Overview
China’s domestic compute base is iterating fast in 2026: at WAIC 2026 Huawei debuted the Ascend 950 SuperPoD with 1 EFLOPS of FP8 compute per node; concurrently, a Beijing forum on neuromorphic, intelligent-compute and chip technology targeted the full chain from bottom-layer compute to robot applications.
Background & Interpretation
1. Technology Evolution
Compute energy use is climbing — 2025 domestic data-centre power hit 170 TWh, +30% yoy, with 2030 forecasts above 800 TWh — forcing high-efficiency hardware. National plans raise domestic-chip share, maturing local AI chips on the inference side.
2. Core Drivers and Mechanism
The Ascend 950 node provides the FP8 base for LLM training and complex embodied simulation; neuromorphic and intelligent-compute chips decide whether robot “brains” can truly think and act — the bottom-layer gap the industry must fill. Open-compute paradigms lower the dev barrier via unified architectures.
3. Market Structure and Impact
Localisation extends from training to inference; Huawei, Cambricon and Enflame are accelerating适配. Edge scenarios like space compute (37 advanced GPUs in orbit) are emerging, diversifying supply and easing single-vendor risk.
Implications & Outlook
Near term, better domestic price-performance accelerates inference-side migration. Mid term, energy efficiency and software ecosystem are the real battleground. Watch framework适配, green-compute policy and embodied workload curves.
Takeaways for Industry Participants
Cloud and AI firms: pilot domestic-compute substitution on inference to control supply risk. Robot firms: bind early to compute and simulation platforms to shorten real-hardware cycles. Procurement: include efficiency in TCO.
This column compiles industry information and shares technical perspectives; it does not constitute investment advice.