Iluvatar Says GPU Matches Nvidia H200; Racks Cut Costs 40%

Chinese chipmaker Iluvatar CoreX showed its Tiangai 300 GPU running live for the first time at a Shanghai media day on 28 October 2026, and said it is the first domestic general-purpose GPU whose measured performance reaches the level of Nvidia’s H200, ITHome and ifeng Tech reported. The headline figure is a third-party test that put per-card end-to-end throughput at 1.05 times the H200.

The same reports say the company’s IXL144 rack puts 144 of the GPUs in one cabinet with about 10.1TB of memory, and that Iluvatar CoreX estimates its cost per million effective tokens can be about 40% lower than international mainstream solutions. By our arithmetic, 10.1TB across 144 GPUs is about 70GB per GPU, roughly half the 144GB per card the company lists for the Tiangai 300.

Short explainer · Iluvatar Tiangai 300 vs H200

A vertical news Short narrated from this piece’s headline and opening line.

Business Pill · Sovereign AI

An evergreen Business Engineer explainer on sovereign AI. It teaches the general idea only and does not discuss Iluvatar CoreX or the Tiangai 300.

What Iluvatar CoreX Says the Chip Does

ifeng Tech reported that the 1.05x figure comes from a test completed in July by Shanghai Instrument Electric (INESA), in a prefill-decode (PD) disaggregated deployment, which it described as typical for large models. ITHome noted that the specific test methods, model loads and workload conditions had not been published.

The rest of the H200 comparisons come from Iluvatar CoreX’s own performance lab, per ifeng Tech: Attention at 64K context at 1.05 times the H200, MoE throughput on GLM-5.3 at 64K at 1.15 times, agent first-token latency 10% to 30% lower at 8K to 128K context, and multimodal generation throughput about 2% to 5% higher. A training demo, an 8-card FP8 run on Qwen3-32B, was compared with the H100, not the H200, at 1.33 times per card.

The company launched the Tiangai 300 at the World Artificial Intelligence Conference on 19 July. TrendForce reported then that it supports FP4 and FP8 instructions and that the company said it meets the conditions for scaled commercial deployment. ifeng Tech lists 144GB of memory per card and FP4, FP8, FP16 and BF16 support. Neither report gives a price or a shipping date.

Our Arithmetic: The Memory Numbers Do Not Line Up

Nvidia’s product page says the H200 is the first GPU to offer “141 gigabytes (GB) of HBM3e memory at 4.8 terabytes per second (TB/s)”. On paper the Tiangai 300’s 144GB is about 2% more, by our arithmetic. The reports we read give no memory bandwidth for the Tiangai 300.

The rack figure is the one that does not reconcile. 144 GPUs at 144GB each would be about 20.7TB, by our arithmetic, against the roughly 10.1TB that ITHome and ifeng Tech report for the IXL144 cabinet. That is about 70GB per GPU, about 49% of the card spec. The reports we read do not explain the difference, and we did not contact the company.

How the Lever Works

Iluvatar CoreX chose the H200 as its yardstick. On Nvidia’s own filings, China’s share of its revenue fell from 19.2% in fiscal 2025 to 9.1% in fiscal 2026 under its headquarters-based count, as set out in our analysis of how much of Nvidia’s revenue comes from China, and the recount itself is in our piece on the Singapore line.

The 40% cost estimate is for the IXL144 system, against unnamed international mainstream solutions, by the company’s own calculation. AASTOCKS’ quote page for the Hong Kong-listed company showed the shares down 8.76% on 29 October, with short-selling data as at 16:25.

The Structural Read

As we read the reports, the claim of H200 parity rests on one third-party test of end-to-end throughput, while the larger multiples, such as 1.15x on MoE throughput, come from the company’s own lab.

As we read them, the 40% cost estimate is a rack-level figure, computed by the company against systems it does not name.

By our arithmetic, the Tiangai 300’s listed 144GB per card is about 2% more than the H200’s 141GB, but the rack-level memory figure the company gives works out to about 70GB per GPU.

Three Implications

FOR BUYERS IN CHINA The company says the Tiangai 300 meets the conditions for scaled commercial deployment, per TrendForce, but the reports we read give no price or shipping date.

FOR COMPARISONS WITH NVIDIA ITHome notes that the test methods, model loads and workload conditions behind the 1.05x figure were not published, so the result cannot yet be reproduced from the reports.

FOR INVESTORS AASTOCKS’ quote page showed the Hong Kong-listed shares down 8.76% on 29 October, with short-selling data as at 16:25.

The Business Engineer Lens

This story maps onto the Business Engineer framework The AI Memory Chokepoint.

The framework puts it this way: “Memory is becoming the new compute.”

As we read it, that is why the memory figures in Iluvatar CoreX’s claims matter as much as the throughput ones. The framework does not discuss Iluvatar CoreX or the Tiangai 300.

Related coverage: How Much of Nvidia’s Revenue Really Comes From China and Nvidia Recount: Singapore Gone, US +$16.2B, China +$7.9B.

What Is Not Established

Only the 1.05x end-to-end figure is attributed to a third party, and ITHome notes that its test methods, model loads and workload conditions were not published. The other comparisons are the company’s own. The 40% cost estimate names no comparison system. ifeng Tech’s report is a self-media post on its platform, carrying the platform’s disclaimer. The 70GB, 20.7TB and 2% figures are our arithmetic from the reported numbers.

The Bottom Line

Iluvatar CoreX says its Tiangai 300 matches Nvidia’s H200, on the strength of one third-party test at 1.05x and several in-house ones. Its rack claims a 40% cost advantage, but the rack’s reported 10.1TB works out to about 70GB per GPU, against a 144GB card spec, by our arithmetic.