Huawei introduced more than 10 new chipsets for artificial intelligence computing infrastructure on Thursday. The Chinese company used its Huawei Connect 2026 event to present the next generation of its Ascend AI accelerators, alongside new central processing units and high-speed connectivity silicon. Coverage of the event put the total at 11 semiconductors.
Rotating chairman David Wang said the Ascend 960DT is scheduled for the first quarter of 2027, followed by the Ascend 960PR in the third quarter of the same year.
From one accelerator to a million-processor system
The roadmap matters less for any single chip than for how Huawei intends to wire the chips together. Huawei says the new parts were developed around its UnifiedBus interconnect, a technology designed to let AI processors exchange data efficiently inside very large systems, and that future superclusters built on the design could connect as many as 1 million AI processors. The company has previously said its 960-generation systems could eventually scale beyond 1 million neural processing units.
That is the comparison that gives the launch its meaning. A supernode combines several AI processors so they can work on the same computing task, and Wang said Huawei has already delivered more than 1,000 supernodes to more than 370 customers. That installed base, not the chip count, is the baseline against which the 2027 systems should be judged. Huawei has also previously outlined an Ascend 970 generation.
Why the packaging matters more than the silicon
Nvidia's GPUs and its CUDA software ecosystem remain deeply entrenched in global AI development, and US export controls have pushed Chinese technology companies toward domestic alternatives. Washington has nevertheless allowed limited, case-by-case shipments of Nvidia's H200 and similar processors to approved Chinese customers, under conditions introduced in January 2026.
Huawei's answer is framed as a full stack: processors, interconnects, servers and software that can scale without relying entirely on US technology. On that reading, the contest is no longer about who builds the fastest accelerator, but about who can assemble the largest and most efficient system around it, where interconnect bandwidth between chips frequently decides how much useful work a cluster delivers.
Cost is the counterweight. Reuters reported in September that Huawei raised the indicative price of its Ascend 950DT accelerator above 250,000 yuan (37,255 dollars), partly because of soaring high-bandwidth memory costs. Interconnect scale does not remove the memory bottleneck that sets that price, and memory supply remains the tightest link in the chain for any accelerator vendor outside the leading foundries.
What the announcement does not show
Huawei did not publish performance benchmarks, unit prices or shipment targets for the 960 chips in the statements cited, and the quarterly dates are targets rather than contractual commitments. The "1 million processors" figure describes a design ambition, not a deployed or contracted system. The figure of more than 370 customers comes without a breakdown by cluster size, revenue contribution or repeat orders.
No independent benchmark compares the 11 chips with Nvidia's current data-centre accelerators, and there is no public indication of which parts would face export controls or stay restricted to the Chinese market. Those gaps matter because the binding constraint on Huawei's data-centre business is not the ambition of its roadmap but manufacturing capacity and access to high-bandwidth memory.
Sources
- Nikkei Asia — Huawei debuts 11 AI-related chips in challenge to Nvidia, Intel, AMD: asia.nikkei.com
- Insights Integration — Huawei New AI Chips: China Sets 2027 Launches as It Takes Aim at Nvidia: insightsintegration.com
