Why this matters if you've tried renting GPUs before
The market for GPU compute is strange right now. The cheap providers want identity verification, a business entity, or a monthly commitment. The ones that don't ask are often reselling consumer cards from someone's basement with no guarantee the machine is there tomorrow.
We're a thin layer over real datacenter hardware: these are DigitalOcean GPU droplets, deployed from our account, and you pay us in crypto instead of giving a card to a US company. That's the whole trade. You get datacenter H100s with an email address, and we handle the part where somebody has to have a billing relationship with a cloud.
What you get on the machine
Every GPU server boots the NVIDIA AI/ML Ready image: Ubuntu with drivers and CUDA already installed. You SSH in as root and start working — no driver hunting, no kernel module compilation at 2 AM.
That also means the OS picker doesn't apply to GPU plans. GPU droplets only boot images built for them, so we don't offer a choice we can't honour. Pick the plan, pick the region, get a working CUDA box.
Reasonable first things to run: a local LLM behind an API, Stable Diffusion or a video model, fine-tuning a small model on your own data, batch inference for a pipeline that would cost more per-token on a hosted API. An H100 with 80 GB of VRAM runs most open-weights models that matter today without quantisation gymnastics.
The one rule that's different
Read this bit before you deploy, because it's the one place GPU servers behave unlike our others.
On a normal NoctHost server, running out of balance is recoverable: we snapshot the disk, delete the instance so you stop paying, and keep the snapshot for seven days. Top up and your server comes back with its data.
**GPU servers are deleted without a snapshot.** When your balance hits zero, the machine and everything on it is gone.
The reason is arithmetic rather than policy. A GPU server has a 500–720 GB disk, and holding that in snapshot storage costs roughly $30–43/month — often more than the server earned while it was running. We'd be paying to store data indefinitely for servers nobody came back to. Rather than build that cost into everyone's hourly price, we don't keep the snapshot.
So treat a GPU box the way you'd treat a spot instance: push your checkpoints and results somewhere else as you go, and keep an eye on your balance. At $4.60/hr a $50 balance runs about ten hours; the dashboard shows the countdown, and we email you before it runs out.
Deploying one
Exactly like any other server: **Deploy server** → provider **DigitalOcean** → pick the GPU plan → region → set a root password. It's live in a couple of minutes, and you SSH in as root.
Over the API it's the same call you already use, with the plan id:
curl -X POST https://api.nocthost.com/servers \
-H "Authorization: Bearer $NOCTHOST_KEY" \
-H "content-type: application/json" \
-d '{"label":"trainer","planId":"gpu-h100x1-80gb","regionId":"ams","password":"..."}'No image field needed — GPU plans always get the CUDA image.
Regions, and why there are only two
H100s run in Amsterdam and Toronto; the RTX 6000 Ada is Toronto only. That's not a product decision, it's capacity: those are the regions where our provider has these cards available to us. As that changes, the region list in the dashboard changes with it — it's generated from what can actually be deployed, so you'll never pick a combination that fails at creation.
Same for the 8×GPU machines: they exist, but not in a region we serve, so we don't list them. If you need multi-GPU, tell support what and where — that's exactly the kind of demand signal that decides what we add next.