Sources: Huawei's Ascend chips still lag far behind Nvidia's for model training and have stability issues, slower inter-chip connectivity, and inferior software
Tech group's Ascend artificial intelligence chips are being widely adopted but Chinese companies complain of performance problems
Context & Ripple Effects
Huawei’s Ascend push was already constrained by reported difficulty scaling production of its leading AI server chip under a new US crackdown. Chinese internet companies had also begun testing the Ascend 910C, making real-world training performance—not only claimed chip specifications—a central test of its positioning against Nvidia.
The report matters because it separates adoption driven by the need for a domestic option from readiness for the most demanding training workloads. It also highlights that interconnects and software are part of the competitive product, not peripheral features.
First-order effects
- Chinese companies using Ascend for model training face lower effective performance and operational risk from stability, connectivity and software shortcomings.
- Huawei’s near-term competitive case is weakened in training workloads, while Nvidia retains an advantage where customers need dependable multi-chip training systems.
Second-order effects
- Customers are likely to segment workloads more carefully, reserving the most demanding training jobs for platforms with stronger software and inter-chip performance while continuing to evaluate Ascend where it fits.
- Huawei must improve the surrounding system stack—not just chip capability—to convert broad adoption into repeat use for large-scale training; that raises the importance of its software and networking execution relative to rival hardware offerings.
Third-order effects
- If the gap persists, China’s AI-compute market could develop as a heterogeneous environment: domestic accelerators serving selected workloads while Nvidia-compatible systems remain the benchmark for frontier training.
- The episode reinforces that supply constraints can create demand for alternatives without eliminating the integration bottleneck; durable competition depends on mature hardware, networking and developer software together.
The trend: AI accelerator competition is shifting from peak-chip claims toward proof that an integrated compute stack can run large training workloads reliably at scale.