Clockwork Systems raises $31 million and launches TorchSnap for AI cluster recovery
The AMW Read
The funding and TorchSnap launch incrementally expand an infrastructure player, while reported customer deployments make cluster fault tolerance relevant across AI infrastructure buyers.
Clockwork Systems raises $31 million and launches TorchSnap for AI cluster recovery
Clockwork Systems raised $31 million in a round co-led by Seligman Ventures, Wing Ventures and Premji Invest, with participation from existing investors New Enterprise Associates and e& Capital. Total funding now stands at $73 million. The company also introduced TorchSnap, which captures multi-node snapshots of distributed AI inference workloads without requiring developers to modify their code. It adds recovery capability alongside LinkPass, its network failover tool, and TorchPass, its GPU migration software.
The investment targets an operational constraint in AI infrastructure: installed GPU capacity loses value when failures leave healthy chips idle or force workloads to repeat completed computation. The article cites Meta's 54-day Llama 3 training run on 16,384 GPUs, which experienced hardware problems roughly every three hours; separately, Clockwork says recovery can take up to 90 minutes. TorchSnap extends the company's resilience offering into inference, where interruptions also affect service continuity. This positions Clockwork in the software layer that improves useful output from existing clusters, with recovery behavior becoming part of infrastructure purchasing decisions.
For builders and investors, the concrete diligence question is how much useful GPU time these tools recover under production conditions. The source reports that LinkedIn deployed LinkPass across all its GPU clusters, eliminating thousands of GPU-hours of downtime monthly, while Together AI offers TorchPass as a service and WhiteFiber uses Clockwork technology to assess cluster reliability before production deployment. Those deployments provide adoption evidence, but the article supplies no quantified TorchSnap results. Buyers should measure recovery time, repeated computation and deployment overhead before extrapolating the benefits of its existing tools to the new feature.