In early November, Microsoft's patent application became public. It envisions sending AI training jobs to computers and phones while they sit idle, with the user's consent, and paying participants in cash, software store credits or priority access to models.
The patent is titled “Opportunistic Micro-training on Edge Devices Using Spare GPU Utilization”, with publication number US-20260349820-A1 and a publication date of November 5. The patent search service Patentlyze spotted it first, and tech outlets including TechSpot and Insider Gaming followed up.
What the patent says
According to the patent description, a participating device must first be authorized by its user. The system continuously reads the device's CPU and GPU utilization, memory usage, temperature, power draw, video memory bandwidth and compute queue depth, and uses those readings to judge when the device is idle and how much compute it can spare.
Training jobs are split into “micro-tasks” and handed out according to how idle each device is and how accurate its past results have been. The training network shown in the patent's diagrams includes servers, desktops and smartphones, which means phones are within scope.
The patent also contains one more detail: when the GPU is busy or running hot, the system checks in less often to cut down on interruptions for the user. The patent says the aim is to avoid slowing down apps, draining the battery too quickly or overheating the device.
On compensation, the patent lists three forms: money, software credits, or access to the very models the user helped train. Insider Gaming reads this as a setup in which third parties short on computing resources could train models on this capacity, with Microsoft acting as the broker in between.
Someone tried this more than two decades ago
Pooling idle computers to crunch numbers is an old idea. Stanford launched Folding@home in 2000, letting volunteers simulate protein folding on their home machines, and Berkeley's BOINC platform carried the same approach to a range of scientific projects in astronomy, climate research and other fields. Those projects ran on volunteer goodwill and paid nothing.
Paid versions followed. Platforms such as Salad pay gamers by the hour for renting out spare graphics cards to run inference and rendering, and some projects have rewarded contributors of compute with cryptocurrency.
The Microsoft patent differs in two ways. First, the workload is model training, which demands tighter synchronization between devices and stronger verification of results. The patent specifically mentions assigning tasks based on a “reputation for result accuracy.” Second, Microsoft controls entry points such as Windows and Xbox with enormous installed bases, so if it did build this, the cold-start problem would be far easier than for a small platform.
Timing and open questions
The timing of the patent is rather curious. Just days earlier, OpenAI CEO Sam Altman said publicly that having memory sit idle inside home PCs is “terrible” and that it belongs in data centers instead, a remark that drew considerable criticism. The backdrop to this round of debate is that AI data centers are soaking up large amounts of memory and graphics card production, pushing up prices for consumer hardware.
The patent also leaves several gaps. How user data would be isolated from training data, how much GPU a single job would take and how compensation would be calculated are all left unexplained in the filing. Microsoft has not publicly responded to the patent, and a patent application does not mean a product will follow.
Sources: TechSpot, Insider Gaming, CocoLoop, US Patent and Trademark Office public filings. Verified items: patent title, publication number, monitored metrics and forms of compensation.