Huawei's 2027 Nvidia Rival Hinges on the Network Between Its Chips

Huawei says it can field a credible Nvidia rival by 2027. The binding constraint is no longer the chip itself but the inter-chip network that lets a million processors act as one computer.

That framing came from rotating chairman Eric Xu on September 17 at HUAWEI CONNECT 2026 in Shanghai, where Huawei unveiled two pieces of new hardware and one architecture meant to make the claim concrete. The Atlas 960E SuperPoD links up to 4,096 of Huawei's NPUs (the company's term for AI accelerators) into a single training cluster through an all-optical interconnect Huawei calls UnifiedBus. The Peerium Computing Architecture, the second reveal, is meant to scale that idea a hundredfold: Huawei wants a million processors acting as one computer by 2027, using a programming model called Nested BSP that breaks the master-slave pattern Nvidia's NVLink and its peer interconnects are built on. All-optical interconnects avoid the signal loss and bandwidth ceiling that copper traces hit at cluster scale, which is why Nvidia's own NVLink switch chips exist and why the industry has moved to treat the inter-chip network as the binding layer, not the silicon.

The hardware claim and the architecture claim are separable bets. The technical detail behind both claims is in a Huawei-authored arXiv preprint on Nested BSP and the Nested Parallel von Neumann Architecture, and the open-source implementation of the underlying Unified Bus protocol lives in a separate clean-room paper, OpenURMA. Treating OpenURMA as independent is honest: the protocol is open, but the implementation that runs on Huawei's own Atlas hardware is the company's, and outside benchmarks are not yet in the source set.

What matters here is whether a self-built Chinese compute stack can hold a credible schedule. The wire will frame it as a US-China scoreboard instead. Independent research group Epoch AI has weighed in with a roadmap analysis titled "Will Huawei catch up to Nvidia by 2030?", the cleanest external counterweight to Huawei's own claims. Epoch's team lands on the same constraint: the inter-chip network, not the transistors. If Epoch's read is that Huawei is behind on process nodes, the 2027 chip slips and the million-processor question becomes academic. If independent benchmarks show the Atlas 960E already matching Nvidia H100 or H200 clusters on real training runs, the 2027 timeline is conservative.

Xu's keynote at HUAWEI CONNECT drew a line on how bullish to read the moment. Chinese AI labs "must accelerate development," he said, and added that domestic models "aren't yet powerful enough to pose frontier risks." Read with the chip schedule, that admission turns 2027 into a precondition for a model ecosystem that does not yet exist. A 2027 chip that trains 10-trillion-parameter models is useless if no Chinese lab is ready to train one.

The Beijing-Washington backdrop is the reason the question is being asked at all. According to wire reports this month, Anthropic CEO Dario Amodei has called on US and Chinese labs to "pace" frontier AI development; Beijing's state-aligned press called the appeal "fear mongering" and "self-serving," and the Trump administration has downplayed the need to slow development. China is not signalling restraint, which is why a 2027 chip schedule and a million-processor architecture read as a strategic gate, not a procurement choice.

The watch item is whether Epoch AI's roadmap and any independent Atlas 960E benchmarks land before HUAWEI CONNECT 2027. Xu has put a date on the promise. The interconnect is the part of that promise an outside reader can actually check.