A Shanghai chipmaker says it has built a GPU that matches Nvidia's H200. According to a South China Morning Post report published October 30, 2026, Iluvatar CoreX demonstrated its new Tiangai 300 GPU to mainland Chinese media and said it can match and in places surpass the H200 on real workloads. The claim is notable because the H200 is, per the report, the most advanced Nvidia chip that can be exported to China under current US rules.
This post sorts the numbers by who said them. The short version: the demo figures are the company's, one outside benchmark is reported but not published in a form we could read, and the usual caveats about software, memory and production scale all apply. SCMP's piece was behind a block when we tried to fetch it, so we rely on secondary summaries and Iluvatar's own earlier disclosures; where that matters we say so.
TL;DR: claimed vs verified
| Claim | Source | Status |
|---|---|---|
| Tiangai 300 matches or beats H200 on real workloads | Iluvatar, via SCMP | Company claim |
| Text-to-image test: 100% completion, 0.101 images/s vs H200 at 87.5%, 0.086 images/s | Live demo, as summarized by tech360 from SCMP | Company demo, one test |
| 1.05x H200 throughput under typical deployment | Shanghai INESA tests, reported by tech360 | Reported third-party number, report not seen |
| 15% higher mixture-of-experts throughput, up to 30% lower first-token latency | A separate financial report cited by tech360 | Unattributed, unverified |
| 1.33x H100 effective throughput fine-tuning an open model | Same report | Unverified |
| Parity across the whole AI stack | Nobody claims this | Not shown |
What was shown
Per the reports, Iluvatar ran a live presentation for mainland media where the Tiangai 300 handled a multimodal text-to-image task. The company's numbers: a 100% completion rate and 0.101 images per second, versus 87.5% and 0.086 images per second for the H200. That is roughly a 17% throughput edge in one configuration.
Two details matter. A 100% versus 87.5% completion rate on a small test is a coarse measure, and a rate like 87.5% (7 of 8) suggests a very small sample, though we have not seen the test design. And who configured the H200 matters: vendor demos routinely use settings tuned for the favored chip.
The third-party number
The reported INESA figure, 1.05 times the H200's throughput under typical deployment conditions, is the strongest independent data point we found. INESA is described as state-backed and Shanghai-based, which makes it local and not neutral in the way an international benchmark body would be. A 5% edge is also inside the range that configuration choices can swing. We could not find the report itself, so treat 1.05x as reported.
What Iluvatar has said before
The company is not new to bold claims. In January 2026, KrASIA reported a four-architecture roadmap (Tianshu, Tianxuan, Tianji, Tianquan) with the last aimed at outperforming Nvidia's Rubin in 2027. It said Tianshu delivered about 20% higher average performance than Hopper on workloads related to DeepSeek V3, with over 90% compute utilization, citing 36Kr. It also reported more than 52,000 GPGPUs delivered to over 290 customers by June 2025, and a Hong Kong IPO on January 8.
Revenue per its prospectus rose from RMB 189 million in 2022 to RMB 540 million in 2024, with widening net losses. That is the context for any performance claim: a company that must show momentum to investors and customers, selling into a market where Nvidia's best chips are restricted.
Why the H200 is the yardstick
China's buyers cannot freely buy Nvidia's newest hardware. Our coverage of the export-control fight, including the remote-access rules debate and ByteDance and Alibaba's Nvidia approval, shows a market where approved Nvidia parts, domestic chips and workarounds compete. A domestic chip that equals the H200 on common workloads would be a credible substitute inside China even if it trails Nvidia's latest abroad.
What a fair test would need
To believe parity, a reader would want:
- The model, batch size, precision and software versions for both chips.
- Tests on standard workloads: LLM serving, mixture-of-experts, fine-tuning, long-context, not only one image task.
- The same measurement on both chips, repeated, with variance.
- Multi-card scaling and interconnect results, since AI clusters are limited by networking as much as by a single chip.
- Memory capacity and bandwidth, which usually decide what model fits.
- Software: how much of CUDA-style tooling and popular frameworks works without porting.
- Price, power draw and delivery dates.
Why matching one chip is only part of the story
Training and serving modern models depends on more than one accelerator's arithmetic. A cluster is limited by how fast chips talk to each other, how much high-bandwidth memory each holds, and how well the software stack schedules work. Nvidia's lead has long rested as much on its CUDA software ecosystem and networking as on raw silicon, and that is the part a single demo cannot show.
A few specific gaps are worth naming. Mixture-of-experts models, which the unattributed report says the Tiangai 300 handles 15% faster, put heavy pressure on memory bandwidth and interconnect, so a single-card advantage can shrink at scale. Long-context serving, where the same report cites up to 30% lower first-token latency, depends on memory capacity and attention kernels that vendors tune carefully. Fine-tuning claims, such as 1.33 times the H100's effective throughput, say little about stability over days-long runs. None of these is impossible. They are simply the claims a buyer should ask to see reproduced on the buyer's own model.
There is also a generational point. The H200 is a Hopper-era part. Nvidia's newer Blackwell and Rubin platforms are a different class, and Iluvatar's own roadmap targets Blackwell-class competition with its next architectures and Rubin only in 2027. Matching the H200 is therefore a meaningful step for the domestic market and a modest one against the global frontier. Both statements can be true at once.
A short timeline of Iluvatar's performance claims
- June 2025: the company reports more than 52,000 GPGPUs delivered to over 290 customers, per its disclosures cited by KrASIA.
- January 8, 2026: Hong Kong IPO.
- January 2026: public four-architecture roadmap; Tianshu reported at about 20% above Hopper on DeepSeek V3-related workloads, with over 90% utilization, per 36Kr as cited by KrASIA.
- October 30, 2026: SCMP reports that the Tiangai 300 matches the H200, based on a live text-to-image demo for mainland media.
The pattern is a company publishing numbers on a schedule that invites scrutiny. That is a good sign for accountability: roadmaps with dates can be checked. What is missing so far is the independent confirmation step that would turn each claim into a fact.
What to watch next
- Publication of the INESA report or any other third-party test with full configuration.
- Participation in standard benchmark programs, which would allow like-for-like comparison.
- Customer deployments at scale, beyond the reported 10,000-card-scale installations of the Tiangai series, which are themselves a company statement.
- Whether the software stack supports the frameworks and models Chinese developers actually use, including recent open-weight releases such as those from DeepSeek and Qwen.
- Supply: who fabricates the chip, which memory it uses, and how many units ship per quarter.
- A response from Nvidia or independent analysts testing the demo workload.
What it means for people who build with AI
If you run workloads outside China, this changes little today. If you track the AI supply chain, it adds to a pattern we have covered: memory limits, packaging limits and demand shaping who can build what, as in this look at TSMC, Samsung and GlobalFoundries and the CXMT revenue report. Domestic accelerators are only useful with domestic memory and packaging, and those are the bottlenecks to watch. For builders inside China, the practical question is porting cost and supply, not peak numbers.
How to check claims like this yourself
Ask where the number first appeared: company slide, media day, or filing. Check whether the comparison chip is named with its configuration. Look for a reproducible benchmark such as MLPerf-style submissions. Compare to the vendor's prior claims and see whether they held up. Finally, wait for customers to report production results.
What we could not confirm
We could not find an Nvidia response, a published INESA report, or an independent lab replication. We also could not read the full SCMP article, so details we cite come from secondary summaries of it. The Tiangai 300 name, test figures and the MoE and latency numbers should be read with that in mind.
Bottom line
Iluvatar's claim is plausible as a company-selected result and worth watching, since a chip that equals the H200 on selected tasks would matter for China's self-sufficiency push. It is not yet verified parity. Wait for independent tests, workloads beyond one demo, and evidence of shipping volume.
Figures reflect public reporting as of October 30, 2026 and may change.