Intel patent puts AI processing inside network switches

Intel patent application US 2026/0312885 A1 describes network switches with built-in AI inference resources that can handle supported requests without sending them to distant data centers. When a request matches a registered AI service, the switch can process it locally and return the result, while other traffic continues through the network as usual.

The system is designed for cloud providers serving multiple tenants. A service operator registers an AI model over a separate management link, providing loading instructions, a tenant ID, a service ID, an SLA and a billing cost. The switch stores the registration, loads the model onto local inference hardware such as GPUs, and uses the tenant’s requirements to decide which resources to assign.

Incoming requests are checked against registered models, queued and scheduled using weighted round robin adjusted by tenant SLAs. If a model is not already loaded, it can be fetched from storage before processing. Intel’s examples include autonomous driving, video analysis for threat detection, augmented reality and speech recognition, where shorter network trips can matter for response time.