AXERA unveils 1000+ TOPS AI inference cards and robot brain at WAIC 2026

AXERA, the Chinese AI chip startup, used this year’s World Artificial Intelligence Conference in Shanghai to show off hardware that targets a specific pain point: running large AI models outside the cloud.

The company’s new Yuanxi product line includes an A-series AI inference card rated at over 1000 TOPS, paired with large-capacity memory and high-bandwidth interconnects. AXERA designed the card for high-concurrency workloads — think server clusters processing thousands of video streams simultaneously, or enterprises running private knowledge bases that can’t leave their own infrastructure.

The card builds on AXERA’s self-developed Tongyuan mixed-precision NPU architecture, the same foundation the company uses in its smaller M.2 compute cards and standardized core modules. Together, the three form an edge computing matrix that AXERA says covers everything from lightweight on-device inference to data-center-class deployment.

But the more interesting reveal was the embodied intelligence brain controller. This compute module pushes 1500 TOPS with roughly double the memory bandwidth of competing products. AXERA designed it to support custom operator development, which matters for robotics teams that need to run world models or vision-language-action models without being locked into a fixed pipeline. The controller is meant for general-purpose robots, not a single form factor or use case.

There is also the AX8910, a vision perception chip already in production and shipping inside binocular camera modules from several major robotics solution providers. AXERA says the chip is deployed across commercial and service robots today, not just in reference designs.

AXERA is far from a household name internationally, but the company has been building quietly. Its strategy — custom silicon for specific AI workloads rather than trying to compete with NVIDIA on general-purpose GPUs — mirrors what several Chinese chip startups are attempting. The question is whether the software ecosystem can keep up. A 1000 TOPS card is only useful if developers can actually run their models on it without rewriting everything from scratch.

The company did not announce pricing or availability dates for the new Yuanxi products.