We're all used to cloud instances being temporary, but losing an entire day's worth of checkpoints because you forgot to configure persistent storage is a different kind of pain. Most GPU providers bill compute and storage separately, which is fine until you realize your data only lives as long as the instance does. The platforms below flip that model and give you persistent volumes by default, so saving your GPU workspace between sessions isn't something you have to remember to do.
TLDR:
When you spin up a GPU instance and start training a model, everything lives in temporary disk space tied to that instance. Shut it down and the data disappears with it.
Persistent GPU storage keeps your files on a separate volume that outlives any individual instance. You can pause a job, swap to a different GPU, and find your workspace exactly as you left it.
For long training runs, this matters a lot. Losing checkpoints to an unexpected shutdown wastes compute time you already paid for. Persistent volumes make that a non-issue. Model checkpointing best practices highlight the importance of reliable storage infrastructure.
Each option here was assessed against a consistent set of publicly available criteria, pulled from documentation, pricing pages, and product specs to keep the comparison grounded in what's verifiable. When comparing cloud GPU providers, storage persistence often gets overlooked despite its impact on workflow continuity.
Here are the criteria we used:
Thunder Compute Local is built for teams and solo researchers who need GPU sessions that actually persist between runs. When you spin up an instance, your workspace, datasets, and model checkpoints stay exactly where you left them, no re-uploading, no re-mounting, no lost progress.
Storage is attached at the instance level, so read/write speeds stay fast and latency stays low. You get persistent volumes that survive instance stops, making iterative training workflows significantly less painful.
Thunder Compute Local is a strong fit for ML engineers running multi-day fine-tuning jobs or anyone tired of rebuilding their environment from scratch every session.
RunPod offers GPU instances with persistent storage volumes that survive session termination, which is a key differentiator for iterative ML workflows. Users can attach storage volumes across different pod types, keeping datasets and checkpoints intact between runs.
Storage is billed separately from compute, typically at around $0.10 per GB per month for network volumes. This keeps costs predictable when you need to pause training and resume later without re-uploading data.
RunPod supports both community and secure cloud tiers, with the secure tier offering more reliability guarantees for production-grade workloads that depend on consistent storage availability.
Crusoe is a GPU cloud provider built around a sustainability angle, powering its data centers with stranded or flared natural gas rather than grid electricity. That framing appeals to teams that want to reduce the carbon footprint of compute-heavy workloads without changing how they work.
On storage, Crusoe offers persistent volumes that attach to GPU instances and survive instance termination. Volumes are provisioned separately from compute, so you can spin down a GPU, keep your dataset and model checkpoints intact, and reattach when you resume. Pricing for persistent storage sits around $0.04–$0.08 per GB per month depending on region and tier.
Crusoe supports NVIDIA H100 instances alongside its persistent volume offering. The workflow is straightforward: create a volume, attach it to an instance at launch, and detach it when you stop. There is no automatic snapshotting by default, so teams need to build their own backup cadence if data durability beyond volume persistence is required.
Crusoe is a reasonable choice for teams where sustainability reporting matters and who need basic persistent storage without a complex managed storage layer on top.
Lambda Labs offers on-demand and reserved GPU instances with persistent storage attached to each instance. Storage persists between sessions by default, so you don't lose your dataset or model weights when you stop a machine.
Lambda provides up to 1 TB of persistent SSD storage per instance. Pricing is competitive for research teams, with H100 instances starting around $2.49/hr. The storage is tied to the instance, so there's no separate volume management required.
This simplicity works well for solo researchers but can limit flexibility for teams that need shared or independently scalable storage volumes.
Atlas Cloud is a GPU cloud provider focused on enterprise and research workloads, offering persistent storage as a standard feature across its instances. Users can attach block storage volumes that survive instance termination, making it a practical choice for teams that need to preserve model checkpoints and datasets between sessions.
Storage is provisioned separately from compute, so you only pay for what you use. Volumes are available in SSD-backed tiers, and snapshots can be scheduled for automated backups.
NodeAI is a GPU cloud provider built with researchers and developers in mind. It offers persistent storage as a core feature rather than an afterthought, meaning your datasets, checkpoints, and environment configurations survive between sessions without any manual intervention.
Storage is attached at the volume level, so you can mount the same workspace across different GPU instances. This is useful when scaling from a smaller GPU for prototyping to a larger one for full training runs.
Pricing sits in a mid-range bracket relative to hyperscalers, making it a reasonable option for teams that need reliable storage without committing to AWS or GCP contracts.
| provider | storage type | persistent? | storage included | pricing model |
|---|---|---|---|---|
| Thunder Compute Local | NVMe SSD (local) | yes | yes, with instance | pay-as-you-go |
| AWS | EBS / S3 | yes | add-on | per GB + I/O fees |
| Google Cloud | Persistent Disk / GCS | yes | add-on | per GB provisioned |
| Azure | Managed Disk / Blob | yes | add-on | per GB + transactions |
| CoreWeave | SSD / HDD volumes | yes | add-on | per GB/month |
Storage persistence and billing structure vary widely across GPU cloud providers. Some attach storage at the instance level with no extra configuration, while others treat it as a separate billable resource with its own provisioning steps.
A few things worth knowing about the table above:
Thunder Compute Local is built around a simple idea: your GPU workspace should persist between sessions without any extra configuration. When you spin up an instance, your files, datasets, models, and environment variables are all exactly where you left them.
Storage is backed by high-performance NVMe SSDs, so checkpoint reads and dataset loading stay fast even during long training runs. There are no manual snapshot workflows or re-upload steps between sessions.
Pricing is straightforward, and you only pay for compute time. Persistent storage is included rather than billed as a separate line item, which keeps costs predictable for teams running recurring workloads.
Persistent GPU storage matters most when you're deep into a training run and can't afford to lose progress. The best setup is one where your workspace just persists without extra steps, separate charges, or surprise I/O fees. Thunder Compute Local built storage persistence directly into every instance, so your datasets and checkpoints are always there when you spin back up. Find a provider that treats storage as part of the workflow, not an afterthought you have to configure yourself.
Ready to stop worrying about lost checkpoints? Get started with Thunder Compute Local and experience persistent GPU storage that just works.
Start by identifying whether you need shared volumes across multiple instances or instance-attached storage. If you frequently switch between GPU types during training, look for providers where volumes mount independently of compute. For simpler workflows, instance-attached storage with automatic persistence requires less configuration overhead.
Thunder Compute Local and RunPod both offer persistent volumes that survive instance stops, making them practical for long training runs where you need to pause overnight and resume without losing checkpoints. Look for providers that include storage with the instance rather than billing it separately to avoid cost surprises during extended jobs.
Yes, most providers on this list support persistent volumes that detach from one instance and reattach to another. RunPod, Atlas Cloud, and NodeAI all provision storage independently from compute, so you can mount the same workspace across different GPU configurations without re-uploading your datasets.
With Thunder Compute Local and Lambda, your storage persists automatically without manual snapshot workflows, so your data survives instance termination by default. Crusoe and some other providers require you to manage your own backup schedule, meaning you could lose uncommitted work if an instance stops unexpectedly.
Choose separate volumes when you need shared access across multiple instances, independent scaling of storage capacity, or when your storage needs outlast individual training jobs. Instance-attached storage makes more sense for solo researchers or small teams who want faster setup and predictable billing without managing volume lifecycles manually.
