Huawei Ascend AI chips target 10 trillion parameter models by 2027

The launch of the Atlas 960E SuperPoD signals a shift toward optical interconnects to solve the bandwidth bottlenecks inherent in massive AI clusters. By utilizing near-packaged optics and the Tau Scaling Law, Huawei is attempting to maintain technical parity with Nvidia despite U.S. export restrictions on advanced lithography. For the streaming and media ecosystem, this localized hardware stack ensures that Chinese firms can continue developing high-parameter generative AI and recommendation engines without relying on Western silicon. Watch for the Q1 2027 release of the 960DT to see if real-world training benchmarks match Huawei's performance claims for 10 trillion parameter models.

Huawei's Ascend roadmap extends beyond individual chip launches into a broader ecosystem strategy centered on optical interconnects and proprietary scaling architectures. At Huawei Connect 2025, the company unveiled the Atlas 960E SuperPoD as a rack-scale system designed to connect up to 100,000 Ascend processors through optical links, positioning the architecture as a direct alternative to Nvidia's NVLink and NVSwitch interconnect fabric. The near-packaged optics approach reduces electrical signal degradation across large clusters, a critical bottleneck when scaling beyond single-rack deployments. Huawei's Tau Scaling Law, which the company claims governs performance gains across its Ascend generations, is intended to provide a predictable cadence comparable to Moore's Law for the AI training market. The business context for these launches is shaped by ongoing U.S. export controls that restrict Huawei's access to advanced lithography below 7 nanometers. In September 2025, the U.S. Commerce Department expanded entity-list restrictions affecting semiconductor equipment suppliers to Chinese AI chip developers, tightening the constraints that Huawei's optical-interconnect strategy is designed to circumvent. Huawei has responded by investing heavily in chiplet-based designs and optical I/O to compensate for process-node limitations. Eric Xu, Huawei's rotating chairman, stated at the conference that the company plans to ship Ascend 950-series processors in volume by late 2026, with the 960DT targeting 10 trillion parameter model training by 2027. This timeline places Huawei in direct competition with Nvidia's Rubin architecture, which is expected to reach volume production in the same window. On the technical side, Huawei's optical interconnect strategy mirrors approaches being pursued by other AI infrastructure vendors seeking to overcome bandwidth walls in large-scale training clusters. Broadcom demonstrated 51.2 Tbps co-packaged optics switches at OFC 2025, while Nvidia's own NVLink 6.0 specification incorporates optical links for its next-generation GB300 systems. For streaming and media companies operating in China, the Ascend ecosystem's maturation means that generative recommender training and generative AI workloads can remain on domestic hardware without performance regression. The key question for buyers evaluating Ascend-based clusters is whether Huawei's software stack, particularly its CANN (Compute Architecture for Neural Networks) framework, can match the developer productivity that Nvidia's CUDA ecosystem provides for video AI pipelines including content recommendation, automated metadata tagging, and real-time transcoding optimization.