, throughput was about 2% to 5% higher, and for online recommendation inference it was 21% higher.

The company also reported that eight-card training of Alibaba’s Qwen3-32B model in FP8 precision reached 1.33 times the per-card effective throughput of Nvidia’s older H100. Iluvatar said its IXL144 system fits 144 GPUs and about 10.1 terabytes of pooled memory into one standard cabinet, and that this design lowers the cost per million effective tokens by about 40% compared with mainstream international alternatives. InfoQ notes that the comparison baseline and cost accounting behind that 40% figure are not fully disclosed.

The edge-computing line was part of the same event. Iluvatar said its TY1200 module, rated at 300 trillion operations per second of dense compute, reached 1.2 times Nvidia’s Thor T5000 on two large language models, according to InfoQ.

What is not public

IT Home pointed out that the specific test methods, model loads and workload conditions had not yet been released. InfoQ added its own caution: the 1.05 times figure applies to the stated prefill-and-decode deployment, depends on the model, concurrency and the balance of compute and communication, and does not represent every workload. The article says the stability result, a step-time fluctuation of 0.43% against 0.76% for the Hopper-based comparison, depends on test duration, load and hardware setup.

Lens analysis: the H200 comparisons in the company’s own list are leads of roughly 2% to 30% on selected measures, not across-the-board gains, and the multimodal result is the smallest. A 5% lead in one deployment style is narrower than a claim of parity on all tasks. Until the tester and the test conditions are named, outsiders cannot check whether the baseline H200 systems were configured the same way. The 21st Century Business Herald account reports no results for text-to-image or text-to-video generation.

A chip first shown in July

Iluvatar launched the Tiangai 300 on July 19 at the 2026 World Artificial Intelligence Conference in Shanghai, according to IT Home. At that time the company compared it with Hopper-based systems without naming the H200. It said the chip, built on what it calls a SIMT general-purpose computing architecture, delivered attention efficiency above 90%, about 10% higher efficiency than the Hopper solution for 64,000-token attention and for mixture-of-experts workloads, about 20% lower time to first token and about 13% lower average communication latency. IT Home noted that the memory capacity, process node and absolute compute figures were not stated.

The company has set out a public timetable. Kr-Asia reported that Iluvatar published a roadmap of four GPU architectures, each tied to a Nvidia product, with the last, called Tianquan, aimed at Nvidia’s Rubin platform in 2027. The same article says Iluvatar completed its Hong Kong initial public offering on January 8, that it had delivered more than 52,000 general-purpose GPUs to more than 290 customers by June 2025, and that its revenue rose from 189 million yuan in 2022 to 540 million yuan in 2024 while net losses widened. Kr-Asia notes that the performance and ranking claims in the roadmap are the company’s own.

Wikipedia’s entry on the company says it was founded in December 2015 by Li Yunpeng, who previously spent about 10 years at Oracle, and that its first training chip, the 7-nanometer Tiangai 100, was released in January 2021.

The market for domestic GPUs

Iluvatar’s claim arrived in a week when Chinese GPU makers were in the market spotlight. Bloomberg reported on October 28, according to AI Weekly, that Moore Threads, a Shanghai-listed rival, posted third-quarter revenue of about 1.6 billion yuan, up 1,773.92% from a year earlier, and its first quarterly profit of 49.26 million yuan. Briefs Finance, also relaying Bloomberg, says management credited demand for its MTT S5000 chip from AI startups and government bodies, and that the company and JD.com are building a domestic cluster of 100,000 GPUs.

The share prices did not follow the revenue. Manuals.dev, which links to a South China Morning Post article as its source, said Moore Threads and the storage-technology company Biwin each fell more than 10% in morning trading on October 29, reversing the previous day’s gains, while the STAR 50 index dropped more than 4% during the session. The article said it could not establish why investors sold. Briefs Finance had noted that Moore Threads shares, after rising roughly eightfold in their first days of trading, later gave back 59% of those gains.

Demand for local chips is also shaped by what Chinese companies may buy from abroad. Reuters reported in January, citing The Information, that Beijing had told some technology companies to halt orders for Nvidia’s H200 and was expected to require purchases of domestic AI chips, after the Trump administration approved H200 exports to China late last year on condition of a 25% revenue-sharing payment to the U.S. government. Tom’s Hardware has since reported that Trump said China was blocking H200 purchases in favor of domestically developed chips. Lens did not find a current account of how many H200 chips have shipped to Chinese buyers.

What to watch

The first question is whether Iluvatar or the tester publishes the models, software versions and system configurations behind the 1.05 times result, which would let other labs repeat it. The second is whether customers outside the company’s own demonstrations deploy the Tiangai 300 and the IXL144 cabinets at scale. Kr-Asia’s account of the roadmap shows the company has committed to dates for later chips, so its next architecture may face the same kind of scrutiny. Moore Threads’ share-price swings show that strong results do not guarantee that investors in China’s chip stocks will follow.

Sources

Image: Pudong skyline at dusk.jpg by Lloyd Tudor, CC BY-SA 4.0, via Wikimedia Commons. Illustrative photo; it does not show the events described.