Huawei Declares Ecosystem Inflection Point for Ascend AI and Outlines 2029 Roadmap
Huawei announced that its Ascend AI ecosystem has reached a major inflection point, with over 5,200 monthly active developers on its open-source CANN software stack and non-Huawei contributors now exceeding internal ones. The company also disclosed its multi-year chip roadmap extending through 2029, starting with the commercial cloud launch of the Ascend 950 cluster. As export restrictions limit Chinese tech companies' access to advanced Nvidia GPUs, Huawei's progress underscores Ascend's emerging role as a viable domestic alternative for large-scale AI model training. The rapid growth of its open-source developer ecosystem and annual hardware roadmap demonstrate China's aggressive push toward AI hardware self-reliance in the emerging era of Agentic AI. Over 40 large language and multimodal models have completed pre-training on Ascend, making it China's only domestic AI computing platform supporting full pre-training workflows. The future roadmap includes the Ascend 960 supernode featuring Near-Package Optics (NPO) technology scheduled for 2027, followed by annual iterations for the Ascend 970 in 2028 and Ascend 980 in 2029.
## BACKGROUND
Ascend AI hardware relies on Huawei's Compute Architecture for Neural Networks (CANN), a heterogeneous computing software stack that bridges low-level hardware with popular frameworks like PyTorch. To sustain performance scaling for Agentic AI—autonomous systems designed to execute complex tasks with minimal supervision—Huawei is adopting Near-Package Optics (NPO), which places optical engines directly on the PCB near the chip package to boost bandwidth and reduce interconnect power consumption.