Runpod pricing: every GPU, Serverless and storage rate explained
Short answer: Runpod charges by the second for GPU compute, with no monthly subscription required. As of October 2026, Pods (full GPU machines) range from about $0.16 an hour for an entry card on Community Cloud up to around $8.99 an hour for its most powerful Secure Cloud hardware. Serverless workers, which scale to zero when idle, bill from about $0.58 to just under $12 an hour depending on the card, only while a request is actually being handled. Storage is billed separately by the gigabyte per month. There's no free trial, but you only pay for what you actually use.
You want to know exactly what Runpod will cost you before you commit a project budget to it.
I'm Ranjit, and I run Blur Studio: 15+ years building software, 120 apps taken live, and clients who have gone on to raise $10M+. I build AI systems and processes that help businesses do a lot more with a lot less time. I also work differently from most of what else is out there, more on that further down this page. Say hello.
This page may contain affiliate links from which Blur Studio earns commissions.
Runpod is a GPU cloud rental service: a platform where you rent a graphics card by the hour or the second to run AI workloads, instead of buying the hardware yourself. It has no flat monthly subscription. Every price below is confirmed against Runpod's own pricing page as of October 2026, and I'll flag anywhere a number depends on your specific setup rather than a fixed rate.
Runpod Pod pricing by GPU
Pods are full GPU machines, billed per second while running, on two tiers: Community Cloud (third-party hosted, cheaper) and Secure Cloud (Runpod's own data centers, pricier, more consistent). Here's the current range, cheapest to most expensive on Community Cloud.
| GPU | VRAM | Community Cloud | Secure Cloud |
|---|---|---|---|
| RTX A5000 | 24GB | $0.16/hr | $0.27/hr |
| RTX 3090 | 24GB | $0.22/hr | $0.50/hr |
| RTX A6000 | 48GB | $0.33/hr | $0.59/hr |
| RTX 4090 | 24GB | $0.34/hr | $0.89/hr |
| A40 | 48GB | $0.35/hr | $0.59/hr |
| L4 | 24GB | $0.44/hr | $0.59/hr |
| RTX 5090 | 32GB | $0.69/hr | $1.19/hr |
| L40 | 48GB | $0.69/hr | $0.82/hr |
| RTX 6000 Ada | 48GB | $0.74/hr | $0.99/hr |
| L40S | 48GB | $0.79/hr | $1.09/hr |
| A100 PCIe | 80GB | $1.19/hr | $1.79/hr |
| A100 SXM | 80GB | $1.39/hr | $1.79/hr |
| RTX Pro 6000 | 96GB | $1.69/hr | $2.49/hr |
| H100 PCIe | 80GB | $1.99/hr | $2.89/hr |
| H100 NVL | 94GB | $2.59/hr | $3.19/hr |
| H100 SXM | 80GB | $2.69/hr | $3.99/hr |
| H200 | 141GB | $3.59/hr | $5.29/hr |
| B200 | 180GB | $5.98/hr | $7.99/hr |
| B300 | 288GB | $6.94/hr | $8.99/hr |
These rates were current on Runpod's pricing page on 2026-10-30. Confirm current pricing directly before budgeting a large job, since GPU cloud rates shift as new hardware comes online.
Runpod Serverless pricing
Serverless deploys your model as an endpoint that spins up a worker when a request arrives and scales back to zero when there's none, so you stop paying the moment traffic stops. Worker pricing runs by active hour, roughly $0.58 an hour for smaller cards (L4, A5000, RTX 3090 class) up to around $5.39 for an H100, $6.99 for an H200, and just under $12 an hour for its largest B300 workers. The exact rate depends on the card your endpoint is configured to use, and only the active time counts toward your bill, not the idle time while it's scaled down. My Runpod Serverless guide walks through setting one up.
Storage pricing
Storage is billed separately from compute, by the gigabyte per month. Container disk and volume disk while a machine is running cost $0.10 a gigabyte a month. An idle volume disk (attached but not running) costs $0.20 a gigabyte a month. Network storage, used to persist files across machines, costs $0.07 a gigabyte a month under 1 terabyte, dropping to $0.05 above that, or $0.14 a gigabyte a month for a high-performance tier. How to stop re-downloading big AI model files every time covers the storage side of this.
Instant Clusters pricing
For training jobs too large for a single GPU, Runpod's Instant Clusters product provisions several GPUs connected together for a multi-GPU job. Published rates here are narrower than the main Pod lineup: an H200 SXM cluster runs $6.35 an hour and an A100 SXM cluster runs $2.15 an hour, while larger configurations (L40S, H100 SXM and B200 clusters) move to a "contact sales" model rather than a published self-serve rate. This tier is aimed at genuinely large training runs, not the typical small business workload, and most readers of this guide won't need it.
Worked monthly cost examples
A solo developer fine-tuning a small model a few hours a week: renting an RTX 4090 on Community Cloud at $0.34 an hour for roughly 10 hours a month costs about $3.40 in compute, plus a few dollars of storage for model files left between sessions. Even generous use rarely crosses $20 to $30 a month at this scale.
A small business running an inference endpoint with light, irregular traffic: a Serverless endpoint on an L4-class worker at roughly $0.58 an hour, active for maybe 40 hours across a month of real but irregular usage, lands near $23 in compute, with no charge during the idle time in between, which is the entire point of choosing Serverless over a Pod running continuously.
A team running a heavier workload nearly continuously: an H100 SXM Pod on Secure Cloud at $3.99 an hour, running 12 hours a day for a month (roughly 360 hours), comes to about $1,436 in compute for that month. At that usage level, it's worth comparing a reserved or dedicated option against continuing on-demand, since sustained heavy use is where the per-hour rate adds up the fastest.
Savings tips
- Use Community Cloud for anything that can tolerate occasional variability in availability, since it's reliably cheaper than Secure Cloud for the same card.
- Set a spend limit on the account before your first real job. A forgotten running Pod is the most common way this costs more than planned.
- Move bursty or unpredictable traffic to Serverless rather than a Pod you leave running continuously "just in case," since you only pay while a request is actually being handled.
- Match the card to the job, not the other way around. An oversized GPU for a small model wastes money per hour for no speed benefit on a job that doesn't need the extra VRAM.
How billing actually works
Runpod bills by the second rather than rounding up to the nearest hour, which matters more than it sounds like once you're running a lot of short jobs. A Pod that runs for 11 minutes and 20 seconds is billed for exactly that, not for a full hour, which is a meaningfully different experience from older cloud billing conventions that round every partial hour up. You add funds to your account balance up front, and compute and storage both draw down against that balance as you use them, rather than receiving an invoice after the fact. This makes it easier to cap your exposure: once the balance runs out, machines stop, rather than a bill quietly accumulating past what you intended to spend.
When a free or cheaper alternative is enough
If you're only experimenting with a small model and don't need dedicated hardware, a free-tier notebook service or your own laptop (if it has a decent GPU) may genuinely cover early testing before you need to pay for anything. Runpod earns its cost once you need a specific card, more VRAM than your own hardware has, or a model running continuously for other people to use, not for casual one-off experimentation that a free tool already covers. The Runpod alternatives guide compares Vast.ai, Lambda and Modal on price.
How I'd approach this
When I build an AI feature for a client that needs real GPU compute, I size the hardware to the actual job rather than defaulting to the biggest card available, and I build in the spend controls and monitoring so the bill matches real usage instead of a guess. Get in touch if you want this costed out properly before you commit a budget.
Questions people ask
Does Runpod have a free trial?
Runpod doesn't advertise a free trial or free credits as of October 2026. You pay for compute from the first second, though rates are low enough that a short test job costs very little.
What's the cheapest GPU on Runpod?
The RTX A5000 is the cheapest listed card at $0.16 an hour on Community Cloud as of October 2026, suitable for smaller models and lighter workloads.
Is Runpod Serverless cheaper than running a Pod continuously?
For irregular or bursty traffic, yes, since Serverless only bills while a request is being handled and scales to zero in between. For traffic that's genuinely constant all day, a Pod can work out more predictable.
Does Runpod charge for storage separately from compute?
Yes. Storage (container disk, volume disk and network storage) is billed by the gigabyte per month, separately from the hourly GPU compute rate.
How do I avoid a surprise Runpod bill?
Set a spend limit on your account and build "stop the machine" into your actual workflow once a job finishes. A running Pod bills by the second until you stop it.
Is Secure Cloud worth the higher price over Community Cloud?
If your job needs consistent, guaranteed availability, Secure Cloud's higher price buys that reliability. For jobs that can tolerate some variability, Community Cloud is reliably cheaper for the same hardware.