Clockwork.io raises $31 million
Clockwork.io raises $31 million; LinkedIn, Together AI, and WhiteFiber adopt its fault-tolerant software to avoid wasting GPU hours
LinkedIn saves tens of thousands of GPU downtime hours per month; TorchPass’s new innovation preserves AI workload progress without code changes and accelerates reinforcement learning.
Palo Alto, California, October 5, 2026 /PRNewswire/ — Clockwork.io, a fault-tolerant software provider enabling AI training, reinforcement learning, and inference workloads to continue running during infrastructure failures, today announced a new $31 million funding round, along with production deployments at LinkedIn and Together AI, and further adoption by WhiteFiber.
The company has also launched two new features for TorchPass, each capable of capturing the state of running distributed AI tasks. The multimodal platform snapshot is an industry first in training, enabling the saving of complete tasks across all nodes and subsequent restoration without modifying training code. Fast asynchronous application checkpoints execute in the background during task execution, accelerating reinforcement learning by providing updated model weights to inference replicas responsible for generating rollouts—the samples upon which the model learns—thereby reducing wait times for these replicas or preventing them from working with outdated models.
Fault tolerance has become a fundamental requirement for large-scale AI.
Large-scale distributed AI workloads may span thousands of GPUs, which must remain synchronized: a single GPU failure, link disruption, or server outage can halt the entire task. Meta reported that during a 54-day training of Llama 3 using 16,384 GPUs, unexpected interruptions occurred on average every three hours.
The standard approach is to reload the checkpoint, which is a saved copy of the task’s progress. The recovery process can take up to 90 minutes, leaving otherwise operational GPUs idle and forcing the system to repeat work already completed. Customers pay for both the idle GPUs and the redundant computations, and model completion times are extended. As task sizes grow, each restart puts more GPU usage time at risk.
At this scale, ensuring continuous operation even in the event of failures has become an infrastructure requirement. Clockwork.io meets this need through a suite of fault-tolerant tools, deployed by the platform team as an intermediary layer between hardware and workloads. LinkPass reroutes traffic to bypass faulty links, allowing tasks to remain unaware of the failure. TorchPass migrates workloads from faulty GPUs to healthy ones, enabling training to resume without reverting to earlier states. Both solutions are now in production. Today’s announcement of TorchPass’s new platform snapshot feature enables capturing the state of running distributed tasks; this capability can be used to restore entire tasks when failures are severe enough to exceed simple migration.
“At AI scale, failures are inevitable; but losing hours of productive work as a result should not happen,” said Suresh Vasudevan, CEO of Clockwork.io. “Fault tolerance acts like a performance multiplier: it allows GPUs to continue doing productive work rather than waiting for recovery processes or repeating tasks already completed. We developed this software in collaboration with enterprises and cloud providers operating some of the largest GPU clusters, ensuring it handles the actual failures they encounter. This protection should be a foundational part of the infrastructure these companies and providers rely on daily.”
Scope expanded to include enterprises, hyperscale clouds, and neoclouds.
Enterprises building their own GPU clusters, hyperscale cloud providers, and "neocloud" providers all adopt Clockwork.io for the same reason: to put more GPU hours to productive use.
LinkedIn has deployed LinkPass network fault tolerance across its entire AI infrastructure, preventing tens of thousands of GPU downtime hours each month.
Raghu Hiremagalur, Senior Vice President and Chief Technology Officer of LinkedIn’s Infrastructure, said: “At the scale of AI infrastructure, a single network issue should never take a functioning GPU offline or interrupt running workloads. Before adopting Clockwork.io, jitter in InfiniBand NICs could take down a server with eight GPUs, and switch port jitter could disable a second server, doubling the impact to 16 GPUs. Clockwork.io has transformed this operational model. Its network fault-tolerance technology automatically reroutes traffic to available paths, allowing tasks to continue running during repairs to links, optical components, cables, or network cards. Overall, Clockwork.io prevents tens of thousands of GPU downtime hours across our entire cluster each month. It turns disruptive operational events into manageable maintenance events, helping improve infrastructure utilization and operational efficiency.”
Together AI is launching TorchPass as a service on its GPU clusters. The two parties will demonstrate a multi-node training task live at the PyTorch Conference, showing that the task can continue running without restarts even when network and GPU failures are intentionally introduced.
“Customers measure us by goodput—the proportion of GPU hours that actually advance the model,” said Pavneet Ahluwalia, Head of Product at Together AI. “Our node recovery feature automatically detects failures and allocates replacement capacity. Clockwork.io’s TorchPass and LinkPass build on this infrastructure to ensure tasks continue executing despite GPU or link failures, preserving progress already made. We are now bringing them to market as the next layer of fault tolerance on our platform.”
WhiteFiber (NASDAQ: WYFI), an existing customer of the company, is expanding its use of Clockwork.io software within its growing global GPU-as-a-Service infrastructure.
WhiteFiber’s Chief Technology Officer, Tom Sanfilippo, said: “It’s critical to stress-test the reliability of a cluster before putting it into production, because if customers receive a system with hidden interconnect failures, they ultimately pay the price in failed tasks and lost GPU hours. Subpar optical components, misconfigured network interface cards (NICs), and links that pass basic tests but degrade under load can all be overlooked. Clockwork.io’s automated cluster audits validate every link and every node simultaneously, pinpointing failures within minutes and allowing us to fix them before delivery. We can launch clusters faster, ensuring that our customers’ first training runs occur on a network that has undergone end-to-end validation—not just on a system that happens to be powered on. Given the rapid market demand, delivering verified capacity to customers quickly is essential to our business; therefore, we are deploying Clockwork.io across all our clusters.”
TorchPass's new feature delegates workload protection to platform teams.
Clockwork.io has extended TorchPass capabilities beyond GPU migration to include two features previously unavailable to the platform team: the ability to generate snapshots of entire distributed tasks and application checkpoints that are fast enough to run in the background.
In training scenarios, platform snapshots capture the execution state of running tasks across all nodes, enabling task recovery after interruptions. Platform teams and AI infrastructure engineers can deploy this solution to compatible workloads without requiring application owners to modify code or add checkpointing logic. Enterprise teams can protect training tasks across their entire cluster through a single mechanism, and cloud providers can safeguard customer tasks even when they don’t control the underlying code. When teams implement checkpointing at the application layer, TorchPass’s fast checkpointing capability enables more frequent checkpoints, reducing lost progress after failures and minimizing the need for redundant computation.
In terms of inference, large models typically run on two or more servers, so link failures can cause the entire replica to fail and interrupt ongoing user sessions or agent tasks. LinkPass enables these multi-server replicas to continue operating even during link failures.
Reinforcement learning relies on both training and inference. Multiple model replicas generate rollouts, and the training system learns from them; updated weights must first be returned to these replicas before they can generate results using the latest version. TorchPass's application checkpoints enable faster transmission of updated weights back to the execution replicas, while LinkPass ensures these replicas remain operational during link failures.
Dylan Patel, founder, CEO, and chief analyst at SemiAnalysis, said: “Cluster fault tolerance was previously only a training issue; now it’s also an inference issue. Based on our testing of ClusterMAX, TorchPass reduced effective performance loss during training from 14% to less than 3% at a neocloud provider with a ‘Gold’ rating. Reinforcement learning (RL) connects these two processes: inference replicas generate simulations, which the training system learns from, and the updated weights are then sent back to the replicas. Clockwork.io enables replicas to continue operating despite link fluctuations and network failures. Its ultra-fast checkpoints also accelerate the return of weights to the inference cluster, preventing stagnation in either direction. This solution must be deployed as a universal fault-tolerance layer for training, inference, and reinforcement learning.”
Enterprise platform teams and cloud service providers can contact Clockwork.io to evaluate how their software can be applied to their workloads or to explore collaboration opportunities.
Funding Round and Use of Funds
This round was co-led by Premji Invest, Wing Venture Capital, and Seligman Ventures, with participation from existing investors NEA and e& Capital. With this funding, Clockwork.io has now raised a total of $73 million. The company will use the funds to accelerate the deployment of its fault-tolerant tool suite—covering training, inference, and reinforcement learning phases—expand enterprise adoption, and scale delivery through cloud partners.
“The only thing that scales perfectly is unreliability: the more GPUs you pack onto a single machine, the more things will break,” said Greg Papadopoulos, partner at NEA. “Traditional approaches—stopping tasks and reloading checkpoints—are meaningless at today’s scale. Clockwork treats failure as normal: TorchPass can migrate training in real time away from faulty GPUs, and now it can capture the state of an entire running task without requiring any code changes. It’s already saving tens of thousands of GPU hours per month. We invested in Clockwork.io back in 2021 and are thrilled to continue supporting them as they define the performance layer for AI clusters.”
About Clockwork.io
Clockwork.io is the pioneer of Software-Driven AI Fabrics™, a programmable layer between hardware and workloads that enables GPU clusters to achieve observability, fault tolerance, and maximum utilization—regardless of the accelerators, networks, or cloud environments used. AI workloads require the entire cluster to operate as a single machine; however, failures and bottlenecks often leave GPUs idle. Clockwork’s FleetLens platform recovers this lost capacity: with nanosecond-level telemetry, it precisely identifies which GPU, node, or link is slowing down or blocking tasks, and validates clusters before deployment. Additionally, through LinkPass and TorchPass, it ensures continuous operation of workloads—from training to inference—even during infrastructure failures or degradation. Independent benchmarking by SemiAnalysis shows that TorchPass outperforms major open-source frameworks and traditional checkpoint-and-restart methods in fault recovery speed. Companies such as LinkedIn, Together AI, WhiteFiber, Wells Fargo, Nebius, NScale, and DCAI use Clockwork.io to power their AI infrastructure. Learn more at www.clockwork.io.