Clockwork.io
by Clockwork.io
GPU fault-tolerance and fabric observability software that keeps AI training jobs running through failures
Clockwork.io sells software that keeps distributed AI jobs running when GPUs or network links fail, and shows where a cluster is slow or misconfigured. It is for neoclouds and enterprise AI infrastructure teams losing expensive GPU hours to restarts, and it needs no new hardware.
Clockwork.io is a Palo Alto company founded in 2018 by Balaji Prabhakar, Deepak Merugu and Yilong Geng, growing out of Stanford research on clock synchronization and traffic control. It sells a software layer between AI workloads and GPU hardware. LinkPass, delivered as an NCCL plugin, reroutes traffic across a node's other NICs when a link fails and fails back on recovery, with no model changes and an install the vendor puts at under 30 minutes. TorchPass migrates a failed worker's state to a spare GPU over RDMA and resumes from the same iteration, falling back to checkpoints if migration fails. New in October 2026 are multi-node platform snapshots, which the vendor says take under 20 seconds without code changes, and asynchronous checkpoints for moving weights to inference replicas. FleetLens and FleetAudit cover telemetry, fault localization and pre-launch validation. In a March 2026 vendor benchmark on 64 NVIDIA H200 GPUs with injected failures, TorchPass finished 3,000 steps in 405 minutes, against 818 for checkpoint restart and 930 for Meta's TorchFT. Clockwork announced a $31 million round on October 5, 2026, bringing total funding to $73 million. It says LinkedIn runs LinkPass across its AI fleet, Together AI will offer TorchPass as a service on its GPU Clusters, and WhiteFiber is expanding its deployment. Its customer list also names Wells Fargo, Nebius, Nscale and the Danish Centre for AI Innovation.
The head of AI infrastructure at a neocloud or large enterprise running distributed training or inference on hundreds of GPUs or more, where each failure costs measurable compute time.
Fewer lost GPU hours, because jobs route around failed links and migrate off failed GPUs instead of restarting from the last checkpoint.
At a Glance
- Category
- Infrastructure & Cloud
- Pricing
- Contact for pricing
- Target Market
- CTOs, Heads of AI Infrastructure, Neocloud Operators, ML Platform Engineers
- Deployment
- Self-hosted, Multi-cloud
- Founded
- 2018
- Headquarters
- Palo Alto, United States
Key Features
- ✓LinkPass
An NCCL plugin that reroutes traffic across a node's other NICs when a link fails, so a flapping link does not take a server out of the job.
- ✓TorchPass Migration
Moves a failed worker's state to a spare GPU over RDMA and resumes at the same training iteration, avoiding recomputation since the last checkpoint.
- ✓TorchPass platform snapshots
Captures a running distributed job across all nodes without code changes, which the vendor says takes under 20 seconds, for failures too large to route around.
- ✓Asynchronous application checkpoints
Takes checkpoints in the background to speed reinforcement learning weight updates to inference replicas, and needs a lightweight code integration to set up.
- ✓FleetLens observability
Detects slow or failing jobs, decides whether the fabric is at fault, and pinpoints the link, switch or NIC, using sub-microsecond clock synchronization.
- ✓FleetAudit
Checks links, NICs and GPUs before a job starts, so operators find bad components in validation instead of during an expensive training run.
Use Cases
- •Keeping large training runs alive
A model team running multi-node training migrates ranks off failed GPUs and resumes at the same step, which cut wall-clock time by about half against checkpoint restart in the vendor's benchmark.
- •Surviving network link flaps
LinkedIn runs LinkPass across its AI fleet and says automatic rerouting prevents tens of thousands of GPU hours of downtime monthly, a figure the company reports itself.
- •Reselling fault tolerance as a neocloud feature
Together AI plans to offer TorchPass as a service on its GPU Clusters, giving tenants recovery from hardware faults as part of the rental.
- •Validating a new cluster before launch
WhiteFiber's CTO says the automated fleet audit validates every link and node at once and localizes faults in minutes during bring-up.
Ideal For
Best For
- ✓Neoclouds competing on training goodput and wanting a fault-tolerance feature to resell
- ✓Enterprise AI platform teams running multi-node PyTorch training on large GPU clusters
- ✓Operators who lose jobs to InfiniBand or Ethernet link flaps and want automatic rerouting
- ✓Infrastructure teams that need to localize faults to a specific link, switch or NIC before launching a cluster
Not Ideal For
- ✗Teams on a handful of GPUs, since the vendor's failure figures come from clusters of 1,024 GPUs and up, and a smaller fleet has far fewer failures to recover from
- ✗Buyers who need independently audited performance data before purchase, since the published benchmark is vendor-run on a single model and cluster configuration
- ✗Shops training outside the PyTorch ecosystem, since the published test and model-layer patches cover TorchTitan, Megatron-LM and DeepSpeed, so other frameworks need confirming with the vendor
Deployment
Market Analysis
Pros
- ✓LinkPass installs as an NCCL plugin with no model changes, which keeps adoption cost low if the vendor's claims hold in your environment
- ✓Named production users include LinkedIn, with Together AI bringing TorchPass to its GPU Clusters and WhiteFiber expanding its deployment
- ✓The March 2026 benchmark published its method and setup, and showed zero lost steps and about half the wall-clock time of checkpoint restart under injected failures
- ✓Runs as software across GPU and network vendors, according to 2025 coverage of FleetIQ
Cons
- ✗Performance evidence is mostly vendor-sourced: the benchmark covers one model and one cluster configuration, and LetsDataScience notes the claimed gains have not been independently audited
- ✗The headline 14% to under 3% goodput figure comes through a SemiAnalysis quote in Clockwork's release, and the underlying SemiAnalysis data could not be located
- ✗TorchPass migration needs spare GPUs to move onto and reverts to checkpoint recovery if migration fails
- ✗No pricing, security certifications or social links appear on the company's homepage, so procurement needs direct vendor answers
- ✗TorchPass launched only in March 2026 (version 0.2.1 in that month's benchmark), so production history is short
Pricing
Custom
Contact for pricing
- ✓LinkPass network fault tolerance
- ✓TorchPass migration and snapshots
- ✓FleetLens and FleetAudit observability
- ✓Free consultation offered on the website
Clockwork.io publishes no list prices or metering model (checked November 2, 2026); its website offers only a free consultation, so cost has to come from a sales conversation. Tenants of Together AI GPU Clusters are due to get TorchPass as a service there, and that price is also unpublished in the sources reviewed.
Security & Compliance
Sources
This page was written from 10 sources, 8 on domains other than clockwork.io.
- 1.clockwork.io — clockwork.iovendor
- 2.clockwork.io — keeping distributed training running through failuresvendor
- 3.prnewswire.com — clockworkio raises 31m as linkedin together ai and whitefibe
- 4.sdxcentral.com — clockworkio secures 31m to keep ai workloads resilient again
- 5.hpcwire.com — clockwork io announces 31m in funding and deployments at lin
- 6.techstartups.com — stanford spinout clockwork raises 20 6m in funding launches
- 7.letsdatascience.com — clockworkio raises 31m for ai workload resilience software 3
- 8.nvidia.com — mission control
- 9.docs.pytorch.org — run
- 10.hn.algolia.com — search
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Runpod
Per-second GPU cloud for training, fine-tuning and serverless inference, from RTX 4090s to B300s
LiveKit
Open-source real-time voice, video and AI agent infrastructure behind ChatGPT's voice mode, with a managed LiveKit Cloud
Akamai Inference
Managed open-weight model inference on Akamai Cloud with an OpenAI-compatible API, serverless or dedicated GPUs
Grafana Assistant
AI observability agent in Grafana Cloud that writes PromQL, LogQL and TraceQL from plain English, investigates incidents and proposes code fixes as pull requests