Clockwork.io raises $31 million for AI cluster fault tolerance

The $31 million round was co-led by Seligman Ventures, Wing Ventures and Premji Invest. Existing investors New Enterprise Associates and e& Capital also participated, bringing Clockwork’s total funding raised to $73 million.[1] Clockwork said it will use the funding to expand fault-tolerance deployments across training, inference and reinforcement learning, including through cloud partners.[2] SiliconANGLE dated its report on the round October 5, 2026.[1] The announcement describes Clockwork as a data-center infrastructure startup focused on improving the efficiency of AI chip clusters.[1]

Clockwork announced TorchSnap, a feature designed to minimize wasted compute.[1] It captures multinode snapshots of distributed AI inference workloads across cluster nodes without requiring developers to change their code.[1] Separately, fast asynchronous application checkpoints are aimed at reinforcement learning.[3] These are distinct details: the evidence describes TorchSnap in connection with distributed inference snapshots, and describes asynchronous checkpoints as aimed at reinforcement learning.[1][3] Clockwork’s broader fault-tolerance software is designed to keep training, reinforcement-learning and inference workloads running through infrastructure issues.[4]

LinkedIn has deployed Clockwork’s LinkPass functionality across its entire GPU fleet.[1] LinkPass reroutes AI traffic around optical-link and switch failures to keep jobs running.[1] Hiremagalur said the software prevents tens of thousands of GPU-hours of downtime each month across LinkedIn’s fleet.[3] WhiteFiber uses Clockwork technology to audit and validate cluster reliability before new AI workloads enter production.[1] These examples show different uses described in the evidence: routing around network-related failures and checking cluster reliability before production workloads are added.[1]

Clockwork’s software layer sits between GPUs and running AI workloads, using nanosecond-accurate telemetry to identify failures before a full cluster restart.[1] Fault tolerance is intended to keep GPUs doing useful work rather than waiting for recovery or repeating completed work.[1] Clockwork said conventional recovery from a saved snapshot can take up to 90 minutes.[1] The evidence also notes that hardware failures can occur daily, and cites Meta’s report of hardware issues every three hours on average during Llama 3’s 54-day training run across 16,384 GPUs.[1]