While the US and China keep playing tug-of-war over who gets to buy which chips, Huawei showed up to WAIC 2026 with a rebuttal roughly the size of a shipping container. Its message: sanctions or not, we built a bigger machine.

1,024 Chips, One Very Confident Spec Sheet

Huawei unveiled its Atlas 950 SuperPoD at the World AI Conference in Shanghai on July 16, showing off a prototype that links 1,024 Ascend NPUs in its base configuration and scales up to 8,192 chips. The system runs on Huawei's proprietary Lingqu interconnect with a claimed 3-microsecond round-trip latency, delivers 1 EFLOPS of FP8 compute and 2 EFLOPS of FP4, and packs 256TB of unified memory — enough, Huawei says, to train trillion-parameter models without breaking a sweat.

The headline number is the comparison Huawei chose to make out loud: 6.7 times the compute and 15 times the memory of Nvidia's next-gen NVL144 supernode. The inference-focused Ascend 950PR chip has already been in mass production since April; the training-focused Ascend 950DT is slated for Q4 2026.

Domestic Silicon Just Got a Marketing Department

This isn't just a spec flex — it's a statement about self-sufficiency. Huawei has spent years building an Ascend ecosystem specifically because US export controls make Nvidia's top-tier chips hard to buy in China, and the Atlas 950 is the clearest evidence yet that "we'll just build our own" wasn't an empty threat.

Whether the 6.7x claim survives contact with independent benchmarks is the part nobody can verify yet — it's still a prototype, and vendor comparisons against a competitor's not-yet-shipped hardware should be read with a raised eyebrow. But the direction of travel is unmistakable: China's AI compute stack is racing to become fully vertically integrated, from chips to interconnects to models.

Nvidia's moat has always been software and ecosystem, not raw FLOPS. Huawei just made a very loud case that the FLOPS gap, at least, is closing.

Source: Huawei Central