RunPod

I rented GPUs on RunPod for a batch of inference work rather than buying a card, and the arithmetic was easy: an RTX 4090 at $0.34 an hour on Community Cloud is about $1.40 for an afternoon. A card that does the same job sits on my desk costing money whether I use it or not.

Then I looked at the bill and the GPU was not the interesting line.

Verdict

RunPod is genuinely good at the thing it advertises. You get a real Linux box with a real GPU, root SSH, per-second billing and no subscription, and for occasional heavy work that beats owning hardware by a wide margin. I would recommend it to anyone who needs a big card a few times a month.

What nobody tells you is that the storage model runs on completely different logic to the compute model, and if you treat it like the GPU you will quietly pay for nothing. Compute stops when you stop. Storage does not, and it gets more expensive the moment you stop using it.

What it’s good at

The per-second billing is the whole pitch and it works. Community Cloud is substantially cheaper than Secure Cloud for identical silicon, and for anything you can restart safely it is the obvious choice.

Root SSH is what makes it useful to me rather than merely cheap. It is an ordinary remote Linux machine, so my normal tooling works, and I can point Claude Code at it the same way I would at any other host. No notebook interface to fight, no proprietary runtime to learn.

What it’s bad at

Here is the number that reorganised how I use the service.

Volume disk is $0.10/GB/month while the pod runs and $0.20/GB/month while it is stopped. Storage costs double when you are not using it.

So the instinct everybody has, stop the pod to save money, is wrong in two directions at once. Your storage bill doubles, and stopping also releases the GPU back to the pool while your pod stays pinned to that specific machine. RunPod’s own FAQ has an entry for restarting with zero GPUs attached, because somebody else rented the card while you were away.

Put a number on the idle disk and it gets worse. 100 GB sitting stopped is $20 a month, which is roughly 59 hours of the 4090 you are not using. Break-even against five hours of 4090 time a month arrives at 8.5 GB of idle volume.

How I use it

Keep the pod’s own disk small. Whatever you actually need to keep goes on a network volume at $0.07/GB/month, which is cheaper than idle volume disk, does not double, and is not tied to one physical machine.

Then terminate rather than stop. Redeploy a fresh pod when you next need a GPU, reattach the same network volume, and you land on whatever hardware is free instead of queueing for one specific card that may be taken.

Everything else happens over SSH from my own terminal. The model weights get pulled from inside the pod rather than pushed from my laptop, which sidesteps the fact that the basic proxied SSH cannot do scp or sftp. The pod is disposable. The volume is the only thing that persists, and it is the only thing I pay for continuously.

When I’d skip it

If you need a GPU most days, rent the hours for a month, work out what you actually spent, and compare it against a used 3090. At sustained load the maths flips toward owning.

If you only ever call a model rather than running one, skip GPU rental entirely and use a hosted inference API. RunPod is worth it when you need the machine, not just the model.