Microsoft patented a technology that will let AI be trained on users’ GPUs: how it will work

Microsoft patented a technology that will let AI be trained on users’ GPUs: how it will work

Microsoft has patented a technology that allows using idle GPU capacity on users’ computers to train artificial intelligence. 

Device owners will be able to voluntarily provide computing resources and potentially receive rewards in return.

The company filed the application on April 30, 2025, and on November 5, 2026 it was published under number US 2026/0349820 A1. Microsoft was granted the patent on November 10.

How the technology will work

Microsoft proposes using computing resources of devices that are not operating at full capacity at a given moment. For example, when a person leaves a computer on but is not running games or other demanding applications, its GPU could perform part of the computations for training AI.

To do this, the system will break large tasks into small pieces and distribute them among connected devices. After completion, it will collect the results and use them to update the AI model.

Diagram of the distributed AI training system described by Microsoft. Source: U.S. Patent and Trademark Office

At the same time, Microsoft foresees a mechanism intended to prevent excessive load on the hardware.

Specifically, the system will monitor:

• CPU and GPU utilization;
• RAM and video memory usage;
• component temperatures;
• power consumption;
• available computing resources of the device.

Based on these metrics, the technology will determine when a computer can perform additional tasks and how many resources are available for that. For example, if a user launches a game or another resource-intensive program, the system can reduce the load on the GPU so as not to interfere with the device’s primary work.

In addition, Microsoft proposes verifying computation results coming from different computers. The system would take into account devices’ previous reliability and compare their results to reduce the risk of errors during AI training.

According to the company’s documents, device owners must voluntarily agree to the use of their computing resources. Microsoft also describes the possibility of rewarding participants depending on the work performed.

Compensation options include special credits that can be exchanged for money, free software, or access to the AI models the user helped train.

Why Microsoft needs this

Training modern AI models requires significant computational power. Technology companies use data centers with thousands of GPUs for this.

In the patent, the company explains that traditional AI training mostly depends on specialized hardware and dedicated compute resources. Distributing tasks across users’ devices would allow using additional capacity without having to place all computations on separate servers.

Read also: OpenAI launches Business Hub in ChatGPT: how it will work