Thunder Compute sells GPU virtual machines with a deliberately small catalog and a developer experience built around VS Code. Runpod runs a broader fleet across Pods, Serverless and Clusters.
Both are aimed at developers rather than enterprise procurement, and on several cards their prices sit within pennies of each other. So price is not what separates them. What separates them is what happens after you have rented a machine.
| Capability | Thunder Compute | Runpod |
|---|---|---|
| What you rent | GPU virtual machines | Pods, Serverless and Instant Clusters |
| GPU models published | 4 | 24, including AMD MI300X |
| Serverless for your own container | Not published | Serverless, sub-200ms cold starts via FlashBoot |
| Multi-node training | Not published, 1 to 8 GPUs per server | Instant Clusters, self-serve |
| Billing granularity | Per minute | Per second |
| Instance snapshots | Yes, restore an exact environment | Network volumes |
| Editor-first workflow | Built around VS Code | Shell, notebook or your own image |
| Ingress and egress charges | None | None |
Thunder Compute capabilities and rates were read from thundercompute.com/pricing on 21 August 2026 and may change. Runpod rates pull live.
What Thunder Compute does well
A small catalog is a feature for some teams. Four models means no decision paralysis, and the four chosen cover most common workloads from a 48GB A6000 up to an H100.
Developer experience. Thunder's tooling centers on connecting an editor directly to a remote GPU, which suits the write-run-iterate loop well. If that is how you work, it is a genuinely nice product and the reason to choose them.
Honest, legible pricing. Rates are on one page with no tiers to decode, and their comparison table quotes competitors accurately, including us.
Snapshots. Instance snapshots let you restore an exact environment later without rebuilding it, billed hourly by size.
Where Thunder Compute stops
Four GPU models. As of 21 August 2026, there is nothing below an RTX A6000, nothing above an H100 PCIe, and no H200, B200 or B300. If your workload needs a smaller card to be economical or a newer one to be possible, it is not on the menu.
No serverless tier. Thunder rents virtual machines and publishes no scale-to-zero inference endpoint, so a served model bills continuously whether or not requests arrive. For an endpoint with uneven traffic that is a structural cost, not a rate difference.
No multi-node training. Their published configurations run from one to eight GPUs in a single server. Training that needs more than one node has no path there.
Where Runpod fits
Runpod covers three workload shapes on one account, with the same container images across all of them. Pods for development and long-running jobs. Serverless for inference that scales to zero, with sub-200ms cold starts via FlashBoot. Instant Clusters for multi-node training, provisioned without a sales call.
That matters most at the transition points. A team that prototypes on a VM, then needs to serve the result, then needs to train something larger, changes configuration on Runpod rather than changing provider. On Thunder, the second and third steps are not available.
The lineup runs 24 GPU models from an RTX A5000 through to a B300, including AMD MI300X. Runpod is SOC 2 Type II, independently verified for HIPAA and GDPR, and runs across 31 global regions.
Billing granularity is the difference that shows up on the invoice
Thunder Compute bills per minute. Runpod bills per second. On a long training run the difference is noise. On short, bursty inference calls it is not: a nine-second job billed to the minute costs the same as a fifty-nine-second one.
Storage differs in shape too. Thunder includes the first 100GB while an instance is running, then charges $0.03 per 100GB per hour, with snapshots at $0.05/GB per month, additional vCPUs at $0.04 each per hour and 8GB of RAM included per vCPU. Runpod charges $0.10/GB per month for container disk and $0.07/GB per month for network storage under 1TB, dropping to $0.05 above it. Neither charges for egress.
Pricing compared, card by card
On the four cards Thunder publishes, neither platform wins outright.
| Card | Thunder Compute | Runpod Community | Runpod Secure |
|---|---|---|---|
| H100 PCIe | $2.19/hr | {{gpu:h100-pcie:community}}/hr | {{gpu:h100-pcie}}/hr |
| A100 80GB | $1.09/hr | {{gpu:a100-pcie:community}}/hr | {{gpu:a100-pcie}}/hr |
| L40 | $0.79/hr | {{gpu:l40:community}}/hr | {{gpu:l40}}/hr |
| RTX A6000 | $0.35/hr | {{gpu:rtx-a6000:community}}/hr | {{gpu:rtx-a6000}}/hr |
Thunder Compute rates were read from their own pricing page on 21 August 2026 and change without notice. Runpod rates pull live.
Thunder's own comparison table quotes Runpod at $1.19 for an A100 and $1.99 for an H100. Those are our Community Cloud rates and they are quoted accurately, which is worth saying because plenty of vendor comparison tables are not.
Which one should you choose
Choose Thunder Compute if your workload fits one of their four cards, you work primarily in an editor connected to a remote machine, and you want the simplest possible pricing page.
Choose Runpod if you need a GPU they do not carry, if you are serving inference and want to stop paying between requests, if you need more than one node for training, or if per-second billing materially changes your cost on short jobs.
The honest split: for a single developer on a single VM running an A100, Thunder Compute is a good choice and the editor integration is a real draw. The case for Runpod is not that it undercuts them, because on that card it does not. It is that the platform keeps working when the workload changes shape.
Frequently asked questions
What is the main difference between Runpod and Thunder Compute?
Scope. Thunder Compute rents GPU virtual machines from a four-card catalog. Runpod covers Pods for development, Serverless for inference that scales to zero, and Instant Clusters for multi-node training, across 24 GPU models on one account.
Does Thunder Compute offer serverless GPUs?
No. Thunder Compute provides GPU virtual machines from one to eight GPUs, with no published scale-to-zero serverless tier. Runpod Serverless runs your own container, scales to zero between requests and cold starts in under 200ms via FlashBoot.
How many GPU models does Thunder Compute offer?
Four: RTX A6000, L40, A100 80GB and H100 PCIe. Runpod publishes 24 models, from an RTX A5000 through to a B300.
Can I run multi-node training on Thunder Compute?
Not from their published offering. Configurations scale from one to eight GPUs within a single server. Runpod Instant Clusters provide multi-node training self-serve, at {{cluster:h200-sxm}}/hr for H200 SXM and {{cluster:a100-sxm}}/hr for A100 SXM.
Is Thunder Compute cheaper than Runpod?
On the A100 80GB, yes: $1.09/hr against Runpod's {{gpu:a100-pcie:community}}/hr Community rate. On the H100 PCIe, Runpod Community is lower at {{gpu:h100-pcie:community}}/hr against $2.19. On the RTX A6000 the two are close. Compare the specific card you need, and factor in per-minute against per-second billing if your jobs are short.
Does Thunder Compute offer persistent storage?
Yes, through instance snapshots billed hourly by size at $0.05/GB per month. The first 100GB is included while an instance is running, with additional storage at $0.03 per 100GB per hour. Runpod offers network volumes that persist independently of any pod.
Do either charge for data egress?
Neither does. Both state no egress charges.
