Why Huawei Says Ascend Has Reached an Ecosystem Inflection Point

At HUAWEI CONNECT 2026, Huawei computing strategist Zhu Zhaosheng described Ascend as having passed an ecosystem inflection point. The argument was about more than chip capability: Huawei pointed to an expanding software and developer base, alongside shared compute that partners could use to build on Ascend. Zhu paired that ecosystem claim with a proposed computing approach centered on SuperPoDs and SuperClusters. 221

Why long-running agents change the computing problem

An agent working through a lengthy task may need to resume execution, retain state and repeatedly produce a timely first token. As workloads span more NPUs, moving data between them and accommodating growing memory demands become part of the application problem—not just infrastructure details. Huawei’s proposal treats the SuperPoD or SuperCluster as the scale at which those resources should be programmed. Its separate UnifiedBus announcement likewise frames interconnect and per-NPU memory requirements as cluster-level challenges. 26

The five proposed shifts

1. Compose resources across nodes. Huawei proposes APIs for combining compute, communication and memory across a SuperPoD or SuperCluster. The aim is to make cross-node work easier to express while addressing the transfers and memory demands that can constrain larger agent workloads. That is a design goal, not evidence that interconnect costs disappear. 24
2. Give agents an execution model that can retain state. ThinkPro and openEuler’s ThinkProcess are intended to manage an agent’s execution across a long task. Preserving useful state could reduce work when an agent resumes and help responsiveness, including time-to-first-token; the cited material does not establish a measured latency improvement. 2
3. Support more than one NPU programming style. Huawei positions Ascend 950’s SIMD and SIMT capabilities, together with Regbase programming, as ways to handle both dense computation and less regular operations. The practical benefit would depend on how well particular workloads map to those capabilities. 2
4. Match precision to the workload. Huawei proposes HiF4 as a lower-precision option intended to improve throughput and reduce memory use where a model can tolerate it. Any speed or capacity gain has to be weighed against accuracy on the workload being run. 2
5. Use agents to help develop operators. Agent-assisted creation and optimization of operators is meant to help new model techniques reach efficient Ascend implementations faster. It remains a proposed development approach rather than a demonstrated fix for every software bottleneck. 2

What makes this an ecosystem claim

The technical plan depends on developers being able to use it. Huawei points to an open PTO instruction set, community-driven CANN development, model and framework support, developer tools, and shared 10k-NPU-scale compute for partners. A report on Zhu’s remarks also cited more than 5,200 monthly active users in the CANN community. These are indicators of access and activity, not proof that every workload is portable or fast. 2421

The distinction matters: Zhu’s inflection point is Huawei’s assessment that Ascend now has enough surrounding software, contributors and compute access to support a broader agent ecosystem. The five shifts explain how Huawei intends to tackle first-token responsiveness, cross-node communication and memory pressure. Whether they deliver those benefits for long-horizon agents requires workload-specific results beyond the announcements cited here. 26