DigitalOcean sells GPUs the way it sells everything else: one product page, one price, no quote form. For a lot of people that alone is worth paying for — you already have the account, the object storage, and the networking, and the GPU lands next to them. This post is the part the product page does not tell you: what a GPU Droplet costs per GPU-hour against the rest of the market, and which of its two disks is going to lose your work.
TL;DR: GPU Droplets are billed per second and priced per GPU: $4.41/hr for an H100, $4.47 for an H200, $2.59 for an MI300X, $1.57 for an L40S or RTX 6000 Ada, $0.76 for an RTX 4000 Ada, as of DigitalOcean's August 1, 2026 pricing. Against 1,044 live offers from nine providers, that is roughly double the market low on the small cards and about a third above the H100 median — but it is genuinely competitive on H200, where the market median is $4.59. The thing to actually plan around is storage: the 5 TiB "scratch" disk attached to every multi-GPU config is documented as non-persistent, is excluded from snapshots, and is lost when the Droplet is destroyed — and destroying is the only way to stop the meter, because a powered-off GPU Droplet keeps billing in full.
What a GPU Droplet is
A GPU Droplet is an ordinary DigitalOcean Droplet with a GPU attached, in seven shapes. Every configuration below is from DigitalOcean's own GPU Droplet pricing page, which states its rates are "effective as of August 1, 2026":
| GPU | VRAM | Count | vCPU | RAM | Boot disk | Scratch disk | Per GPU/hr |
|---|---|---|---|---|---|---|---|
| NVIDIA HGX H200 | 141 GB | 1 or 8 | 24 | 240 GiB | 720 GiB NVMe | 5 TiB NVMe | $4.47 |
| NVIDIA HGX H100 | 80 GB | 1 or 8 | 20 | 240 GiB | 720 GiB NVMe | 5 TiB NVMe | $4.41 |
| AMD Instinct MI325X | 256 GB | 1 or 8 | 20 | 164 GiB | 720 GiB NVMe | 5 TiB NVMe | $3.80 |
| AMD Instinct MI300X | 192 GB | 1 or 8 | 20 | 240 GiB | 720 GiB NVMe | 5 TiB NVMe | $2.59 |
| NVIDIA RTX 6000 Ada | 48 GB | 1 | 8 | 64 GiB | 500 GiB NVMe | — | $1.57 |
| NVIDIA L40S | 48 GB | 1 | 8 | 64 GiB | 500 GiB NVMe | — | $1.57 |
| NVIDIA RTX 4000 Ada | 20 GB | 1 | 8 | 32 GiB | 500 GiB NVMe | — | $0.76 |
Two details worth pulling out of that table before the price comparison. The multi-GPU shapes are sold as 1 or 8 — there is no 2, 4, or 6. And DigitalOcean's own two sources disagree on the billing floor: the pricing page says GPU Droplets are "billed per second with a 5-minute minimum billing period", while the Droplet pricing documentation says they are "billed per second with a minimum charge of 60 seconds or $0.01, whichever is higher". Both are DigitalOcean's; we have not been able to reconcile them, and on a workload of short bursty runs the difference is real, so budget for the five-minute floor.
What it costs against the rest of the market
We run a marketplace across providers, so we can price the same card in a lot of places at once. The figures below come from our public GPU price index, generated 2026-08-28 from 1,044 live offers across nine providers and 126 regions. Every number is a per-GPU hourly rate — no row is a whole-node price divided by nothing, which is the single most common way these comparisons go wrong.
| GPU | DigitalOcean | Market low | Market median | Offers / providers |
|---|---|---|---|---|
| H100 80GB | $4.41 | $2.20 | $3.29 | 81 / 6 |
| H200 141GB | $4.47 | $3.93 | $4.59 | 45 / 4 |
| MI300X 192GB | $2.59 | $2.39 | $2.39 | 7 / 2 |
| L40S 48GB | $1.57 | $0.79 | $0.99 | 39 / 3 |
| RTX 6000 Ada 48GB | $1.57 | $0.66 | $0.74 | 14 / 3 |
| RTX 4000 Ada 20GB | $0.76 | $0.28 | $0.28 | 5 / 2 |
Read that honestly and it splits in three.
On the small cards, DigitalOcean is roughly double the market. An RTX 4000 Ada at $0.76 against a $0.28 low is a 2.7x spread; RTX 6000 Ada at $1.57 against $0.66 is 2.4x. If your workload is a single 20-48GB card doing inference or fine-tuning, this is the most expensive way to buy it that still counts as a fair price.
On H100 it is above the median but not absurd — $4.41 against a $3.29 median across 81 offers. You are paying about a third over the middle of the market for the integration.
On H200, DigitalOcean is genuinely cheap. Its $4.47 sits below the $4.59 median of 45 offers from four providers. H200 supply is still thin enough that the market has not compressed, and DigitalOcean has priced it as if it had. If you want an H200 this week, this is a real option and we will say so.
Two caveats in DigitalOcean's favour, because leaving them out would make this a sales page. MI325X is not on our index at all — 256GB of HBM3E in a mainstream cloud is not something most of the market carries, and at $3.80 it has no comparison here to lose. And our MI300X row is thin: 7 offers from 2 providers, so a $2.39 "median" is two suppliers agreeing, not a market.
The two disks, and which one loses your work
Every multi-GPU GPU Droplet ships with two local disks, and they behave nothing alike.
The boot disk (720 GiB on the 8x shapes, 500 GiB on the single cards) is the persistent one — OS, frameworks, and anything you meant to keep. The 5 TiB NVMe scratch disk beside it is not. DigitalOcean's documentation is direct about this. It is "a local, non-persistent disk to store data for staging purposes, like inference and training". "The scratch disk is not included in snapshots of the Droplet." "If you destroy or recreate a GPU Droplet, the scratch disk is lost." "The scratch disk has no redundancy, so hardware failures may also lead to loss of data." And it is "not automatically mounted" — you get a pre-formatted ext4 partition and you mount it yourself (DigitalOcean docs).
None of that is a defect. A scratch disk that says it is scratch is doing its job. The problem is what it collides with next.
Powering off does not stop the bill
DigitalOcean's billing documentation states it plainly: "You are still billed for GPU Droplets that are powered off. The GPU and other compute resources stay reserved on the hypervisor even when the Droplet is not running." And then the remedy: "To end billing, destroy the GPU Droplet."
Put the two facts side by side and the trap is structural rather than sneaky. The only action that stops you paying is the same action that erases the 5 TiB disk. Your options at the end of a training run are to keep paying for a machine you are not using, or to destroy the machine and lose whatever was staged on scratch.
The arithmetic is not small. An 8x H100 GPU Droplet is 8 x $4.41 = $35.28/hr. Leave it powered off across a weekend you were not working — 48 hours — and that is $1,693.44 for an idle machine. Leave a single H200 powered off for a week and it is $751. Powering off buys you nothing but the illusion of thrift.
The documented escape is a snapshot: take one, destroy the Droplet, restore later. That works, with two costs. Snapshots are charged at $0.06 per GB per month (DigitalOcean snapshot pricing), so a full 720 GiB boot disk is about $43/month to hold. And the snapshot covers the boot disk only — the scratch disk is explicitly excluded, so anything you staged there has to be re-staged from object storage on the way back up, which for a multi-terabyte dataset is the slow part of your morning.
If you got here from Paperspace
Paperspace is part of DigitalOcean now, and a lot of the people searching for GPU Droplets are Paperspace users working out where they land. Two things carry over that are worth knowing in advance.
The billing shape is the same trap in a different wrapper: a powered-off Paperspace machine keeps billing storage, its public IP and its add-ons by the hour, which we went through in detail in Paperspace storage: powering off won't stop billing. If that is the thing that pushed you to look around, note that GPU Droplets do not fix it — they make it larger, because on a GPU Droplet the full compute rate keeps running too, not just storage.
The persistence model, though, is a real downgrade in one specific way. A Paperspace machine keeps its disk. A GPU Droplet keeps its boot disk and throws away the 5 TiB you were most likely to put a dataset on. If your Paperspace habit was "leave everything on the machine", that habit does not port.
When a GPU Droplet is the right call
It is the right call more often than the price table suggests, and it is worth saying which cases.
If the rest of your application already lives on DigitalOcean, one vendor, one bill and one VPC is worth real money in operational time. If you want an H200 specifically, the pricing above is competitive today. If your work is a bounded run — start, train, snapshot, destroy — the per-second billing and the 5 TiB of local NVMe are exactly the right shape, and the non-persistence never bites you because you were not planning to come back to that disk anyway.
It is the wrong call if your pattern is the other one: a box you return to, with an environment you built up over weeks, that you would like to stop paying for overnight. That workload is where $0.76 for an RTX 4000 Ada against a $0.28 market low compounds into something worth moving over, and it is where the destroy-to-stop-billing rule turns into an actual data-loss event rather than a line in the docs.
What we do about it
We build the second case. Aquanode is a marketplace across providers rather than one fleet, so the price you pay is the market's, not a single vendor's — the live marketplace and the price index above are the same data this post priced DigitalOcean against.
The part that matters more here is what happens when you stop. You can turn on workspace backups and snapshot the whole box — filesystem, environment, installed packages at their versions, model files at their real paths — on a schedule you set or by hand, then restore it onto a different GPU, or a different provider, when you next deploy. The unit is the environment rather than the machine, which is what makes stopping cheap instead of destructive.
One boundary, stated plainly because the industry is loose about it: this covers the box you stopped. If a provider reclaims capacity out from under you, the only honest protection is a snapshot you already took — nothing brings back a machine you did not capture. That is a real limit and it applies to us too.
If you want to compare the specific numbers rather than the argument, cheapest cloud GPU providers has on-demand rates from eight providers' own pricing pages, and our comparison with Paperspace covers the persistence model side by side.
