Microsoft patent describes paying gamers to train AI on idle GPUs

Felix Laurent 11 Nov 2026 4 min read

Big tech

Microsoft has patented a way to train AI models on idle phones, laptops and gaming PCs, and to reward their owners for it. The application, titled "Opportunistic Micro-training on Edge Devices Using Spare GPU Utilization," was made public on November 5, 2026. TechPowerUp reported it on November 10.

The filing, US 2026/0349820 A1, comes from Microsoft Technology Licensing and was submitted in April 2025. It describes owners who opt in and earn credits for the work their hardware does. Microsoft has not announced any product based on it.

How Microsoft's idle GPU training would work

The system takes a large AI training job and splits it into small micro-tasks, each sized to what a single device can handle. According to Wccftech, it first analyzes the model's computational dependencies. It weighs communication costs, memory needs and data locality before deciding how to divide the work.

The tasks then go to consumer devices whenever they have spare capacity. The patent names gaming desktops, laptops and smartphones. Assignments follow each device's free resources and its predicted idle periods.

To stay out of the owner's way, the system watches several signals:

GPU and CPU load;
memory use;
temperature;
power draw.

It checks less often when the GPU is busy or hot. Training can slow down or stop when temperatures rise or the owner launches a demanding app, such as a game.

Some devices will be slow, faulty or dishonest, so the results are cross-checked against other machines. Devices with a good track record get more trust. The system also combines contributions without needing every computer to stay online, and the finished results go back to a central service.

The most talked-about part of the filing is the reward.

What PC owners would get

Owners would have to explicitly opt in. The system would then track how much work each device completes and add credits to the owner's account. According to the filing, those credits could be redeemed for free software, for access to the trained AI model or for money.

That part is weaker than headlines suggest. The application has 20 claims, which define what the patent would legally protect. None of them mentions paying participants, according to tbreak, citing FourWeekMBA's reading of the filing. Payment appears only in the description, as one possible way to use the technology.

The filing gives no payout rate and lists no supported hardware. It does not mention Windows or Xbox. A published application also is not a granted patent, since publication does not mean the claims have been examined.

The open question for owners is whether any reward would cover the cost.

Would it be worth it for gamers

The electricity bill would land on the owner. PC Games Hardware, cited by Wccftech, estimates that an extra 250 watts for four hours a day adds about 30 kWh a month. At a typical US home rate of around 18 cents per kWh, that is roughly $5 a month. A high-end card such as the RTX 5090 can draw close to 600 watts under heavy load.

Other costs are harder to price. Long training runs add heat and wear to a graphics card that many gamers paid well over $1,000 for. Data privacy is another gap, since the filing leaves open what data the devices would see and process.

TechPowerUp also points to trust. Plenty of people already have doubts about AI training, and running it on their own hardware may be an even harder sell. The site also cites Microsoft's own "Copilot for Gaming fiasco" as a reason to read the fine print before opting in.

The idea itself is not new.

Paying for spare compute is an old idea

Volunteer computing has existed for decades. Folding@home has used home computers to simulate protein folding since 2000. Services such as Salad already pay gamers to rent out idle graphics cards for AI and other workloads. Startups like Prime Intellect have trained language models across GPUs spread around the world.

Microsoft's version would bring that model to a company with a huge installed base. Windows runs on most of the world's PCs, and Microsoft sells the Xbox. The timing also fits a compute shortage. New data centers face growing local pushback, as with the $68 billion in stalled projects we covered after the midterms. GPU and memory prices are climbing again. OpenAI CEO Sam Altman has even argued that the memory sitting in home PCs is a poor use of scarce supply.

Distributed training still has technical limits. Large models need fast links between chips, and home internet connections are slow compared with a data center. That is likely why the filing focuses on small, independent micro-tasks instead of training a whole frontier model.

Will Microsoft pay PC owners to train AI

Not soon, and maybe never. This is a patent application, and big tech companies file many that never become products. Payment is not even part of the claims. Microsoft has not commented and has not announced a Windows or Xbox service.

The filing does show that Microsoft sees consumer devices as a possible pool of AI compute. If a service like this ever launches, the real test will be the rate per hour against the extra power, heat and wear. Until Microsoft names a price, gamers have little reason to want their GPU working for someone else's model.