Deployments

A deployment turns contracts into machines: gpus of a BASIS contract paired with gpus of the GPU contract for the same GPU and month. While it is open both sets of contracts are locked: they can’t be sold. Ops claims the request, and the machines are provisioned on Northflank, our hosting partner: a JupyterLab notebook on CUDA by default, or a public image you name, with the ports you ask to reach. Their addresses appear on the deployment once it is running. Where Northflank cannot run the configuration, ops provisions by hand and sends access by email.

Only the current month can be deployed. Later months trade but wait for their month.

Request a deployment

POST /api/v1/compute/deployments · scope trade

Field	Type	Description
basis_symbol	string	The provider’s contract, e.g. BBA-10-26.
gpu_symbol	string	The GPU contract for the same GPU and month, e.g. H100SXM-10-26.
gpus	integer	How many GPUs. You need this many free contracts on both books.
image	string	Optional. A public container image to run instead of the default notebook.
ports	array	Optional. 1 to 8 of { "port", "protocol" } (http or tcp) to reach; the notebook’s port 8888 when absent. HTTP ports get a public address; TCP ports are reached through the Northflank CLI.
env	array	Optional. Up to 32 { "name", "value" } environment variables for the image. They are passed to the machines and not stored.

REST

curl -X POST "$CAPITALAI_API_URL/api/v1/compute/deployments" \
  -H "Authorization: Bearer $CAPITALAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"basis_symbol": "BBA-10-26", "gpu_symbol": "H100SXM-10-26", "gpus": 8}'

200 with the deployment:

{
  "id": "dep_7f3a",
  "basis_symbol": "BBA-10-26",
  "gpu_symbol": "H100SXM-10-26",
  "gpus": 8,
  "status": "requested",
  "ends_at": "2026-11-01T00:00:00+00:00",
  "created_at": "2026-10-21T18:40:12+00:00",
  "activated_at": null,
  "ended_at": null,
  "ops_note": null,
  "parent_id": null,
  "url": null,
  "image": "quay.io/jupyter/pytorch-notebook:cuda12-latest",
  "ports": [{ "port": 8888, "protocol": "http" }],
  "northflank": null
}

url is the first HTTP port’s address and each port gains its url once the deployment is active. northflank names the project and service for the Northflank CLI once you are a member of the project.

Refusals are 400 with a code and the field it concerns:

code	Field	Meaning
unknown_basis	basis_symbol	No such BASIS contract.
month_over	basis_symbol	The month has ended.
not_current_month	basis_symbol	Only the current month deploys.
unknown_gpu_contract	gpu_symbol	No such GPU contract.
gpu_mismatch	gpu_symbol	Not the same GPU and month as the basis.
invalid_gpus	gpus	Fewer than one.
not_whole_contracts	gpus	The books do not deliver whole GPUs for that count.
basis_not_free, gpu_not_free	gpus	You don’t hold that many free contracts on that book. Cancel a resting sell, or buy more.

503 when the exchange is unreachable.

Status

status	Meaning	Contracts
requested	You asked; ops hasn’t claimed it. You can still cancel it.	locked
provisioning	Ops is setting up the machines.	locked
active	The GPUs are running and their addresses are set. activated_at is set.	locked
ending	You stopped it, or the month ended; ops is tearing it down.	locked until ended
ended	Done: cancelled before a claim, torn down, or the month ended unclaimed. ended_at is set.	released

Ending a deployment releases its contracts; it does not consume them. They stay in your account and can be sold or deployed again while their month lasts. Deployments that are ending, or whose month is over, are torn down by a job that runs every ten minutes.

ops_note carries anything ops wrote for you, such as how access was sent.

Stop a deployment

POST /api/v1/compute/deployments/{id}/end · scope trade

Stops all of the deployment’s GPUs, or gpus of them.

Field	Type	Description
gpus	integer	Optional. How many to stop, from 1 to all. Absent: all.

REST

curl -X POST "$CAPITALAI_API_URL/api/v1/compute/deployments/dep_7f3a/end" \
  -H "Authorization: Bearer $CAPITALAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"gpus": 4}'

200 with the deployment of the GPUs stopped. Stopping some of them splits those GPUs into a deployment of their own, with parent_id pointing at the one that keeps the rest, so ops tears down exactly those machines.

• A requested deployment ends at once (ended): nothing was provisioned.
• A claimed one goes to ending until ops has torn it down.

Status	Meaning
400 invalid_gpus	Fewer than one, or more than the deployment has.
404	No such deployment on your account.
409	The deployment has already ended.

List your deployments

GET /api/v1/compute/deployments is what the Compute page’s Active Capacity table reads. It is session-only today and is coming to the API.