Billing for compute hours

Compute hours represent the available computing resources for training AI models.

The available compute quota depends on the license type used and any optionally assigned Resource Packs. At the start of a new subscription period, the compute quota included in the license is replenished. Unused compute hours expire at the end of the respective runtime. Additional compute hours can be added using resource packs.

If the availability of compute quota is no longer sufficient, no further training sessions can be started. In this case, you must either wait until the next subscription period or assign an additional resource pack.

You can find the current value of your compute quota in the top-right corner of the header; it is displayed in hours and minutes.

Reserving compute hours

When a model training session starts, the estimated training time is first calculated based on the training configuration and the data set used.

If the estimated training time is less than or equal to the available compute quota, the training is started. In this process, the estimated compute time is initially reserved. This mechanism is similar to temporarily setting aside a budget, such as when making a credit card payment.

Once the training session is complete, you will be billed only for the actual training time used. Unused reserved compute hours are automatically released.

Insufficient compute quota

If the estimated training time exceeds the currently available compute quota, the training cannot be started.

In this case, additional compute quota must be allocated.

Billing date

A training session is always billed based on the compute quota that was available at the time the training started. The reserved compute hours remain tied to the original subscription period for the entire duration of the training.

Example:

If training is started on the last day of a subscription period and does not end until the following subscription period, the entire training duration will still be charged to the original subscription term. The reason for this is that the compute quota was already reserved when training started.