Clockwork resumes live AI training job after GPU server is unplugged

At the 70th IT Press Tour, startup Clockwork demonstrated fault-tolerant AI training: an engineer cut power to a GPU server at step 5,040, and the job resumed from the same step on a different machine in about a minute, with no checkpoint reload. Its LinkPass and TorchPass features are part of the FleetIQ platform, generally available since September 2025; the company announced a $31 million round a day earlier.

• The job stopped at step 5,040 and resumed from the same step on another server
• LinkPass shifts traffic to healthy NICs in one to two seconds
• Meta puts mean time to failure at 7.9 hours for 1,024-GPU jobs
• 42.5% of ByteDance jobs ran at least 10% slower due to stragglers