Get paid for your GPU? What Microsoft's AI patent promises, and what it leaves out

What this is about

"Microsoft wants to pay you for your graphics card": that's how a patent application is currently making the rounds worldwide, from PC Games Hardware and TechPowerUp to IT之家 in China. On November 5, 2026, Microsoft Technology Licensing published application US 2026/0349820 A1, titled "Opportunistic Micro-training on Edge Devices Using Spare GPU Utilization". The idea is micro-training that runs whenever gaming PCs, laptops or phones are sitting idle. According to RuntimeWire, the application was filed back on April 30, 2025.

We wanted to know whether this is a technical breakthrough. Our thesis: the patent mainly describes a billing and brokerage model. The genuinely hard problems are currently being solved by others, and their numbers show what a home graphics card is actually good for today.

What the patent says, and what it doesn't

According to reports from Insider Gaming, PCGH and IT之家, the application describes the following:

• Splitting: A large training job is broken into independent micro-tasks. The system picks the split points based on dependencies within the model, communication costs, extra memory requirements and where the data lives. Task size is matched to the device.
• Monitoring: The system tracks GPU and CPU utilization, memory, temperature and power draw. If the card is hot or busy, it checks less often; if it's cool and idle, more often.
• Verifying: According to PCGH, identical tasks go to multiple devices. The results are compared so that faulty or tampered contributions stand out.
• Rewarding: Volunteers receive credits, which could be exchanged for software, access to AI models or money. Microsoft would act as a broker for third-party companies that need compute.

Several things are missing. There's no payout rate and no list of supported hardware. Gurugamer and FourWeekMBA also point out that payment appears only in the description, not in the legally binding claims. The reports also don't answer the key technical question: how do the partial results get merged back into one model? We weren't able to access the full text on Google Patents or the USPTO site, so we're relying on the trade press reports.

Why training over the internet is so hard

During training, every machine has to sync its gradients after each step. These are the correction signals indicating which direction each individual value in the model should move. For a 10B model in BF16, that's about 20 GB per step. Epoch AI assumes a typical home upload speed of around 60 Mbit/s. By our math, a single sync would take about 44 minutes; in a data center, it takes a fraction of a second. Epoch estimates that this approach can sensibly train models of at most around 600 million parameters.

Three more hurdles come on top:

• Not enough VRAM: As a rule of thumb, training with the standard Adam optimizer needs about 16 bytes per parameter: weights, gradients and optimizer states combined. With 24 GB and no tricks, you top out at just over 1.5 billion parameters, and that's before accounting for activations.
• Machines drop out constantly: Anyone who launches a game leaves. During the Pluralis run, there were 669 joins and 607 departures over 40 days.
• Cheating: If people get paid, they can send random numbers. Running everything twice, as the patent describes, halves the usable compute. On top of that, different GPUs rarely produce bit-for-bit identical results, so a simple comparison isn't enough.

The tricks that changed a lot in 2025/26

• DiLoCo: Each participant runs hundreds of steps on its own before syncing. According to Epoch, this cuts data traffic by up to 500x.
• SparseLoCo: Only a tiny fraction of the updates gets sent, coarsely quantized to 2 bits. The authors report 97 to 99 percent sparsity while still beating DiLoCo. One example: 4.25 MB instead of 512 MB per message.
• Protocol Models (Pluralis): The model is split into stages, and each machine holds just one of them. The data passed between stages can reportedly be compressed by up to 100x.
• Verification: At Templar, "Gauntlet" assesses whether a contribution actually improved the model on its assigned data. Prime Intellect uses TOPLOC checksums, which reportedly work even on non-deterministic GPU hardware.

The real runs compared

The fine print matters. The largest run, Covenant-72B, used data center machines despite running over the internet. For every 20 minutes of compute, there were only about 70 seconds of communication, with data compressed by more than 146x. The result lands roughly at LLaMA-2-70B level (MMLU 67.1 vs. 65.6), but falls short on HellaSwag and WinoGrande. DiLoCoX and SPES are lab experiments with throttled and fast networks, respectively. So far, the only genuine swarm of home PCs is the Pluralis run, and according to the paper it reached 63 percent of the efficiency of a centralized H100 run.

Epoch AI has quantified the gap: the largest decentralized runs sit at about a thousandth of frontier models. That matches our math: 5 × 10²⁶ divided by 4.8 × 10²³ comes to about 1,040. For Pluralis-8B, the gap is roughly 16,000x. According to Epoch, decentralized training is growing about 20x per year, while centralized training grows only 5x. Even so, Epoch considers it unlikely that the swarm will catch up with the frontier this decade.

What an RTX 4090 could do

On paper, an RTX 4090 delivers about 165 TFLOPS in BF16. Let's optimistically assume half of that is usable, which is what the LLMQ project achieves on consumer cards. Our back-of-the-envelope math then gives:

• Covenant-72B's compute: about 1.6 million GPU hours. 10,000 cards with 4 idle hours a day each would need about 40 days, or more like 60+ with Pluralis-style losses. The bottleneck isn't compute, though, it's memory. By the rule of thumb, the training state of a 72B model alone takes up just over 1 TB, which works out to about 48 cards with 24 GB each for a single copy.
• Grok 4's compute: about 1.7 billion GPU hours. Even a million gaming PCs with 4 idle hours a day would be busy for well over a year.

What's realistic instead comes down to three things:

• RL rollouts: These are trial answers a model generates while learning from rewards. For INTELLECT-2, about 4.5 times as much compute went into these answers on average as into the actual training. According to the technical report, this work can also run on consumer GPUs and is easy to verify.
• Fine-tuning and mid-sized training on your own machine: LLMQ trains a 7B model on a single 16 GB gaming card and a 32B model on four RTX 4090s.
• Individual pipeline stages in Pluralis-style swarms: According to the documentation, you need at least 24 GB of VRAM, 80 GB of RAM and 200 Mbit/s.

The closest match to Microsoft's idea is the research system Cleave. It breaks training into small compute blocks and distributes them from a central server. Its strong results up to 70B, however, come from a simulation, and the central server with 200 Gbit/s is a single point of failure.

What it could earn

These are customer rental prices according to computeprices.com. Platform fees and periods without a renter still need to come off. Vast.ai cites typical earnings of $0.30 to $0.60 per GPU hour for an RTX 5090, but demands high reliability, which is exactly what an "only compute when idle" model can't deliver. Crypto networks are even less predictable. According to taodaily.io, the Templar subnet at times burned 95 percent of rewards, with the rest going only to the top participant. Covenant AI, the team behind Covenant-72B, left Bittensor in a dispute in April 2026, calling the network "decentralization theater".

Our take: at 4 hours a day, an RTX 4090 would net more like €10 to €30 a month after electricity. And that's only if there's a steady stream of jobs, which nobody has demonstrated so far. Extra wear and noise aren't factored in.

The regulatory question

A May 2026 paper (arXiv 2605.29359) worked out how distributed training could be used to sidestep reporting thresholds. According to the paper, the EU AI Act's 10²⁵ FLOP threshold could be crossed with data center hardware for about $31 million, spread across small, inconspicuous nodes. That's only about three times the cost of centralized training. As countermeasures, the author proposes an additional memory threshold for clusters, chip registries and whistleblower programs. The paper didn't model consumer GPUs at all, though.

The bigger picture

Microsoft's patent combines familiar pieces: load monitoring à la Folding@home, verification through redundant computation and a credit system. How the results come together into one large model over slow home connections, the genuinely hard problem, remains unanswered in every report. That's what Prime Intellect, Pluralis and Templar are working on, and a lot really has moved there in 2025/26: 146x compression, a 72B model at LLaMA-2 level, a swarm of home PCs at 63 percent efficiency.

Our take: for frontier pretraining, the gaming PC will remain too small, too poorly connected and too unreliable for the foreseeable future. For RL rollouts, fine-tuning and individual model stages, though, it's a real resource. Just don't expect to get rich. Whether Microsoft ever turns this into a product is entirely open. A patent application is not an announcement.