DigitalOcean¶

DigitalOcean rents GPU Droplets: full VMs with NVIDIA (H100, L40S, RTX 4000/6000 Ada) or AMD (MI300X) GPUs. Which GPUs are offered depends on the region, and GPU Droplets may first need to be enabled for your account. gpuprov boots DigitalOcean's AI/ML-ready GPU images, which are Ubuntu with the GPU drivers preinstalled.

export DIGITALOCEAN_ACCESS_TOKEN=...      # or [digitalocean] api_key in config.toml
gpuprov provision --cloud digitalocean --name dev --gpus L40S:1
gpuprov provision --cloud digitalocean --name big --gpus H100:8 --region nyc2

Create a token with read and write scope at https://cloud.digitalocean.com/account/api/tokens. $DIGITALOCEAN_TOKEN is also accepted.

Acceptance testing pending

This adapter is covered by unit, mock-API and vendor-stub tests. It has not yet passed the end-to-end acceptance run against a live account (allocate → SSH → k3s/GPU ready → terminate → nothing billable left), so treat it as preview.

Planning: live sizes¶

One call, GET /v2/sizes, returns every Droplet size with its hourly price, its GPU (gpu_info), and the regions it can be created in. Each available size × region becomes one offer:

instance_type = the size slug, e.g. gpu-h100x1-80gb or gpu-l40sx1-48gb
region = the region slug (tor1, nyc2, atl1, …); there are no zones
ordered by hourly price, cheapest first

Without --region, only North American regions (New York, San Francisco, Toronto, Atlanta, Richmond) are considered. Pass --region explicitly for any other region, e.g. --region ams3. --zone also pins a region, because DigitalOcean has no zones.

--gpus matches the GPU model by whole tokens: L40 does not match L40S, RTX4000 and RTX-4000-Ada both match nvidia_rtx_4000_ada, MI300X matches amd_mi300x, and the count must equal the size's GPU count (H100:1 vs H100:8). --instance-type takes an exact size slug, including CPU sizes such as s-2vcpu-4gb.

Availability is re-checked right before launching, so a stale offer fails before anything is registered or created. A create that DigitalOcean rejects because the size is not available in the region fails over to the next offer.

What gets created¶

Resource	Notes
Droplet	named `gpuprov-<name>`, tagged `gpuprov` and `gpuprov-vm-<name>`
tag `gpuprov-vm-<name>`	the VM's identity; created before the firewall so the firewall can reference it, deleted on teardown
cloud firewall	always one per VM, named `gpuprov-<name>` and applied to the VM's tag; allows inbound SSH, ICMP and any `--port`s from anywhere (IPv4 and IPv6), and all outbound traffic; deleted on teardown
SSH key	your gpuprov public key, registered once and reused; never removed

The firewall is always created because a cloud firewall drops any inbound traffic it doesn't list, and gpuprov needs to know exactly which ports are open.

--image takes an image slug (e.g. ubuntu-24-04-x64) or a numeric image id. Without it, gpuprov follows DigitalOcean's GPU guidance:

Size	Image
NVIDIA, 1 GPU (any model)	`gpu-h100x1-base`
NVIDIA, multiple GPUs	`gpu-h100x8-base`
AMD	`gpu-amd-base`
CPU	`ubuntu-24-04-x64`

--disk has no effect. A Droplet's disk is fixed by its size, and for GPU sizes it includes a large local scratch disk.

SSH¶

The login is root on port 22. DigitalOcean installs the SSH keys passed at creation for root on every image, so the login doesn't depend on cloud-init. gpuprov's cloud-init user data still creates the usual ubuntu user with your key.

Kubernetes¶

Droplets accept cloud-init user data, so the k3s installer runs while the Droplet boots and gpuprov verifies it over SSH. On NVIDIA images the installer sees nvidia-smi and adds the NVIDIA runtime and device plugin. The cloud firewall attaches per VM, so clusters use direct API access: ports 6443, 80 and 443 are opened in the VM's firewall.

Billing and health¶

DigitalOcean bills a Droplet for as long as it exists, including while it is powered off. gpuprov health reports a powered-off Droplet as UNHEALTHY and says it is still billing. Run gpuprov down to stop the charges. A Droplet that is locked while an action runs (resize, snapshot) is reported as information.

Limits¶

Feature	Why
`--spot`	DigitalOcean offers on-demand Droplets only

API rate limits are 5,000 requests per hour and 250 per minute. When gpuprov is throttled (429), it retries the step.

Teardown¶

gpuprov down deletes the Droplet and waits until DigitalOcean no longer lists it. It then deletes the VM's cloud firewall and its gpuprov-vm-<name> tag.