Microsoft Wins Patent for AI Training Across Idle Personal Devices

Microsoft received a patent for a proposed system that would distribute AI training work across personal devices using their spare computing capacity.

U.S. Patent No. 12,792,414 B2 was granted Tuesday, Nov. 10. The application was filed as No. 19/195,047 on April 30, 2025, and published Thursday, Nov. 5, 2026, as US 2026/0349820 A1. The grant lists 20 claims and names Aditya Vasanth Challapally as inventor.

The design breaks model training into independently trainable “micro-tasks,” then sends those jobs to opted-in owners of smartphones, laptops and gaming computers when their devices have spare capacity. That setup would distribute individual pieces of model-training work among participating consumer devices.

Before assigning work, the system would monitor GPU and CPU utilization, memory use, temperature and power consumption. It would use these readings to estimate whether a device can handle a task and manage workloads as the device's available capacity changes. Together, those readings would help the system gauge capacity and manage workloads across devices.

Once a device finishes a task, the system would validate the returned work before using it to update the model. It would then weight updates according to validation results, how stale each update is and the reliability of the device that produced it. These criteria make the quality and timing of the work, as well as the contributing device's record, part of how model updates are combined.

The patent description offers account credits as one possible incentive for an owner who contributes computing time. Microsoft Technology Licensing LLC's patent application says each computation ‘may result in a credit being applied to an owner's account.’ Those credits could be exchanged for free software, access to the model being trained or money. The document does not assign a value to a credit or state when an owner would receive one. Its claims do not set a payment rate, schedule or guarantee of compensation.

The claims address training and device management; they do not define a compensation mechanism. The patent also does not establish a consumer service or give a launch date. It names no eligible device models and gives no details about electricity reimbursement or handling training data on third-party devices.