Huawei pulls Ascend 960DT AI chip launch forward to Q1 2027 in push against Nvidia

Huawei speeds up the Ascend timeline

Huawei says its next-generation Ascend 960DT AI accelerator will be ready in the first quarter of 2027, two quarters earlier than the company had signalled. The revised schedule was announced at Huawei's Connect conference by David Wang, the firm's rotating and acting chairman, and confirmed to TechCrunch by a spokesperson, who said the Ascend 960 family is "doubling performance and advancing year by year." Huawei had previously pointed to a third-quarter 2027 launch. The change is part of the Chinese company's effort to compete with Nvidia in the market for AI training and inference hardware.

Building one machine out of many chips

Alongside the chip news, Huawei detailed the systems architecture it intends to build around its silicon. Eric Xu, Huawei's rotating chairman, said the company is using what it calls the Peerium Computing Architecture to construct larger AI computing systems for both training and inference. The idea is to stitch together huge fleets of accelerators — hundreds of thousands now, millions eventually — so that they behave as a single computer. A central component is UnifiedBus, Huawei's technology for connecting processors with memory, storage and networking hardware. The first products built on the architecture are the Atlas 950 SuperPoD and SuperCluster; Huawei says a SuperCluster can link as many as 256,000 accelerator cards.

The chip is early, the system is smaller

The faster chip schedule arrives with an inconsistency in Huawei's system-level claims. Rui Ma, a China tech analyst, pointed out on X that Huawei had earlier said its Atlas 960 SuperPoD would scale to 15,488 Ascend 960 chips, while this week's presentation described a system with 4,096 chips. In her assessment, the chip itself is landing far sooner, but the SuperPoD now on the table is "much smaller than what they originally laid out." Whether that reflects engineering constraints, a change of plans or a staged rollout is not clear from the announcement.

Sanctions and a summit in the background

The timing is hard to separate from politics. According to TechCrunch, the announcement came a week before US President Trump and Chinese President Xi Jinping were scheduled to meet in Washington on September 24. Huawei has pressed ahead with its chip ambitions despite US restrictions limiting China's access to advanced semiconductor technology, and Ma argued that such curbs are unlikely to halt the country's progress, because the stakes of achieving semiconductor self-sufficiency are now too high for China to back off.

The two governments also hold mirror-image positions on AI development. Trump has rejected calls from AI industry figures to slow the technology on safety grounds, framing the US lead over China as the priority. Xu, as reported by the Financial Times, has argued the opposite direction: China must accelerate AI development to catch up with the US, and Chinese companies need to advance further before they can fully understand and address the risks posed by more capable AI systems.

Why it matters

The AI hardware market has largely been treated as Nvidia's to lose, and Huawei's accelerated schedule is a concrete data point that a sanctioned competitor is still iterating on a yearly cadence rather than falling behind. For Chinese cloud providers and AI labs that cannot easily buy top-end Nvidia hardware, an earlier Ascend 960DT shortens the wait for a credible domestic alternative.

But the detail Rui Ma flagged matters just as much. The chip is arriving early while the largest system described around it appears smaller than earlier claims. Frontier AI performance increasingly comes from clusters rather than individual chips, so Huawei's ability to deliver the full Peerium vision — the interconnect, the SuperPoDs and the 256,000-card clusters — not just the silicon, will determine how real the challenge to Nvidia turns out to be.