# Crusoe cuts GLM-5.3 startup to 31 seconds _Crusoe and NVIDIA cut GLM-5.3 startup from 8 minutes to 31 seconds by restoring GPU snapshots from local NVMe._ **Published:** 2026-11-10 **Source:** https://www.startuphub.ai/ai-news/technology/2026/crusoe-cuts-glm-5-3-startup-to-31-seconds --- Spinning up another copy of a huge inference worker should be instant, yet teams often wait minutes while weights load and runtimes warm. [Crusoe](https://www.crusoe.ai/resources/blog/nvidia-snapshot-crusoe-cloud) says it cut that wait for GLM-5.3 from about eight minutes to 31 seconds with NVIDIA Snapshot, in work done with NVIDIA on Crusoe Cloud. Storage was the fix. Snapshot is a Kubernetes-native checkpoint system that freezes a fully initialized GPU worker into an artifact a new pod can revive. Checkpoint is a saved copy of live CPU and GPU memory that avoids a cold rebuild. A privileged agent on each GPU node takes the checkpoint by invoking CRIU, which is the Linux tool that freezes and revives live processes, alongside [NVIDIA](https://www.startuphub.ai/ai-news/artificial-intelligence/2026/ai-capital-race-nvidia-earnings-softbank-alibaba) cuda-checkpoint for GPU state. To scale, a new pod points at the artifact and the agent replays GPU memory first, then lets CRIU rebuild the process and its mounts. On NVIDIA HGX B300 hardware, which is an eight-GPU baseboard built around B300 modules, Crusoe said a 594 GB restore took 15.0 seconds from cold local NVMe, or about 40 GB/s, versus 33.3 seconds from NFS at about 18 GB/s, per Crusoe. Total agent restore fell from 41.4 seconds to 22.9 seconds, while CRIU stayed near 5.1 seconds on both backends, per Crusoe. Pod created to Ready fell from 50 seconds to 31 seconds, for a 15.5x gain over the roughly eight-minute cold start, per Crusoe. Eight seconds still sit outside the restore, in pod startup. Crusoe also hit the part Snapshot leaves unfinished. A restore failed with binfmt_misc: Operation not permitted, which looked like permissions but was an early CRIU warning, not the cause, per Crusoe. The real block was a missing /cache/vllm mount in the restore pod, so CRIU could not rebuild the mount namespace, and Crusoe’s fix in pull request 532 surfaces the last fatal error instead, per Crusoe. Teams can run Snapshot today on Crusoe Cloud with release v0.2.0-rc.2 plus Crusoe’s tuning in pull request 602, per Crusoe. If you adopt it, match the restore pod spec to the source mounts and put the artifact on local NVMe first, because that is where the seconds are. --- Original analysis from [startuphub.ai](https://www.startuphub.ai), the #1 AI startup directory. © StartupHub.ai. All rights reserved. You may not republish this article in full without a license. Search engines and AI research tools may crawl and summarize for reference. Bulk reproduction or model training on this content requires a license. See https://www.startuphub.ai/terms.