On-demand GPUs

When a task needs a GPU, the agent rents one

When your laptop has no CUDA or the lab GPUs are queued, the scientific agent rents a GPU for the task and releases it when the run finishes. Usage is billed per second from your model balance. You can also connect your lab servers to the workspace over SSH.

Start researching

Billing

Billed per second · Released when done

You grant permission before the agent rents GPUs.

How it works

Rented per task, released when done

The agent picks a GPU type and domain image for the task and releases the instance when the task ends. To keep data for later, keep the data disk and continue next time.

Per-second billing
Domain images
Keep the data disk

One balance for models and compute

GPU costs and model usage come from the same balance, with no separate cloud account or payment method.

Single balance
Used after you grant permission
Itemized costs

Connect your lab servers

After you register your own servers in your account, the agent calls them through SSH aliases. Private keys are stored encrypted and the agent never sees them.

SSH aliases
Keys stay away from the agent
Available in every workspace

Start a GPU computation

Start researching

Explore next