The Pulse
Clockwork.io Raises $31M for AI Workload Recovery
Clockwork.io raised $31 million as LinkedIn, Together AI and WhiteFiber expand deployments of its software for AI infrastructure. New TorchPass features snapshot distributed training jobs and checkpoint model updates for reinforcement learn
AI.info Team · Written with AI ·
Clockwork.io announced a $31 million funding round on October 5, 2026, alongside new deployments at LinkedIn and Together AI and expanded use by WhiteFiber. The company is pitching its software at a costly failure mode in large AI clusters: when one GPU or network link breaks, coordinated workloads can stall while teams repair infrastructure or restart jobs.
Checkpoints can leave GPUs waiting
Clockwork says recovery from a conventional checkpoint can take up to 90 minutes, during which healthy GPUs may sit idle and the job may have to repeat work completed since its last saved state. The company points to Meta’s Llama 3 training as an example of the operating pressure: Meta reported unexpected interruptions averaging about one every three hours during a 54-day run on 16,384 GPUs.
Clockwork’s existing LinkPass software reroutes traffic around a failed network link; TorchPass moves work away from a failing GPU. The company says both products are in production. The new funding announcement adds two TorchPass capabilities intended to protect more of a job’s progress and support reinforcement-learning systems.
TorchPass adds job snapshots and background checkpoints
Multi-node platform snapshots capture a distributed job’s execution state across its nodes, allowing platform teams to restore supported training workloads after an interruption without changing the training code or adding application-level checkpoint logic. Clockwork says this can also help cloud providers protect customer jobs whose code they do not control.
A second feature takes fast application checkpoints asynchronously, while the job continues running. In reinforcement learning, those checkpoints carry updated model weights to inference replicas that generate training examples, or rollouts. Faster updates can reduce how long replicas wait or rely on stale weights, while LinkPass is designed to keep multi-server inference replicas serving through link failures.
Clockwork CEO Suresh Vasudevan described the goal as preserving useful computation rather than simply restoring machines. “Failures are inevitable at AI scale. Losing hours of useful work to them should not be,” he said. The announcement did not provide product pricing or a general availability date for the new capabilities.
LinkedIn reports tens of thousands of GPU-hours saved
LinkedIn has deployed LinkPass across its AI infrastructure fleet. Clockwork says the software prevents tens of thousands of GPU-hours of downtime each month; LinkedIn’s infrastructure chief Raghu Hiremagalur said a network fault previously could take an eight-GPU server out of service, while a switch-port fault could affect a second server as well.
“Its network fault-tolerance technology automatically reroutes traffic onto healthy paths, allowing jobs to continue uninterrupted while link, optic, cable, or NIC faults are repaired,” Hiremagalur said. The deployment described by LinkedIn is LinkPass network protection, not a claim that LinkedIn has adopted the new TorchPass snapshot features.
Together AI plans to offer TorchPass as a service on its GPU clusters and said it would demonstrate a multi-node training job continuing through injected GPU and network failures without restarting. WhiteFiber, an existing customer, is expanding Clockwork’s software across its GPU-as-a-service footprint. Its use includes automated checks of links and nodes before clusters enter production, according to the announcement.
Investors back a wider fault-tolerance push
Premji Invest, Wing Venture Capital and Seligman Ventures co-led the round, with existing investors NEA and e& Capital also participating. Clockwork says the financing brings its total funding to $73 million. It plans to use the capital to expand its software across training, inference and reinforcement learning, grow enterprise adoption and extend distribution through cloud partners.
SemiAnalysis founder and chief analyst Dylan Patel cited a ClusterMAX result in the announcement: TorchPass cut training goodput loss from 14% to under 3% for a gold-rated neocloud. Goodput refers to the portion of GPU time that advances a workload. That figure describes the cited ClusterMAX result, not a reported outcome across every Clockwork customer or deployment.
The announcement’s commercial picture is specific: LinkedIn reports monthly downtime avoided with LinkPass, Together AI plans a TorchPass service, and WhiteFiber is expanding its existing use. Clockwork says enterprise platform teams and cloud providers can contact the company to evaluate the software for their workloads.