A patent application filed by Microsoft describes a concept for using the idle GPUs in PCs to train AI models and rewarding their owners. The application, "US 2026/0349820 A1," was published on November 5, 2026. It describes a system that runs small training tasks during periods when a GPU has capacity to spare. However, this is currently only a concept described in a patent filing. Microsoft has not announced a paid service.
Evaluating the concept requires separating three questions: whether idle GPU time can be detected, whether that computing power can be turned into useful training results, and whether the owner is left with an economic benefit. The application goes beyond collecting spare computing power from devices. It also covers splitting up training tasks and verifying and integrating the results returned from each device. It addresses the technical challenges of distributed AI training, which cannot work by simply borrowing idle GPUs.
Rewards for contributing GPUs are envisioned, but no service is planned
The applicant is Microsoft Technology Licensing LLC, and the invention is titled "Opportunistic Micro-training on Edge Devices Using Spare GPU Utilization." According to a summary reproduced by Minesoft, the filing date is April 30, 2025. Note that the descriptions of rewards and user consent below are not based on a direct review of the full publication. They rely on excerpts from the publication carried by outlets such as Insider Gaming.
According to Insider Gaming's report, the patent application describes a mechanism that records the amount of computation a device has processed and grants credits based on that contribution. The credits could potentially be exchanged for software, use of AI models, or money. The system also envisions requiring the user's explicit approval before a device's computing power is used. These details match excerpts published by VideoCardz and FourWeekMBA, but all of them relay the same patent application's text. None independently confirms actual service development or implementation.
It is also unclear whether a reward program will ever be introduced. FourWeekMBA's analysis points out that the reward description is one of the embodiments in the specification and is not included in the 20 claims the outlet reviewed. From the materials reviewed so far, details such as reward amounts, launch timing, supported GPUs, and target regions are unknown. A reward mechanism appearing in a patent application is clearly distinct from its being offered to users as a real service.
Three steps for putting idle GPUs to work on AI training
According to the summary of the patent application reproduced by Minesoft, the proposed system consists of three main steps. The first step analyzes the dependencies inside an AI model and splits the work into small training tasks that can run independently. The split also takes into account communication costs, additional memory load, and where the required data is placed. The key is not simply to chop up the computation, but to make it possible to run efficiently even on the limited resources of each device.
The next step adjusts the size of the tasks assigned according to the performance of each participating device and its past patterns of idle time. Training runs only when the device has enough headroom. According to the publication as presented by Insider Gaming and VideoCardz, the monitored factors include GPU and CPU utilization as well as memory status, temperature, and power consumption. The system also envisions checking device status less often when the device is under heavy load or running hot, and more often when the load is low and the device is cool.
The final step verifies the training results returned from each device and integrates them into the AI model asynchronously. Asynchronous processing here means taking in results as they come back, rather than waiting for every device to finish and updating all at once. The summary explains that the weight given to each result when it is reflected in the model is adjusted according to the result's reliability, how stale the update is, and the trustworthiness of the device that handled the processing. In other words, the system doesn't just gather computation results from distributed devices. It also builds in a way to judge whether those results are suitable for training.
The summary of the patent application also includes a statement to the effect that the impact on normal device use would be eliminated. However, the materials reviewed contain no measured results for game frame rates, power consumption, or similar metrics. A design that monitors device status and adjusts when training runs is meant to limit the impact on ordinary use, but how much the load can actually be reduced will need to be verified after implementation.
Low GPU utilization doesn't necessarily mean spare training capacity
Low GPU utilization does not mean all of the unused capacity can be put toward AI training. In a paper presented at ICSE 2024, Gao et al. examined 706 problems across 400 jobs, selected from jobs run on an internal Microsoft platform with average GPU utilization of 50% or less. The 50% threshold was a condition for selecting the jobs studied. It does not represent average GPU utilization across Microsoft or in ordinary home PCs. The study is also separate from this patent application and does not demonstrate the performance of the proposed system.
The study cites batch sizes that are too small as one cause of low GPU utilization. Batch size refers to the amount of data fed into the model in a single processing step. It also found cases where saving large models or synchronous remote transfers repeatedly left the GPU waiting, lowering utilization. In other words, low GPU utilization includes not only cases where computing resources are truly idle, but also cases where the GPU cannot be kept fully busy because of how data is supplied or because of transfer waits.
This study shows that gathering the idle time of home PCs and eliminating inefficiencies inside AI training jobs are different technical challenges. Adding more participating devices does not necessarily resolve communication and data-placement problems at the same rate. Assessing the real-world performance of distributed AI training requires considering not only the available GPU time, but also whether tasks can be distributed, computation executed, and valid results collected within that time. That said, no direct technical link has been confirmed between this patent application and the earlier study.
What it would take for idle GPU use to produce real benefits
For users who contribute their GPUs, the amount of computation processed must be distinguished from the economic benefit actually received. VideoCardz's report also notes that the additional power consumption from training would affect profitability. Even if rewards were paid in proportion to computation, it would be impossible to tell whether participation leaves a profit without knowing the increase in electricity costs. The impact on normal PC operation would also be an important factor for users deciding whether to take part.
For the concept to work, three conditions must each be met: a device's spare computing power must be usable, the aggregated results must be useful for AI training, and participants must receive compensation they find acceptable. The table below organizes the items to be evaluated based on the summary of the patent application published on November 5, 2026, statements quoted in press reports, and VideoCardz's points on profitability. The right-hand column lists matters that would need verification during implementation. It does not mean performance has already been confirmed.
| Item to evaluate | Mechanism described in the patent application | Matters to confirm during implementation |
|---|---|---|
| Detecting idle time | Monitors device status and runs tasks during periods with spare capacity | Impact on normal PC operation, game performance, and responsiveness |
| Training results | Splits training into small tasks, verifies results, and integrates them according to staleness and reliability | Whether valid training results are actually obtained, including communication and verification overhead |
| Participant benefit | Envisions recording processed computation and granting credits as an option | Relationship between reward amount and added power consumption; actual profitability |
These cannot be compared on a single metric. Each is a separate challenge that must be solved to make the concept practical. Even if idle GPU time can be detected accurately, the results returned from each device will not necessarily improve the AI model. Likewise, even if useful training results are obtained, that alone does not guarantee profitability for GPU owners.
If Microsoft turns this concept into an actual service, the points to watch most closely will be the reward structure and measured data on the added burden from training. Once it is clear how computation would be converted into rewards, and how much power consumption and PC responsiveness change before and after participation, users will be able to judge whether contributing their GPUs is worthwhile. This patent application presents a technical concept for putting computing resources lying idle in PCs to work on AI training, but its practicality and economic viability will have to be determined through concrete implementation and testing.