WebNN Becomes First Web API to Access NPU Hardware for AI Inference

The W3C published an updated Candidate Recommendation of the Web Neural Network API (WebNN) on 22 January 2026, marking it as the first and only web standard capable of reaching a device's Neural Processing Unit (NPU). NPUs are power-efficient chips built specifically for neural network inference and are found in most laptops sold in recent years, yet browsers have had no way to utilize them until now. Unlike WebGPU, which targets the GPU and competes with other system tasks, WebNN describes a full computational graph that the browser compiles and hands off to platform-specific accelerators such as DirectML on Windows or Core ML on Apple devices. Current support is limited to Chromium-based browsers — Chrome and Edge — across ChromeOS, Linux, macOS, Windows, and Android, while Firefox and Safari have yet to implement the spec. Since two independent passing implementations are required before WebNN can become a full W3C Recommendation, developers are advised to treat it as a progressive enhancement and provide fallbacks to WebGPU or WASM.