# Cloud TPU

> Cloud TPU nodes bill per chip-hour from provisioning until deletion, whether or not a training job is running, and rank among the most expensive per-hour resources on GCP. ZopNight discovers TPU nodes through Cloud Asset Inventory, watches utilization in Cloud Monitoring, and flags idle or long-running TPUs in its recommendations.

Source: https://zop.dev/integrations/gcp/cloud-tpu
Updated: 2026-08-19

---

Cloud TPUs are Google's custom accelerators for training and serving large machine learning models. TPU nodes are among the most expensive resources per hour on GCP, so idle TPUs are a critical cost signal.

## Chip-hours are the TPU meter

A TPU node bills per chip-hour from the moment it is provisioned until it is deleted. The meter runs whether a training job is executing or the node sits idle between experiments; accelerator generation and topology set the rate, and multi-host slices multiply it. Committed use discounts on TPUs bill for the commitment regardless of usage, and a spot node trades a lower rate for the risk of interruption. Because the hourly figure dwarfs almost everything else in a typical project, a single forgotten node can outweigh the rest of the bill on its own.

## ZopNight's view of TPU nodes

ZopDev discovers TPU nodes via Cloud Asset Inventory, watches utilization through Cloud Monitoring, and flags idle or long-running TPUs through its recommendation rules. TPU nodes are not schedulable in ZopNight, because there is no stop verb that preserves a node, so the leverage is fast detection: surfacing nodes whose utilization has collapsed, and long-lived nodes that look like they were meant to be temporary.

## How TPU spend goes wrong

Three patterns dominate. A researcher provisions a node for an experiment, and the experiment ends before the node does. A slice gets sized for a model that never ships, then kept around "in case". And reservations are purchased on utilization assumptions nobody verified. The standing advice is to verify utilization before committing to long reservations, because a committed TPU bills whether or not it ever trains anything.

## Reaching TPUs in the console

Google Cloud console → Compute Engine → TPUs lists every node with its accelerator type, topology, and status. Cross-check that list against Cloud Monitoring utilization before assuming a running node is earning its rate.