Somewhere between a gamer's quiet afternoon and a data center's endless hum, Microsoft has sketched an idea in a patent filing: that the idle graphics chips in ordinary homes might one day lend their strength to training artificial minds, and be paid for it. The document was made public in early November, and it describes a system that would split large AI training jobs into small pieces and send them to phones, laptops and gaming PCs that have spare capacity, with owners opting in and earning credits, free software, model access or cash. Yet it stops short of the things that matter most to a
Microsoft Patent Would Pay PC Owners to Train AI on Their Idle GPUs
Getting paid for idle silicon sounds tempting, until you read the fine print.
Mark
So what exactly is Microsoft proposing here?
Mimi
A system that splits big AI training jobs into tiny pieces and sends them to ordinary devices, phones, laptops, gaming PCs, when their GPUs aren't busy. Owners opt in and earn credits redeemable for software, model access or money.
Luke
Worth stressing it's a patent application, filed April 2025 and published November 5. Microsoft hasn't announced any product, so nothing here is confirmed to be coming.
Mark
How does it avoid slowing down someone's machine?
Mimi
It watches GPU and CPU load, memory, temperature and power draw, and checks less often when the GPU is hot or busy. Tasks can be paused or stopped when limits are crossed.
Luke
That's what the filing describes, not something tested. We don't know how well it would work in practice.
Mark
What about trust in the results from random devices?
Mimi
Results get cross-checked across devices to catch faulty or tampered work.
Mark
And would people actually make money?
Mimi
That's the open question. The filing gives no payout rate and no hardware list.
Luke
Right, so the electricity-cost worry from commenters can't be answered yet. Anyone quoting an earnings figure would be making it up.
Mark
Why does the idea resonate now?
Mimi
Graphics card prices are rising again, and Sam Altman has argued that home PC memory used about 5% of the time is a poor use of scarce supply.
Luke
That's his argument, not a measured fact I'd lean on. And the report itself says last year's Copilot for Gaming fiasco is a reason to read the terms carefully.
Le Pouls
- A Microsoft patent filing made public on November 5 imagines idle gaming PCs and phones quietly training AI models for pay.
- The filing leaves the hardest questions open, with no payout rate, no hardware list, and no earnings figure to set against electricity costs.
- Safeguards that monitor load, temperature, memory and power draw are meant to keep the work from ever slowing the person at the keyboard.
- Graphics card prices climbing again make the offer of paid idle silicon appealing, even as readers recall the Copilot for Gaming fiasco and ask whether to trust it.
- Nothing has been announced as a product, so the story now waits on whether Microsoft turns the patent into a service and on what terms.
Somewhere between a gamer's quiet afternoon and a data center's endless hum, Microsoft has sketched an idea in a patent filing: that the idle graphics chips in ordinary homes might one day lend their strength to training artificial minds, and be paid for it. The document was made public in early November, and it describes a system that would split large AI training jobs into small pieces and send them to phones, laptops and gaming PCs that have spare capacity, with owners opting in and earning credits, free software, model access or cash. Yet it stops short of the things that matter most to a person deciding whether to participate: no payout rate, no list of supported hardware, and no announced product. Like many patents, it may remain a sketch, a record of what was once imagined and perhaps never built.
A Microsoft patent application, published on November 5 under the title "Opportunistic Micro-training on Edge Devices Using Spare GPU Utilization," imagines a future where the unused power of everyday machines becomes useful. The idea is to break large AI training jobs into small micro-tasks, sized to fit whatever a single device can handle, and send them to phones, laptops and gaming PCs whenever they have spare capacity. Nothing runs without the owner's explicit consent, and owners would earn credits redeemable for software, model access or money.
The design leans heavily on restraint. The system watches processor and graphics load, memory, temperature and power draw, backing off when a machine is busy or running hot, and pausing work when limits are crossed. Results are cross-checked between devices so faulty or tampered output can be caught. The intention is clear: the person using the machine should never feel the cost.
What the filing does not say matters just as much. It names no payout rate and no supported hardware, and Microsoft has announced no Windows or Xbox service. Whether rewards would outweigh the electricity a PC consumes is simply unknown. Patents are also not products, and many never leave the drawing board.
The timing gives the idea some pull. Graphics card prices are rising again, and the notion of being paid for silicon that sits idle is attractive. Readers responded with a mix of interest and caution: some would accept it if it stays opt-in and pays fairly, others compared it to the cryptomining era, and many recalled last year's Copilot for Gaming fiasco as a reason to read the fine print. Trust, and the broader unease about AI training, will be a steep hurdle. For now, the thing to watch is whether a real service appears, and if it does, what it pays and on what terms.
A Microsoft patent application made public on November 5 describes a way for idle gaming PCs to earn money for their owners by training AI models. The filing, numbered US-20260349820-A1 and titled "Opportunistic Micro-training on Edge Devices Using Spare GPU Utilization," was submitted in April 2025 under the name Microsoft Technology Licensing.
The core idea is to break a large AI training job into many small micro-tasks, each sized to what an individual device can handle. Those pieces would then go out to phones, laptops and gaming PCs whenever they have capacity to spare. Owners would have to opt in explicitly before any of their hardware was used.
Once a device is taking part, the system would keep track of the work it completes and apply credits to the owner's account. According to the filing, those credits could be redeemed for free software, access to the trained model, or money.
The design includes safeguards meant to keep the work from getting in the way of the person using the machine. The system monitors GPU and CPU load, memory use, temperature and power draw, and checks less often when the GPU is busy or running hot. Tasks can be paused or stopped when set limits are crossed. Results are also cross-checked across devices, so that faulty or tampered work can be caught.
What the filing leaves out is considerable. It gives no payout rate and does not list which hardware would be supported. Microsoft has not announced a Windows or Xbox service built on the idea. That means there is no way to tell whether the rewards would cover the extra electricity a PC would draw while training a model. Patent applications are also not products, and many never become one.
The timing fits a wider conversation about hardware. Graphics card prices are climbing again, which makes the prospect of being paid for idle silicon more appealing. TechPowerUp's report also points to OpenAI's Sam Altman, who argued that DRAM in home PCs, used around 5% of the time, is a poor use of limited supply.
Trust is a hurdle. The report cites last year's Copilot for Gaming fiasco as a reason to read the fine print before opting in to any such service. It also notes that plenty of people already have reservations about AI training in general, and running it on their own hardware will probably be a tougher sell still.
Readers of the original article reacted with a mix of interest and skepticism. One commenter said they would accept it as long as it is opt-in and payments are competitive. Another asked whether payments would outweigh electricity costs. A third compared it to GPU cryptomining, saying people had already done this once before.
The things to watch are whether Microsoft announces an actual service, and if it does, what it pays and on what terms.
Citations marquantes
Argued that DRAM in home PCs, used around 5% of the time, is a poor use of limited supply.— Sam Altman, OpenAI, as cited by TechPowerUp
Envie de l'histoire complète ?
Lire l'original sur TechPowerUp ↗