Shanghai's Iluvatar CoreX put a number on the biggest open question in Chinese AI hardware. On 28 October, at the company's 2026 Media Open Day, it showed the Tiangai 300 accelerator running single-card inference throughput at 1.05 times Nvidia's H200, a figure from a third-party test Shanghai INESA completed in July.

That is a real milestone. It is also the least interesting document Iluvatar published this autumn. On 29 September, four weeks before the demo, the company filed a Hong Kong interim report showing revenue up 191.6% and gross profit up 0.1%, and a first-ever profit that came almost entirely from a stock investment.

Read both, in that order, and the buying decision changes.

The claim, precisely

Shanghai INESA, a Shanghai state-owned technology group, tested the Tiangai 300 in July 2026 across four mainstream models and three scenarios under a typical prefill-decode disaggregated deployment. INESA has published no report of its own. The number reaches the public through Iluvatar's disclosure.

IT之家 flagged the gap in plain terms: the test method, model load and workload conditions have not been published alongside it. One outlet, 新浪科技, chose on 4 November to call the comparison target "某国际顶尖GPU", a top international GPU, rather than naming the H200 as every other report does.

Iluvatar's own framing from 新华财经 is the most honest line in the pack: the comparison is not against theoretical peak, but against the actual output capability of the whole inference chain under a specific model and deployment condition. Read that as scope, not a defect. It is also the whole claim.

The lab numbers published by 天数智芯性能实验室 cover more ground:

Metric versus H200 Result Condition
Attention performance 1.05× 64K context
MoE throughput 1.15× GLM-5.3, 64K context, same user experience
Agent time-to-first-token 10% to 30% lower 8K to 128K context
Multimodal generation throughput 2% to 5% higher
Recommendation inference throughput +21% search, ads, recsys online workloads
Light-compute latency 26% lower
Full-load step-time jitter 0.43% vs 0.76% on Hopper

The MoE and recommendation figures are much bigger than 5%, which suggests the 1.05× end-to-end headline is a blended average across a workload set that includes plenty of easy cases. The win concentrates in particular shapes of work.

The training number, which gets less attention, has a mismatched baseline: on an eight-card FP8 Qwen3-32B test, Iluvatar measured per-card effective throughput at 1.33 times the H100. The H100, not the H200.

The interim report is the harder story

For the six months to 30 June 2026, Iluvatar's revenue was RMB945.7 million, up 191.6% year on year. Gross profit was RMB162.6 million, up 0.1%. Group gross margin went from 50.1% to 17.2%. Research and development costs were RMB558.9 million, up 23.8%, which is three and a half times the entire company's gross profit.

The product mix explains the margin collapse.

Line H1 2026 revenue Share Year on year Gross margin
Training series (Tiangai) RMB261.9m 27.7% +38.0% 64.5%
Inference series (Zhikai) RMB654.2m 69.2% +651.8% 35.6%
GPGPU products RMB916.1m 96.9% +231.0% 43.9%
AI computing solutions RMB13.6m 1.4% −68.1% 27.7%
Semiconductor components RMB14.8m 1.6% new line −1,157.3%
Impairment −RMB72.2m
Group RMB945.7m 100% +191.6% 17.2%

Inference became the business. The Zhikai line grew 651.8% and now carries 69.2% of revenue at a 35.6% gross margin, against 64.5% for the flagship Tiangai training parts. That mix shift alone explains most of the margin fall.

Then there is the semiconductor components line: RMB14.8 million of revenue producing a RMB171.0 million gross loss, a margin of minus 1,157.3%. The company says it offered server components to certain customers below its own procurement costs, to accommodate their needs in the hope of longer-term business. Alongside that went a RMB72.2 million impairment, up from RMB0.9 million, mainly a write-down of work-in-progress inventories following "our new market and product strategy". Add the two and the RMB401.8 million of gross profit on general-purpose GPUs becomes RMB162.6 million of group gross profit.

And the profit. Iluvatar reported RMB106.3 million for the period against a RMB609.3 million loss a year earlier, with adjusted net profit of RMB246.0 million. It had lost RMB1.00 billion on RMB1.03 billion of revenue across 2025. The interim report is explicit about the cause of the turnaround: it was "primarily due to a significant increase in unrealized net gain arising from the Group's investment in the shares of a company listed on the Shanghai Stock Exchange". Other income and gains jumped from RMB39.5 million to RMB873.3 million, of which about RMB760.2 million was fair-value gains. The stake is in the notes: RMB99,999,984 invested during the half for roughly 0.27% equity interest, carried at RMB855,898,000 on 30 June against nil at the end of 2025. Income tax expense was nil in both periods, "primarily due to our loss positions".

The balance sheet tells the other half of the story. Inventory went from RMB709.8 million to RMB1,968.1 million across the half, trade receivables from RMB576.6 million to RMB1,237.7 million, and prepayments and other assets from RMB630.1 million to RMB1,942.3 million. Cash rose to RMB2,910.8 million. Headcount went from 685 a year earlier to 916, and the customer count from over 340 at the end of 2025 to over 380 at 30 June.

On 2 November, five days after the demo, Iluvatar filed a voluntary announcement: chip shipments for the ten months to 31 October exceeded 80,000 units. The company notes the figure is unaudited and "do(es) not constitute, represent or indicate the revenue or financial performance of the Group".

The hardware

What Iluvatar has published about the Tiangai 300, which it launched in July as the Tiangai Gen 4:

  • 7nm process, in-house SIMT general-purpose compute architecture rather than a Transformer-only design
  • More than 100 billion transistors across more than 200 AI compute cores
  • FP16 above 1,000 TFLOPS, INT8 above 2,000 TOPS, with FP4 and FP8 supported but no compute figures published
  • 144GB of memory at roughly 4TB/s bandwidth
  • Under 600W typical power
  • PCIe card and OAM module form factors
  • Native support for CUDA semantics, with around 100 accelerator libraries built since 2018

Die size, FP8 numbers and the interconnect protocol are undisclosed. Memory type is contested: Morgan Stanley's supply-chain research records HBM3E, Baidu Baike records HBM3, and the company has never said.

The system story is the IXL144 supernode: 144 GPUs and 18 CPUs in one standard rack, roughly 10.1TB of aggregate memory, with a peer-to-peer interconnect engine taking single-transfer latency from 3 to 5 microseconds down to 1. Iluvatar estimates about 40% lower cost per million effective tokens than international mainstream solutions. That is a company estimate, not a customer result.

The scale question

Iluvatar listed in Hong Kong on 8 January 2026 under stock code 9903 at a market capitalisation near HK$35.4 billion, after raising more than HK$3.5 billion in the offering. Frost & Sullivan put it at 0.3% of China's general-purpose GPU market by revenue in 2024, fifth place, in a market forecast to reach RMB715.3 billion by 2029.

That 0.3% sits next to the benchmark. The chip may be excellent. The company that supplies it at volume was a rounding error in 2024, and has since doubled its inventory and receivables in six months.

The comparison target is not as available as it sounds

The H200 is the most advanced Nvidia AI processor approved for export to China under current US rules, and even that approval is partial. Washington cleared H200 shipments on 13 January 2026 with a fee of roughly 25% on the sales, and by late February Nvidia had generated no revenue under the licence programme. Beijing then restricted Chinese firms from buying the card, before starting to allow selected companies, including Alibaba, limited purchases. The eight-card H200 module was going out through channels at around RMB1.4 million.

Iluvatar has never published a price for a Tiangai 300. So the benchmark is a card that is hard to buy, compared against a card with no price.

The verdict

The 1.05× is worth a pilot and nothing more. It moves the Tiangai 300 from impossible to testable for the narrow group of buyers priced out of H200 capacity, and the company's own lab numbers are specific enough to be worth repeating on your own traffic.

The filings are why you should not go further than a pilot. Iluvatar's revenue tripled in a half year while gross profit did not move at all, and the gap was filled with a RMB171 million loss on components sold below cost, a RMB72 million inventory write-down and a mix sliding toward its lower-margin inference line. The first profit in the company's history is a paper gain on a Shanghai-listed stock, which is a tradeable thing to have but not an engineering achievement.

Ask for six things in writing before you sign: the four model names, the input and output token lengths, the batch size and concurrency, the precision used on each side, the exact H200 configuration and power limits, and the software versions. Then run it in your own harness.

Two numbers to watch in the next filing. Whether inference gross margin stops falling from 35.6%. And what happens to the RMB856 million stake if the market turns, because that is what the "profit" was made of.

Sources