Huawei Moves Ascend 960DT AI Chip Launch to Q1 2027

Introduction

Huawei is accelerating its effort to build a domestic alternative to Nvidia’s AI computing platform. At its Huawei Connect conference, the company said its next-generation Ascend 960DT accelerator is expected to be ready in the first quarter of 2027. Huawei had previously targeted the third quarter of that year, according to a company spokesperson who spoke with TechCrunch.

The revised schedule is significant because it comes as China’s technology sector continues to operate under U.S. restrictions on access to advanced semiconductor technology. Rather than focusing only on an individual processor, Huawei is presenting a broader approach that combines chips, interconnects, and complete computing systems.

Key points

• An earlier product timeline. Huawei moved the expected readiness of the Ascend 960DT forward by roughly two quarters. The company also said Ascend 960 chips are launching ahead of schedule and described the family as delivering doubled performance and year-over-year advances. The available material does not include test conditions, benchmark results, or a precise definition of that performance claim.
• A system-level strategy. Huawei’s Peerium Computing Architecture is intended to connect processors with memory, storage, and networking equipment. It relies on UnifiedBus, Huawei’s technology for linking those components, and is designed for both AI training and inference.
• Atlas systems as the first implementation. Huawei identified the Atlas 950 SuperPoD and SuperCluster as the first systems based on the architecture. The company says an Atlas 950 SuperCluster can connect as many as 256,000 accelerator cards, reflecting its ambition to combine very large numbers of chips into a single computing resource.
• Unresolved questions about scale. Analyst Rui Ma noted that Huawei had previously described an Atlas 960 SuperPoD capable of scaling to 15,488 Ascend 960 chips, while the latest announcement referred to a 4,096-chip system. The difference suggests that an earlier chip release does not necessarily mean that the largest planned system will be available at the same time.

Why it matters

The announcement strengthens Huawei’s roadmap, but it is not yet proof that the company has closed the gap with Nvidia. For AI developers and infrastructure operators, the practical value of an accelerator depends on more than peak compute. Availability, power efficiency, software support, networking, memory access, and the ability to deploy thousands of units reliably all shape the economics of a cluster.

Huawei’s architecture indicates that the company understands this system-level challenge. By integrating the processor with memory, storage, and networking links, it is attempting to create a platform that can scale beyond individual accelerator cards. The stated focus on both training and inference also suggests an effort to cover the main workloads used by AI companies and research organizations.

At the same time, the revised SuperPoD figures introduce a note of caution. The available report does not provide the Ascend 960DT’s manufacturing process, energy efficiency, software compatibility, or direct comparisons with specific Nvidia products. Nor does it explain why the system scale mentioned this week differs from an earlier plan. Those details will matter when Huawei moves from roadmap announcements to customer deployments.

The broader message is that U.S. restrictions have not ended China’s pursuit of semiconductor self-sufficiency. Huawei is trying to turn earlier chip availability into a larger domestic computing ecosystem. Whether it can do so will depend on manufacturing, packaging, system integration, and software as much as on the accelerator itself.

Source: TechCrunch AI