Clockwork.io Raises $31 Million In Funding

Clockwork.io closed a $31 million funding round, co-led by Premji Invest, Wing Venture Capital, and Seligman Ventures, to scale its fault tolerance software that prevents GPU downtime in large AI training, inference, and reinforcement learning clusters.

Clockwork.io raised $31 million in a late stage funding round, bringing its total capital raised to $73 million. The round was co-led by Premji Invest, Wing Venture Capital, and Seligman Ventures, with participation from existing investors NEA and e& Capital. The capital will accelerate rollout of its fault tolerance suite across AI training, inference, and reinforcement learning; expand enterprise adoption; and scale delivery through cloud and neocloud partners.

What is Clockwork.io?

Clockwork.io, based in Palo Alto, California, originated from Stanford University research. Founded in 2018 as TickTock Networks by Yilong Geng, Deepak Merugu, and Professor Balaji Prabhakar (with Mendel Rosenblum, VMware co-founder, as chief scientist), it initially focused on high precision software based clock synchronization for distributed systems. It rebranded and publicly debuted around 2021–2022, raising a $21 million Series A led by NEA (with angels including John Hennessy, Ram Shriram, and Jerry Yang). A further ~$20.5 million round in September 2025, led by NEA with participation from Lip-Bu Tan, John Chambers, Carl Ledbetter, and e& Capital, coincided with the appointment of Suresh Vasudevan as CEO and the launch of FleetIQ.

Vasudevan, previously CEO of Sysdig (scaled from ~$5M to over $100M ARR) and Nimble Storage (from startup to IPO and HPE acquisition), plus roles at NetApp and Omneon, shifted the company toward AI infrastructure. The core technology evolved from nanosecond accurate clock sync and latency sensing into a “Software Driven AI Fabric”, a programmable layer between hardware and workloads that delivers observability, fault tolerance, and higher GPU utilization across any accelerator, network (InfiniBand, RoCE, Ethernet), or cloud.

Clockwork company leadership team members Anita Pandey, Joe Tarantino, and Gavin Cohen.

Large scale AI workloads (training, inference, reinforcement learning) run across thousands of synchronized GPUs. A single GPU failure, link flap, switch issue, or node problem can stall or crash the entire job. Meta reported unexpected interruptions roughly once every three hours over 54 days while training Llama 3 on 16,384 GPUs. Traditional recovery, reloading a checkpoint, can take up to 90 minutes, leaves healthy GPUs idle, and forces recomputation of work since the last save. Mean time to failure on a 1,024-GPU cluster is cited around 7.9 hours, with daily GPU time losses of 2.3–4.5 hours and detection/repair times of 1–3 hours. As clusters grow, the cost of idle capacity and repeated computation escalates sharply; organizations increasingly prioritize maximizing existing GPU fleets over simply acquiring more hardware.

Clockwork positions a software only layer that treats failure as the normal state:

  • LinkPass: Intercepts link/NIC/switch failures and rebalances traffic across healthy paths (delivered as an NCCL plugin; installs in under 30 minutes with zero model changes). Keeps multi node training and multi server inference replicas running.
  • TorchPass Migration: Live migrates work from a failing GPU or node to a healthy replacement over RDMA and resumes from the exact iteration with no lost compute.
  • TorchPass Snapshots (new): Industry first multi node platform snapshots that capture an entire distributed job’s state across all nodes in under 20 seconds with no training code changes. Enables recovery without application level modifications. Also supports fast asynchronous application checkpoints that run in the background, accelerating reinforcement learning by delivering updated model weights to inference replicas faster.
  • FleetLens / FleetAudit: Observability and pre production validation. Uses sub microsecond clock sync, out of band probing, path decomposition, and triangulation to detect gray failures, congestion, and topology issues that standard tools miss; pinpoints exact links, switches, or NICs; and validates clusters before workloads launch.

SemiAnalysis independently benchmarked TorchPass as the only fault tolerance framework that does not cost training performance, cutting goodput loss (e.g., from 14% to under 3% on a gold rated neocloud in ClusterMAX, or from 28.5% to 10.3% on a 4,096-GPU silver tier cluster) and recovering from failures faster than checkpoint restart approaches. It also offers contractual guarantees (YOCO, You Only Compute Once) against GPU waste.

The funding announcement highlighted production deployments and expanded use:

  • LinkedIn has deployed LinkPass across its entire AI GPU fleet, preventing tens of thousands of GPU hours of downtime monthly. Previously, a single InfiniBand NIC flap could remove an 8-GPU server; a switch port flap could double the impact. Clockwork turns these into non disruptive maintenance events.
  • Together AI is offering TorchPass as a service on its GPU Clusters and will demonstrate live multi node training continuing through injected failures at the PyTorch Conference.
  • WhiteFiber (NASDAQ: WYFI) is expanding Clockwork across its global GPU as a service footprint for automated fleet audits that validate links and nodes before production, accelerating cluster bring-up.

Additional customers and partners include Wells Fargo, Uber, Nebius, NScale, and DCAI (Denmark’s sovereign AI effort). Endorsements from SemiAnalysis founder Dylan Patel and others underscore the shift of fault tolerance from a training only concern to one spanning inference and RL, where weight transfer and replica availability are critical.

Clockwork AI software platform tagline: "AI that never stalls. GPUs that never sit idle."

Recommended: Mojave Energy Systems Raises $19M In Series B Funding Round

Investor Context

  • Premji Invest (Azim Premji’s family office) has been active in U.S. AI bets; this fits its platform of high-conviction infrastructure and software investments.
  • Wing Venture Capital brings enterprise software and infrastructure expertise.
  • Seligman Ventures (launched February 2026, recently doubled deployable capital to $1 billion) focuses heavily on AI infrastructure bottlenecks (compute, networking, connectivity, power, thermal, cybersecurity, and software layers) and has already backed more than a dozen companies in the space.
  • Returning investors NEA (early backer since ~2021, with Greg Papadopoulos emphasizing that “the one thing that scales perfectly is unreliability”) and e& Capital provide continuity.

The round arrives as AI infrastructure economics shift from pure capacity acquisition to utilization and “goodput” (the share of GPU hours that actually advance the model). Clockwork’s software layer sits between hardware and frameworks (compatible with TorchTitan, Megatron-LM, DeepSpeed, etc.), enabling platform teams and neoclouds to protect workloads without requiring model owners to change code. This creates a defensible position as a performance and resilience fabric rather than a point tool.

Use of funds prioritizes product expansion across the full AI lifecycle, enterprise sales, and partner channel scaling, logical given the proof points at LinkedIn scale fleets and the rising cost of downtime. In a market where GPU clusters are among the most expensive computing infrastructures ever built, software that converts failure events into non events and recovers lost capacity represents a direct multiplier on capital already deployed. The combination of technical differentiation (live migration + zero code multi node snapshots), independent validation, and blue chip customer traction positions Clockwork as a defining player in the emerging software defined AI fabric category.

Please email us your feedback and news tips at hello(at)dailycompanynews.com

  • Reading time:7 mins read
  • Post category:News / Popular