Advertise
AIComputers

Microsoft Patent Filing Proposes Training AI on Spare GPU Time

The application, published November 5, 2026, outlines dividing training tasks across gaming PCs, laptops and smartphones, with owners possibly receiving credits.

Microsoft has had a US patent application published for a system that trains AI models using spare GPU capacity on everyday devices, according to gamebusiness.jp. The application, titled "Opportunistic Micro-training on Edge Devices Using Spare GPU Utilization," was published on November 5, 2026.

The document describes splitting training work into small tasks and distributing them to gaming PCs, laptops and smartphones based on each device's performance and expected GPU idle time. Results are then verified and fed into model updates.

The system monitors GPU utilization and device temperature, pausing or aborting tasks when load or temperature exceeds set levels.

Credits for contributed compute

An example in the application suggests device owners could receive credits based on the compute they provide, exchangeable for free software, rights to use the trained AI model or money. The claims section, which runs to 20 claims, does not describe any compensation.

Microsoft filed the application on April 30, 2025. No product or service based on the application has been announced. Users on the overseas forum Reddit questioned whether Microsoft would pay enough to cover electricity costs and hardware wear from increased device operation.

Quick answers

What does Microsoft's GPU-training patent application describe?

It describes splitting AI training work into small tasks and distributing them to gaming PCs, laptops and smartphones based on each device's performance and expected GPU idle time.

Will device owners be paid for lending their GPU?

The application gives an example of credits exchangeable for free software, rights to use the trained AI model or money, but its claims section does not describe compensation. No product based on the filing has been announced.

When was the application published?

It was published on November 5, 2026, after being filed on April 30, 2025.

Source