Axera Unveils 'Yuanxi' High-Compute AI Inference Chip Series at WAIC 2026
The AMW Read
An existing HK-listed edge-inference chipmaker adds a 1,000+ TOPS product tier plus a dedicated, already-design-won embodied-intelligence chip stack — a meaningful silicon-layer update, not a new entrant or debate resolution.
Axera Unveils 'Yuanxi' High-Compute AI Inference Chip Series at WAIC 2026
Axera (爱芯元智), the Hong Kong-listed AI inference chip supplier (0600.HK), showcased its full AI hardware ecosystem at WAIC 2026 in Shanghai on July 20 and unveiled its Yuanxi series of high-compute inference products. The Yuanxi A-series delivers over 1,000 TOPS of inference compute with high memory bandwidth, targeting server-cluster and multi-channel video-analytics deployments, extending Axera's M.2 cards and core modules into a layered edge-compute lineup built on its self-developed mixed-precision NPU. Axera also introduced a 1,500-TOPS embodied-intelligence controller with roughly double the industry's leading memory bandwidth, AEC-Q100 safety qualification, and support for world-model and VLA algorithms, alongside its AX8910 vision chip, already shipping inside binocular-camera modules from multiple robotics vendors.
The launch targets enterprises running models locally to cut cloud token costs, avoid compute bottlenecks, and meet data-localization rules. By pushing four-figure-TOPS compute and large on-card memory to edge hardware, Axera positions its chips as a substitute for cloud GPU inference in high-throughput settings, not just low-power endpoints. The parallel embodied-intelligence push — a dedicated controller paired with a vision chip already carrying camera-module design wins — shows a Chinese edge-silicon vendor building specialized robotics compute as its own product line.
For builders, the layered lineup — core modules through 1,000+ TOPS cards — lets edge-AI teams scale from prototype to production on one vendor without re-architecting around cloud APIs. For investors tracking robotics silicon, AX8910's shipping design wins with camera-module makers are a stronger signal than roadmap claims, worth weighing against rival edge and embodied-AI chip vendors chasing the same integrators.