Clockwork.io Raises $31M for AI Fault Tolerance

AI clusters are sold by GPU count. Operators live with a different number: how many paid GPU-hours still move the model forward after the next link, NIC, server or accelerator fails.

Clockwork.io has raised $31M to push that number higher. The Palo Alto company builds software that keeps distributed AI training, inference and reinforcement-learning workloads running through infrastructure failures. Premji Invest, Wing Venture Capital and Seligman Ventures co-led the financing, while existing investors NEA and e& Capital participated. Clockwork reports $73M in total funding.

The financing arrives with production deployments at LinkedIn and Together AI, expanded adoption by WhiteFiber, and new capabilities for Clockwork's TorchPass fault-tolerance product. Together, those details make the round less about adding another layer to the AI stack and more about protecting the expensive work already moving through it.

What Clockwork.io Announced

Clockwork.io announced the financing on October 5, 2026. The company did not disclose a round label, valuation, investor allocations or ownership terms, so the transaction is best described as new venture funding rather than forcing it into an unsupported series.

Premji Invest, Wing Venture Capital and Seligman Ventures joined as co-leads. NEA and e& Capital returned as existing backers. Clockwork said it will use the capital to accelerate its fault-tolerance suite across training, inference and reinforcement learning, expand enterprise adoption and scale delivery through cloud partners.

Clockwork previously announced a $21M Series A led by NEA in March 2022. The current $73M total indicates additional financing before this round, but the company has not provided a complete primary-source breakdown of every prior transaction. That gap belongs in the record because funding totals can be accurate without making every intermediate round public.

Why GPU Fault Tolerance Has Become an Economic Problem

Large AI workloads coordinate thousands of GPUs as one distributed system. A failed accelerator, cable, NIC, link or server can stop progress beyond the damaged component because the remaining devices still have to reach the same synchronization points. The bill continues while healthy hardware waits or repeats computation lost since the last checkpoint.

The scale of the problem is visible in Meta's Llama 3 research. During a 54-day snapshot of Llama 3 405B pre-training on 16,384 H100 GPUs, Meta recorded 419 unexpected interruptions, roughly 1 every 3 hours. Meta still reported more than 90% effective training time through automation, which shows both sides of the operating reality: failures become routine at scale, and the quality of the recovery system determines how much useful work survives them.

That is where Clockwork wants to sit. The company is selling fault tolerance as part of compute economics, not a maintenance tool used after an outage. The relevant metric is goodput, the share of GPU time that advances the workload instead of waiting, recovering or repeating completed work.

How Clockwork's Software Protects Work in Progress

Clockwork's platform grew from software-based clock synchronization research at Stanford. Fine-grained timing gives the system visibility into the network and workload behavior across distributed infrastructure. The company has expanded that foundation into a software layer that observes the cluster, routes around faults and moves workload state when hardware begins to fail.

LinkPass reroutes traffic around a failed network path so a link problem does not have to interrupt the running job. TorchPass moves work from a failing GPU to a healthy one rather than rolling the job back to an earlier checkpoint. Clockwork's new multi-node platform snapshots capture the state of a distributed job across every node, giving operators a recovery path when a failure is too large to migrate around.

The company also introduced asynchronous application checkpoints for reinforcement learning. Those checkpoints are designed to capture updated model weights while the job continues, then move them to the inference replicas generating new rollouts. The product claim is straightforward: reduce the time replicas spend waiting or learning from stale weights without requiring training-code changes.

Production Adoption Moves the Story Beyond a Demo

LinkedIn says it has deployed LinkPass across its AI infrastructure fleet. Raghu Hiremagalur, LinkedIn's SVP and CTO of Infrastructure, said the software automatically reroutes traffic during link, optic, cable and NIC failures and prevents tens of thousands of GPU-hours of downtime each month. That figure is an attributed customer statement published in Clockwork's release, not an independently audited metric, but it connects the technology to an operating fleet.

Together AI is bringing TorchPass to market as a service on its GPU clusters and plans to demonstrate a multi-node training job continuing through injected network and GPU failures. WhiteFiber is expanding Clockwork software across its GPU-as-a-service footprint. The three relationships cover an enterprise operator, an AI cloud platform and a GPU infrastructure provider, which gives Clockwork several paths into the same underlying customer problem.

AI infrastructure funding has recently flowed toward companies assembling more capacity, including PaleBlueDot AI's $200M Series C, and toward operators giving customers tighter control over deployment, such as Go.AI's $85M Series A. Clockwork occupies a different layer of that market. It is trying to make existing capacity more productive when distributed workloads encounter the failures that large systems cannot eliminate.

The Team Behind the Fault-Tolerance Layer

CEO Suresh Vasudevan joined Clockwork after leading Sysdig and previously serving as CEO of Nimble Storage. The company was founded by Stanford professor Balaji Prabhakar, Deepak Merugu and Yilong Geng, who now serves as CTO. Mendel Rosenblum, a VMware co-founder and Stanford professor, is Clockwork's Chief Scientist.

The leadership mix matters because Clockwork is turning academic systems work into software that has to survive enterprise buying, cloud-partner delivery and production support. Precise timing may reveal the fault, but the commercial product has to integrate with the hardware, networks, orchestration and operational practices already inside each customer's cluster.

What the $31M Has to Change

The financing gives Clockwork more room to distribute its software across training, inference and reinforcement-learning environments. It also raises the standard of proof. The company now has to show that its customer-reported savings can repeat across different accelerators, network fabrics, cloud providers and workload patterns without adding enough complexity to erase the operational gain.

That assignment will keep moving as AI infrastructure changes. Faster chips increase the cost of idle time, larger clusters create more opportunities for individual components to fail, and reinforcement learning connects training and inference in tighter loops. Clockwork has raised $31M to keep the useful work inside those systems moving, while every new deployment adds another failure pattern its software has to recognize before the rest of the cluster becomes an audience.