Clockwork.io Makes GPU Failures a Software Problem

Clockwork.io builds software for a problem that shows up after an AI infrastructure team has already bought the expensive part. The GPUs are installed, the network is fast, and the training job is running. Then one link, accelerator, driver, or node fails, and healthy hardware starts waiting for the recovery plan.

The Palo Alto company turns that failure problem into a software control problem. Its FleetIQ platform combines precise timing, observability, traffic control, and fault tolerance across distributed GPU clusters. TorchPass keeps AI workloads moving through network and accelerator failures, while LinkPass reroutes traffic around broken network paths.

CEO Suresh Vasudevan leads Clockwork.io with a founding and technical team rooted in Stanford systems research. The company matters now because GPU economics are changing the definition of infrastructure performance. Installed capacity looks good in a procurement spreadsheet. Useful work that survives a failure is what determines whether the model ships on time.

Clockwork.io Started With a Better Clock

Clockwork.io began in 2018 as TickTock Networks, built around clock-synchronization research developed at Stanford University. The company renamed itself Clockwork Systems in 2021 and debuted publicly in 2022.

Co-founder Balaji Prabhakar is a Stanford professor whose research includes data-center networking and cloud computing. Co-founder and Chief Engineer Deepak Merugu previously helped build Urban Engines, which Google acquired in 2016. Co-founder and CTO Yilong Geng developed the Huygens clock-synchronization system during his Stanford doctoral work. Chief Scientist Mendel Rosenblum co-founded VMware and spent its first decade as Chief Scientist.

That history matters because distributed systems often fail first as a visibility problem. If thousands of machines do not share a precise view of when an event occurred, operators can see the slowdown without seeing its cause. Clockwork's original timing layer gave the company a way to measure one-way delay and correlate behavior across hosts, NICs, switches, jobs, and networks. The product has since moved from measuring the system to changing what it does next.

FleetIQ Turns Visibility Into Control

The FleetIQ platform combines fleet auditing, fleet monitoring, workload monitoring, dynamic traffic control, and job-aware resilience. Clockwork.io describes the result as a software-driven AI fabric that runs across cloud or on-premises deployments, NVIDIA or AMD accelerators, and Ethernet, RoCE, or InfiniBand networks.

The distinction is operational. Traditional monitoring can tell an engineer that a training run slowed down. Clockwork.io wants to connect the slowdown to the path, component, or workload behavior that caused it, then route traffic or work around the problem.

LinkPass handles network fault tolerance by moving traffic onto another available path when a link or NIC fails. TorchPass extends the same idea to the workload. Its live GPU migration moves the affected portion of a distributed job onto healthy capacity without restarting from the last checkpoint. Clockwork.io also supports multi-node snapshots and asynchronous checkpoints for workloads that need a broader recovery mechanism.

The company is selling less wasted computation, not a prettier dashboard. That makes goodput, the share of GPU time that advances the workload, more important than a simple utilization percentage.

AI Scale Makes Failure a Routine Event

Large AI systems are tightly coordinated. One damaged component can stop progress far beyond its own rack because the remaining workers still have to meet at the same synchronization points. Restarting from a checkpoint adds recovery time and can erase valid computation completed since the checkpoint was written.

Meta's Llama 3 research recorded 419 unexpected interruptions during a 54-day period of training Llama 3 405B on 16,384 H100 GPUs, roughly one interruption every three hours. Meta still achieved more than 90% effective training time through automation. The lesson is not that large clusters are unreliable. It is that failure becomes ordinary, and the recovery system becomes part of model-development velocity.

Clockwork.io has moved beyond laboratory positioning. LinkedIn says it deployed LinkPass across its AI infrastructure fleet and attributes prevention of tens of thousands of GPU-hours of downtime each month to the software. Together AI is bringing TorchPass to market as a service on its GPU clusters. WhiteFiber is expanding Clockwork software across its GPU-as-a-service footprint. Those savings remain customer-attributed claims, but the relationships place the product inside three distinct operating environments.

Funding Expands the Fault-Tolerance Bet

Clockwork.io's October 2026 funding announcement added $31M from co-leads Premji Invest, Wing Venture Capital, and Seligman Ventures, with existing investors NEA and e& Capital participating. Clockwork.io reports $73M in total funding. The company did not disclose a valuation or round label.

The capital is intended to expand fault tolerance across training, inference, and reinforcement learning, deepen enterprise adoption, and scale delivery through cloud partners. That breadth raises the technical burden. Training, inference, and reinforcement learning do not fail in identical ways, and each customer brings a different scheduler, network fabric, accelerator mix, and operating practice.

Clockwork.io therefore has to prove two forms of portability. The software must work across infrastructure combinations, and the economic result must repeat across customers. A reliable migration demo is useful. A repeatable reduction in failed work, delayed jobs, and idle capacity is the commercial product.

Leadership Connects Research to Enterprise Operations

Suresh Vasudevan joined Clockwork.io after serving as CEO of Sysdig and previously leading Nimble Storage through its IPO and acquisition by HPE. His role sits on the other side of the founding team's research depth: turning precise systems work into infrastructure software that enterprises can evaluate, deploy, support, and renew.

The current Clockwork.io job board showed 10 openings on October 5, 2026. The roles span software engineering, network observability, time synchronization, AI infrastructure systems engineering, testing, full-stack technical leadership, human resources, and UK sales.

That mix is an operating signal. Clockwork.io is hiring for the mechanisms underneath the product and the organization needed to carry it into customer environments. The company's careers page describes work on fundamental AI-performance problems and lists benefits, free lunch, a free gym, and TGIFs. Perks describe an office. The open roles describe the work: low-level systems engineering paired with a growing commercial footprint.

Clockwork.io Is Competing Against Lost Work

Clockwork.io is not asking AI infrastructure buyers to believe hardware failures can disappear. It is asking them to stop treating every failure as permission to waste the work around it.

That position becomes more valuable as clusters grow. More accelerators create more opportunities for individual components to fail, while faster and more expensive hardware raises the cost of every idle minute. The winning infrastructure layer will not be the one with the cleanest uptime story. It will be the one that keeps the most useful work moving when the uptime story breaks.

Clockwork.io began by giving distributed machines a more accurate sense of time. Its larger opportunity is giving AI operators a better answer to what happens next.