What You Get
Every GPU instance includes:- Root access — Full
sudoprivileges in an isolated Linux container. - Dedicated GPUs — One or more GPUs exclusively allocated to your instance.
- Configurable specs — Choose your CPU cores, memory, and disk size.
- Public IP address — Direct SSH access and the ability to expose services.
- Pre-installed tools — CUDA drivers, Python, pip, and common ML libraries come ready to use.
Available GPUs
Prices vary by GPU, availability and configuration. See the live GPU pricing page or the Deploy page for current rates.
Datacenter and community hosts
Every machine says who runs it, ashost_tier in the API and a chip in the dashboard:
- Datacenter (
datacenter): a machine in a datacenter. This is what you get unless you ask for more. - Community host (
community): run by an independent host, not a datacenter. The host operates the hardware and can access the machine. Not available on reserved terms.
include_community=true to GET /api/v1/instances/types and "include_community": true when you create the instance, or set include_community on a job or a schedule’s spec. Without it, every list and launch uses datacenter machines only, and the GPU pricing pages always show datacenter prices.
On a community host:
- prepaid and reserved terms are never offered;
- a launch that can’t start is retried only on another community host, and a datacenter launch never moves to one;
- Sensitive secrets are never sent: a launch from an environment that has any is refused, so launch it on a datacenter machine or from another environment.
Spot machines
A spot machine costs less, but its host may take it back after a few minutes of notice (2 to 60, shown asreclaim_notice_minutes and on hover of its Spot chip). The notice comes as the instance.reclaim_notice event and an email; at the deadline the instance ends. A job on a spot machine runs again on another machine from its volume, up to 5 times, without using its max_retries.
Spot machines are off by default. Turn on Include spot machines in the machine picker, pass include_spot=true to GET /api/v1/instances/types and "include_spot": true when you create the instance, or set include_spot on a job or a schedule’s spec. Each one is marked capacity_tier: "spot". A launch that can’t start on a spot machine is retried only on another spot machine, and any other launch never moves to one.
Some machines are offered on prepaid terms only (term_only: true). They launch only with a term: the machine picker lists them once a term is picked, and the API needs term_months.
Minimum charge
A few machines, such as whole bare-metal servers, are billed a minimum however long they run. These carryminimum_billed_minutes in GET /api/v1/instances/types and on the instance, and the machine picker says it before you start, for example “Minimum charge: 8 hours ($310.40 at $38.80/h)”. The minimum is counted from when the machine is handed over to you, and running longer bills per second as usual. A launch that fails before the machine is handed over is not charged, and you can still cancel a launch while it is being set up.
Once the machine is running, you cannot end it before its minimum is up: it is paid for, so it is yours until then. The dashboard shows Paid until HH:MM in your own time zone, and paid_until (ISO 8601) is on the instance in the API. Terminating it before then does not tear it down. The instance keeps running, stays reachable as before (web terminal, SSH, links), and ends by itself when the minimum is up; the dashboard then shows Ends at HH:MM. Reconnect to it instead of starting a new one. DELETE /api/v1/instances/{id} answers 202 with ended: false, code: "PAID_WINDOW", paid_until and ends_at instead of 204, and asking again changes nothing. Its max runtime is set to exactly the end of the minimum, whether yours was shorter or longer, so it neither ends early nor runs past it. No other setting ends it early either: a shorter max runtime you set later is raised to the end of the minimum, and idle stop waits until then (a machine that has already been idle may stop then) and works as usual after it. Ten minutes before it ends you get an instance.runtime_ending alert that says your paid time is ending. After the minimum has passed, terminating works as usual, and running past it bills per second.
Where your machine is
GET /api/v1/instances and GET /api/v1/instances/{id} show host_country: the country in which the machine’s address is registered, from public registry data, as an ISO 3166 code such as DE. It is not an exact location: an address range can be registered in one country and used in another. It is null until the address is known, and when the registry data cannot say.
To keep a launch in the European Economic Area (the EU, Iceland, Liechtenstein and Norway), pass "eea_only": true when you create the instance, or set eea_only on a job or a schedule’s spec. Only types whose region positively names an EEA country are offered. The first check of the machine’s address runs after the machine is rented: an address registered anywhere else, or that cannot be verified, is released before it runs and is never charged, and the launch moves to the next option. When none is left, the instance fails with “No machine in the European Economic Area could be verified for this launch; nothing was charged”. It is off by default and not supported on clusters.
What is withheld, exactly. The secrets your machine pulls from Runcrate, and its other machine calls, are served only to an address registered in the EEA: a machine on any other address, or one whose address cannot be verified, is answered as if it were unknown. The launch script itself (your startup commands) is different: on most machine types it is sent when the machine is created, so it is already on the host when the check runs. On the others it is sent only after the address passes. That is why eea_only is a filter on where machines are rented, and not a guarantee of where your data is processed.
Volumes. A launch script that attaches a volume carries the volume’s keys. With eea_only, a launch with a volume is therefore offered only on machine types that receive the script after the address is checked. If the machine you picked, or every machine left, receives it at creation, the launch is refused with a 400, “A volume cannot be attached to an eea_only launch on this machine type yet”, and nothing is rented. Launch without the volume, or without eea_only.
eea_only reads where an address is registered, not where the machine stands. A range registered in one country can be used in another, so it is strong evidence of the country, not a guarantee of physical location.
Instance Lifecycle
Every instance moves through these states:Billing
Instances are billed by the second on most GPUs, and by the minute on a few, based on the GPU type and configuration you select.- Billing starts when the instance reaches the
deployedstate. - Billing stops when the instance is
terminated. - Charges are deducted from your credit balance automatically.
Next Steps
Deploy an Instance
Step-by-step deployment guide.
Connect via SSH
SSH, VS Code, and port forwarding.