Microsoft Patents AI Training on Idle GPUs of Home PCs

⚡ The gist in 5 seconds

• Microsoft has been granted patent US 12,792,414 B2 for training models on other people's idle GPUs — gaming PCs, laptops, and smartphones.
• A server itself slices the computational graph into chunks, distributes them to devices, and monitors load, memory, temperature, and power consumption.
• No public benchmarks of training speed on a fleet of home machines are provided in the application — only the mechanics.

🔍 What happened

On November 10, 2026, the USPTO published Microsoft's patent grant US 12,792,414 B2, application US 2026/0349820 A1. The text describes a scheme: a server distributes small training tasks to devices with idle compute capacity. The hardware list includes gaming PCs, laptops, and smartphones. It also covers options for voluntary participation and rewards. A breakdown of the patent was published on Habr.

Here's how it works. The server builds a dependency graph and looks for where to cut it. Each cut has three costs: the volume of data exchanged between machines, the extra memory needed for intermediate results, and the frequency of access to data the executor doesn't have. The server picks the cut with the lowest weighted cost. Cutting can be done by model components or by gradient blocks.

An example from the breakdown. Two adjacent transformer blocks sit on different machines. The microbatch is two sequences of 2048 tokens, hidden size 4096, BF16 format. The tensor at the boundary weighs 32 MiB. On an outgoing 20 Mbit/s link, one send takes about 13.4 seconds, and the backward gradient takes just as long.

The scheduler looks at GPU load, memory, temperature, and power consumption. Based on the history of these metrics, it predicts how long a machine will be free: a long window means a bigger task, a choppy schedule means a smaller one. If a threshold is exceeded, the computation is paused or canceled. A separate headache is the late response: if a task was reassigned, both copies may come back, and accepting them as two updates means counting one computation twice.

🧠 Our take

A patent is not a release, and that's the first thing to keep in mind. Microsoft has been patenting schemes for decades that never make it into a product. But the choice of topic itself is telling: the GPU shortage and expensive rentals aren't going anywhere, while the fleet of consumer graphics cards sits idle. Our assessment — the company is planting flags in advance at the crossroads of "distributed training × user hardware," not preparing a launch next week.

What's technically interesting here. 13.4 seconds per 32 MiB at 20 Mbit/s — that is the ceiling of the whole endeavor. Training is bottlenecked not by the card's rated TFLOPS, but by the channel to it. We believe home PCs can handle tasks with long availability windows and low data exchange, but not full training of large models. And one more bit of arithmetic from the breakdown: if a computation takes five seconds while preparation and transfer take fifteen, the executor spends three quarters of the time heating up not the GPU, but the network.

Who wins? In our assessment — whoever already has a cloud and billing, that is, Microsoft itself: other people's graphics cards are cheaper than your own racks. Who loses — low-price GPU rental services, if such a market ever takes shape at all. And it's the PC owner who pays for everything: electricity, fan wear, noise in the room. The patent does mention rewards, but how they're calculated isn't clear from the text.

⚠️ What we don't know

The main weak spot is the numbers. There are no public benchmarks of training speed on a fleet of home PCs in the application — this is stated outright. We read the breakdown, not the full formula from the patent, so details may differ from the primary source. The publication date — November 10, 2026 — looks unusual and should be verified on the USPTO website under US 12,792,414 B2. And it's completely unknown how rewards are calculated, what happens with hardware wear, and whether you can opt out of tasks in the middle of the night.

Prediction: we don't expect a public Microsoft program on consumer GPUs within a quarter after the patent's publication. The first verifiable sign would be a line like "donate GPU" in Windows or Xbox apps; until it appears, this is just talk on paper.

Disclaimer: one source — a patent breakdown on Habr; we did not verify the full text of the application line by line.

Source: Habr