US 2026/0349820 A1 describes a proposal to divide AI-model training into small tasks and send them to devices with spare GPU capacity. Technology coverage published November 9–10, 2026, presented it as a proposal rather than a consumer-service announcement. BetaNews and TechPowerUp reported that Microsoft Technology Licensing filed the application on April 30, 2025, and that it was published on November 5, 2026.
A possible reward described in a patent application is not a payment offer users can claim.
How the proposed GPU sharing would work
The design would break training into smaller jobs and assign them to devices when resources are available. BetaNews reports that the system would map dependencies between parts of a model, then weigh possible task divisions by how much data devices would need to exchange, how much temporary memory the work would require, and how often it would fetch remote data. That means the reported design considers communication and memory overhead, not just whether a GPU has spare capacity.
BetaNews also describes monitoring GPU and CPU load, memory use, temperature, power draw, memory bandwidth, and the depth of computing queues. The proposed controls could pause, stop, or cancel a task when readings or performance measures cross set limits. Results from devices assigned the same task would be compared, with updates weighted according to agreement, age, and each device’s validation history.
What the reports say about rewards
BetaNews reports that the application’s description discusses credits that could be redeemed for money, software, or access to a model. The same report says none of the application’s 20 claims mentions payment. That distinction leaves the reward as a possibility described in the application, not a guaranteed payout or an operating service.