Chinese AI chip vendors spent the past two years answering one question: can you run the workload at all. The question has changed to whether it is worth running it.

At a media day in Shanghai this week, Iluvatar CoreX (TianShu ZhiXin) put its flagship TianGai 300 in front of reporters and let third-party testers run it against Nvidia's H200.

What the numbers say

In a typical prefill-decode disaggregated deployment, the card delivered 1.05x the end-to-end throughput of H200 on a per-card basis, according to tests completed by a third-party lab in July. On 64K long-context attention it was also 1.05x. Multimodal generation inference gained 2 to 5 percent, and recommendation-model throughput rose about 21 percent in online search, ads and recommendation workloads.

Those multipliers come with conditions attached: specific models, specific deployment setups. That is worth stating plainly, because the comparison is a full inference pipeline rather than a peak-FLOPs spec sheet — more meaningful, but also sensitive to parallelism strategy. Buyers should re-run the benchmark with their own models.

On hardware, the card carries 144GB of HBM3e, above H200's 141GB. For models now shipping with million-token contexts, memory capacity is the hard ceiling. Crossing that line means longer contexts or larger batches fit.

The supernode is where the cost math changes

Single-card performance gets you a meeting. Systems win the deal.

TianGai 300's IXL144 supernode packs 144 GPUs into a standard rack with about 10.1TB of aggregate memory. Iluvatar's own estimate puts cost per million effective tokens roughly 40 percent below mainstream international alternatives.

That gap does not come from cheaper silicon. It comes from interconnect and scheduling: a new P2P engine targets single-hop latency of 1 microsecond, down from 3 to 5, supports unified memory semantics, and enables large-scale expert parallelism that spreads MoE weights across more cores. For mixture-of-experts workloads, those details decide real throughput.

Software is the actual barrier

Hardware parity means little if developers cannot use it. Iluvatar framed its stack around three claims: easy to use, easy to optimize, easy to develop.

Two verifiable examples: environment setup and service deployment for PD-separated serving dropped from one hour to 10 minutes, with manual steps cut from dozens to one. A model tuning task that took seven days now finishes in under two hours. Neither is as photogenic as a benchmark chart, but both matter more to a team on deadline.

Edge is coming down too

The company also showed TongYang TY1200, a compact terminal aimed at embodied AI and on-site deployment, rated at 300 TOPS dense rather than the sparse peak figures the industry often quotes. Against Nvidia's Thor T5000, Iluvatar reports 1.2x on Qwen3.8-27B and Qwen3.6-35B-A3B, and 158 to 180 percent on Cosmos-R2 2B and 8B. A next-generation TY2000 is slated at 600T.

What to do with this

Three practical takeaways for teams evaluating domestic accelerators.

First, benchmark with your own models and concurrency. Vendor baselines are a starting filter, not a decision.

Second, price the migration. Missing operators, framework gaps and tuning labor are real costs that do not appear in the token-per-second table.

Third, start where rollback is cheap. Inference and fine-tuning first; pre-training only after the toolchain has proven itself.