What are Huawei’s plans for launching its 960DT and Ascend 960PR AI chips in 2027, how does its UnifiedBus technology enable large numbers of chips to function as integrated AI computing systems—including superclusters of up to 1 million processors and more than 1,000 supernodes

Huawei says it will launch the Ascend 960DT in the first quarter of 2027 and the Ascend 960PR in the third quarter. The strategy is not merely to sell faster chips, but to sell an integrated domestic AI-computing stack—chips, interconnect, supernodes, clusters, and software—to reduce China’s dependence on Nvidia. 1

• Launch roadmap: Huawei’s announced 2027 schedule puts the 960DT first, followed by the Ascend 960PR. The company is positioning the 960 generation as the next step in its Ascend AI-computing roadmap. 1

• What UnifiedBus does: UnifiedBus is Huawei’s high-speed interconnect fabric. It is intended to let many Ascend processors exchange data with sufficiently low latency and high bandwidth that software can treat them as a coordinated AI-computing system rather than isolated accelerators. An upgraded implementation is reported to connect up to 4,000 processors, with optical links extending deeper into the system. 2

• From chips to superclusters: Huawei’s architecture groups processors into SuperPoDs/supernodes and then combines multiple SuperPoDs into a SuperCluster. Huawei has announced an Atlas 960 SuperCluster with more than one million Ascend NPUs; this is a scale-out system claim, not a claim that every processor is directly connected in one flat network. 3

• Commercial proof point, with a caveat: Huawei says it has shipped more than 1,000 Ascend supernodes to over 370 customers. That indicates its systems have moved beyond lab demonstrations, but the shipment and customer figures are company-reported and do not independently establish comparative performance, utilization, or profitability. 4

• Why export controls matter: U.S. restrictions on advanced AI chips limit Chinese access to leading Nvidia accelerators. This creates both a supply constraint and a strategic opening for Huawei: Chinese cloud operators, public-sector buyers, and model developers have stronger incentives to qualify domestic hardware and avoid dependence on a supply chain vulnerable to further controls. 1

• China’s domestic-alternative push: Huawei can align its offer with China’s goal of indigenous AI infrastructure. Its competitive pitch is resilience and availability as much as per-chip benchmark leadership: a locally supported, end-to-end platform that can be deployed at national or enterprise scale.

• Nvidia remains the harder obstacle: Nvidia’s advantage is not just hardware. Its CUDA software ecosystem, mature tools, libraries, developer skills, and broad application compatibility create high switching costs. Huawei therefore has to make migration, model tuning, operations, and scaling on Ascend sufficiently practical—not simply claim higher system-level compute.

• Software ecosystem response: Huawei’s reported 5,270 monthly active developers in its CANN community is meant to signal ecosystem momentum. That developer base can improve frameworks, operators, model ports, and tooling, but it remains far smaller and less entrenched than Nvidia’s global CUDA ecosystem; the key test is whether it generates production-grade software and repeat deployments rather than activity alone. 5

In short, Huawei is trying to offset Nvidia’s hardware and CUDA lead through system-scale interconnect, domestic supply, policy-aligned demand, and an expanding Ascend/CANN developer ecosystem. Its strongest near-term opportunity is China’s constrained market; matching Nvidia globally will depend on software compatibility, reliable mass deployment, and real-world price-performance.