Runpod is mid-priced for training: an NVIDIA H100 SXM Pod lists at $3.99 per GPU-hour on Secure Cloud and from $2.69 on Community Cloud, billed per second. For comparison, the same GPU is listed at $2.60/hr on the Nodus pricing page, and Nodus restores long runs from checkpoints automatically. Runpod bills usually grow past the rate card through storage on stopped Pods and idle hours nobody stopped.
Runpod spelled its name RunPod until 2025, and you will see both. We build Nodus, so weigh our view accordingly. Every Runpod number below links to Runpod's own pricing page or docs, and every sum is shown so you can check it.
What does a Runpod GPU cost per hour?
A Pod's hourly price covers the GPU plus the vCPUs and RAM listed next to it, and Runpod bills it per second with no commitment. Secure Cloud runs in Runpod's T3/T4 data centers; Community Cloud is cheaper and runs on peer-to-peer providers whose reliability Runpod lists as variable (Choose a Pod). Runpod also says it is no longer accepting new Community Cloud hosts, though existing resources remain available. Savings plans have been deprecated and can no longer be bought (Pod pricing), so the listed on-demand rate is what most people pay unless they negotiate with sales.
| GPU | Runpod Secure Cloud | Runpod Community Cloud | Nodus |
|---|---|---|---|
| RTX 4090 24 GB | $0.89/hr | not listed in the FAQ | $0.29/hr |
| L40S 48 GB | $1.09/hr | not listed in the FAQ | $0.793/hr |
| A100 SXM 80 GB | $1.79/hr | from $1.39/hr | $1.09/hr (A100 80 GB) |
| H100 SXM 80 GB | $3.99/hr | from $2.69/hr | $2.60/hr |
| H200 141 GB | $5.29/hr | from $3.59/hr | $3.593/hr |
| B200 180 GB | $7.99/hr | not listed in the FAQ | $5.50/hr |
Prices checked November 7, 2026, from the Runpod pricing page (Pod list and its FAQ for Community Cloud rates) and the hourly supply prices on the Nodus pricing page.
One rule shapes the first bill. To deploy an on-demand Pod, Runpod requires at least one hour of credits for the configuration you picked (Pod pricing). An 8 x H100 SXM Pod on Secure Cloud costs 8 x $3.99 = $31.92 per hour, so that is the minimum balance to start it.
How much does a month of fine-tuning cost on Runpod?
For a realistic month, the GPU hours dominate and storage is a small line, as long as you store data the right way. We priced a team that runs three full fine-tunes in a month, each 12 hours on one 8 x H100 SXM machine. That is 3 x 12 x 8 = 288 GPU-hours.
- Runpod Secure Cloud: 288 x $3.99 = $1,149.12
- Runpod Community Cloud: 288 x $2.69 = $774.72
- Nodus: 288 x $2.60 = $748.80
Add a 500 GB network volume for datasets and checkpoints, kept the whole month: 500 x $0.07 = $35.00. Network volumes attach only to Secure Cloud Pods (Network volumes), so the realistic Runpod total for this plan is $1,149.12 + $35.00 = $1,184.12.
Three 12-hour, 8 x H100 fine-tunes cost $400.32 less on Nodus than on Runpod Secure Cloud at list prices.
Community Cloud closes most of that gap, at $25.92 above Nodus for the month. Runpod's docs describe Community Cloud reliability as variable, which is fine for short experiments and riskier for a 12-hour run without automatic recovery. That trade-off, more than the rate, is what we would weigh. Before a run on Nodus, nodus run --dry-run prints the estimate for the exact request, so you can compare against your own Runpod numbers rather than ours.
Why does Runpod storage cost more when a Pod is stopped?
Runpod charges twice as much for a volume disk on a stopped Pod as on a running one: $0.10/GB/month while running and $0.20/GB/month while stopped (Pod pricing). Container disk is free when stopped, but it is erased when the Pod stops. Network volumes cost the same either way.
Take 500 GB kept for a 30-day month:
| Where the 500 GB lives | Rate | Month |
|---|---|---|
| Volume disk, Pod running | $0.10/GB/month | $50.00 |
| Volume disk, Pod stopped | $0.20/GB/month | $100.00 |
| Network volume, standard | $0.07/GB/month | $35.00 |
| Network volume, high-performance | $0.14/GB/month | $70.00 |
Rates from the Runpod pricing page storage section, checked November 7, 2026.
Parking checkpoints on a stopped Pod's volume disk costs nearly three times a standard network volume.
The decision rule we use: if data must outlive a Pod by more than a day, put it on a network volume from the start. Network volumes must be attached when you deploy a Pod and cannot be added later without deleting the Pod (Network volumes), so this is a choice you make up front. Runpod also says it is not designed for long-term cloud storage and recommends backing up critical data elsewhere.
Is Runpod Serverless cheaper than a Pod?
Serverless is cheaper only when your workers are idle much of the time. Runpod's Serverless list rates are higher per hour than Pods, but flex workers scale to zero and you pay nothing while they are scaled down (Serverless pricing). Billing runs from worker start until it fully stops, rounded up to the second. Model loading at cold start and the idle timeout (5 seconds by default) are billed too.
The break-even is the Pod rate divided by the Serverless rate. If workers would be busy for more of the month than that share, a Pod that runs all month costs less.
| GPU | Pod, Secure Cloud | Serverless | Break-even busy share |
|---|---|---|---|
| A100 80 GB | $1.79/hr | $2.72/hr | 65.8% |
| H100 80 GB | $3.99/hr | $5.39/hr | 74.0% |
| H200 141 GB | $5.29/hr | $6.99/hr | 75.7% |
| B200 180 GB | $7.99/hr | $10.79/hr | 74.1% |
Rates from the Runpod pricing page, checked November 7, 2026. Serverless H100 is the 80 GB tier the page labels PRO. Active workers, which run all the time, can get discounts through Runpod sales, which would move these thresholds.
Serverless costs about 35% more per hour on the H100, so it wins only when workers sit idle more than about a quarter of the time.
For training, this table mostly settles the question: a fine-tune keeps the GPU busy, so Serverless is the more expensive way to run it. Serverless fits the inference side, after the model is trained.
What hidden costs and limits catch people on Runpod?
Three account rules matter more than any rate, and all three are in Runpod's own docs (Pod pricing).
The first is the default spend limit of $80 per hour across all resources. At $3.99 per H100 SXM, that allows 20 H100s on Secure Cloud ($79.80/hr). Two 8-GPU Pods ($63.84/hr) fit, a third ($95.76/hr in total) does not until support raises the limit.
The second is what happens at a $0 balance. Pods stop. Pods with a network volume keep their data; Pods without one are terminated and their data cannot be recovered. Storage keeps billing while the Pod is stopped, so a network volume left unpaid at $0 may eventually be terminated too (Network volumes).
The third is idle time. Per-second billing only helps if someone stops the Pod. Leave an 8 x H100 SXM Secure Cloud Pod running for 14 hours overnight and it costs 8 x 14 x $3.99 = $446.88. That is more than a third of the whole month priced above.
How does Nodus price the same training work?
Nodus bills a run, not a machine, which removes most of the idle and stopped-Pod cases above. You submit a Job or TrainingJob, Nodus places it on the cheapest capacity that fits your limits, and the meter stops when the run finishes. Before paid work starts, Nodus places a funded hold on your prepaid balance and renews it every 5 minutes (When money runs out).
Running out of money is handled differently too. A checkpointed Job takes an urgent checkpoint, suspends, and resumes on its own once you top up, instead of being terminated. A per-run cap works the same way. This is the dry run from the Nodus quickstart:
nodus run --name cli-quickstart --gpu L4 --image nodus/pytorch \
--max-cost 1 --timeout 10m --no-warm --dry-run \
--output metrics=/nodus/outputs/metrics.json -- python train.py
The --dry-run flag shows the estimate and any blockers without creating the Job, and --max-cost 1 caps that Job at $1 (quickstart). Submit the same command without --dry-run once the estimate looks right. Budgets add enforced limits per org, project or person (Budgets). New organizations get a $30 starter credit after email and card verification, valid for 30 days (Credits).
If you already hold Runpod credits, Nodus can place work in a Runpod account you connect (Beta). You spend those credits and keep Nodus checkpoints and caps on top (cloud accounts). The full side-by-side is in Nodus vs RunPod, and other options are in RunPod alternatives.
Where is Runpod the better deal?
Runpod is a good deal in several real cases:
- Short interactive sessions. Per-second billing and a machine you control suit a two-hour debugging session that you start and stop yourself.
- Moving a lot of data. Pods have no ingress or egress fees, which helps when you pull large datasets in and push large checkpoints out.
- Inference with uneven traffic. Serverless flex workers scale to zero, so a lightly used endpoint can cost far less than an always-on machine.
- Community Cloud on the H200. Runpod lists it from $3.59/hr, a fraction of a cent below Nodus at $3.593/hr.
- Multi-node clusters today. Runpod Instant Clusters are generally available, listed at $6.35/hr per H200 SXM and $2.15/hr per A100 SXM on the Runpod pricing page. Multi-node training on Nodus is in Beta.
FAQ
Does Runpod charge for data transfer?
No. Runpod states there are no fees for data ingress or egress on Pods. Storage is billed separately, per second for container and volume disks and hourly for network volumes.
Can I still buy a Runpod savings plan?
No. Savings plans are deprecated. Existing plans stay active until they expire and cover GPU compute only; storage is billed at standard rates.
Is Runpod Community Cloud safe for long training runs?
It works, but plan for interruptions. Runpod describes Community Cloud reliability as variable, and network volumes are Secure Cloud only, so save checkpoints somewhere that outlives the Pod. Our guide to resuming training after a preemption shows the code.
What is the cheapest way to fine-tune a small model?
For a LoRA fine-tune that fits in 24 GB, an RTX 4090 is usually enough. It lists at $0.89/hr on Runpod Secure Cloud and $0.29/hr on Nodus, so a 6-hour run is $5.34 or $1.74.
Originally published on the Nodus blog.
Top comments (0)