Docs · how to

Connect a GPU cloud

Add CoreWeave, Lambda, Nebius, Crusoe or RunPod capacity, with your own key or billed through denv.

GPU clouds often have the hardware a hyperscaler cannot give you this month. Connect one and its hosts join fleets like any other: same denvd, same releases, same endpoints.

Two ways to pay

Mode Who holds the account Who bills you for GPUs
Your key You The GPU cloud, under your contract
--billing denv denv denv, on one bill with everything else

Use your key when you already have a contract, reserved capacity or credits. Use --billing denv to start without opening another account.

Connect with your key

Create an API key in the provider's console with permission to create and delete instances, then:

denv provider connect coreweave --name cw --api-key $COREWEAVE_API_KEY
denv provider connect lambda    --name lambda --api-key $LAMBDA_API_KEY
denv provider connect nebius    --name nebius --api-key $NEBIUS_API_KEY
denv provider connect crusoe    --name crusoe --api-key $CRUSOE_API_KEY
denv provider connect runpod    --name runpod --api-key $RUNPOD_API_KEY

denv stores the key encrypted and uses it only to create, list and delete the instances of your fleets. See Security.

Connect through denv

denv provider connect nebius --name nebius --billing denv

No key needed. Instances are created in denv's account for your organization only; hosts are never shared with another organization.

Find capacity

denv offers --hardware h100 --count 8 --mode spot
denv offers --hardware gb200-nvl72 --provider cw

PROVIDER  REGION    SHAPE              HARDWARE      DOMAIN  MODE       PRICE/H  QUOTA
runpod    us-tx-3   8x H100 SXM        h100          8       spot       …        6
lambda    us-tx-1   8x H100            h100          8       on-demand  …        4
nebius    eu-north1 8x H200            h200          8       on-demand  …        8

DOMAIN is the number of GPUs joined by NVLink in one scale-up domain. It decides which topologies fit; see Hardware.

Create a fleet

denv fleet create h100-spot --provider runpod --region us-tx-3 \
  --shape "8x H100 SXM" --mode spot --min 0 --max 8

Hosts boot with denvd installed by the provider's startup script, then join on their own.

Spot capacity

Spot hosts can be taken back by the provider. When that happens:

  1. denvd reports the notice, if the provider sends one.
  2. The router stops sending new requests to the host's replicas.
  3. In-flight requests finish or fail over to other replicas in the stage.
  4. The provisioner launches a replacement within the fleet's max.

Use spot for burst capacity behind a reserved or owned fleet, not as a stage's only fleet. See Route across fleets.

Check

denv provider list
denv host list --fleet h100-spot

Next