Clockwork.io has raised $31 million as LinkedIn, Together AI and WhiteFiber adopt or expand its software for keeping AI workloads running through infrastructure failures. The funding arrives alongside new recovery features that could help GPU operators cut downtime and repeated computation, though actual savings will depend heavily on workload compatibility and the infrastructure sitting underneath it all.
For cloud buyers, what matters most is completed work, not raw availability. A rented GPU can sit allocated and billing away while contributing nothing to a training run that stalled elsewhere in the cluster. Recovery takes time, and rolling back discards whatever progress existed before the failure. Software that trims those losses could effectively lower the real price of computing, even without touching the advertised hourly rate.
LinkedIn reports that Clockwork.io prevents tens of thousands of GPU hours of downtime each month, a meaningful claim on its own, though the announcement doesn’t share the fleet size needed to calculate a percentage improvement. Buyers should resist converting that number directly into their own savings estimate, since cluster size, failure frequency and existing recovery procedures will all shift the actual result.
Together AI, meanwhile, is bringing TorchPass to its GPU Clusters as a service, which could shift some recovery responsibility away from application developers and onto the infrastructure provider itself. That raises a contract question worth asking directly: does the service protect in-progress work, guarantee a recovery time, or simply offer another recovery option? Those commitments aren’t interchangeable, and conflating them could lead to disappointment later.
On the technical side, Clockwork.io’s LinkPass reroutes traffic around network faults, while TorchPass can migrate work away from a failing GPU. New features now capture distributed job state and support asynchronous checkpoints, aiming to protect workloads without requiring changes to training code. Still, recovery software needs somewhere to actually recover to, so buyers should ask about spare capacity requirements and what happens when several components fail at once.
WhiteFiber takes the idea further, using automated audits to catch infrastructure faults before accepting clusters in the first place. That approach could reduce support incidents down the line, though an audit only means something if it mirrors the actual workload a customer intends to run.
Premji Invest, Wing Venture Capital, and Seligman Ventures co-led the funding. For now, the company has not specified pricing or deployment overhead, so buyers evaluating the technology will need to measure their own incident history before estimating real savings.
