GCP Cloud TPU cost: per chip-hour, with Spot and commitments

Quick answer

Cloud TPU bills per chip-hour by TPU version and pod configuration, with on-demand, Spot (preemptible) (a large discount for interruptible work), and committed-use pricing for steady usage. Like GPUs, TPUs are expensive per hour, so the levers are Spot for interruptible training, shutting down idle TPUs immediately, and keeping utilization high.

Cloud TPU provides Google's custom accelerators for machine-learning workloads. Like GPUs, TPUs cost a lot per hour, so the same discipline applies: an idle TPU is expensive waste, and full-price on-demand for interruptible training leaves large savings on the table.

Per chip-hour, by version

TPUs are billed per chip-hour, with rates depending on the TPU version (newer generations cost more per chip but deliver more performance) and whether you run a single device or a pod slice. A pod of many chips multiplies the hourly cost, so training runtime and pod size together set the bill.

On-demand, Spot, and commitments

Pricing	Best for
On-demand	Interactive, latency-sensitive, short runs
Spot (preemptible)	Fault-tolerant, checkpointed training
Committed use	Steady, predictable, long-term usage

Spot (preemptible) TPUs offer a large discount for interruptible workloads, which most training is if it checkpoints regularly. Committed-use discounts apply to steady, predictable usage over a term. On-demand is for short, interactive, or latency-sensitive work that cannot tolerate interruption.

The same accelerator discipline as GPUs

Because TPUs are expensive per hour, shut down idle TPUs the moment a job finishes, use Spot for checkpointed training, keep utilization high with efficient data pipelines that do not starve the accelerator, and reserve committed-use for genuinely steady workloads. This is the same logic as GPU cost optimization, applied to TPUs.

TPU vs GPU

TPUs can be more cost-effective than GPUs for large-scale training and inference of models that map well to them, particularly large matrix workloads, while GPUs are more flexible across frameworks and model types. The cost comparison depends on your model and how well it utilizes the accelerator, so benchmark cost per training step or per inference, not just the hourly rate.

FAQ

How is GCP Cloud TPU priced?

Per chip-hour by TPU version and pod configuration, with on-demand, Spot (preemptible), and committed-use pricing. Newer TPU generations cost more per chip but deliver more performance, and a pod of many chips multiplies the hourly cost. Runtime and pod size set the bill.

How do I reduce Cloud TPU cost?

Use Spot (preemptible) TPUs for interruptible, checkpointed training (a large discount), shut down idle TPUs the moment a job finishes, keep utilization high with efficient data pipelines that do not starve the accelerator, and use committed-use discounts for genuinely steady workloads.

Are Spot TPUs worth it?

For fault-tolerant training, yes, and it is a large saving. Most training checkpoints regularly, so a preempted TPU just resumes from the last checkpoint. Reserve on-demand TPUs for short, interactive, or latency-sensitive work that cannot tolerate interruption.

Is a TPU cheaper than a GPU?

It depends on the model. TPUs can be more cost-effective for large-scale training and inference of models that map well to them, particularly large matrix workloads, while GPUs are more flexible. Benchmark cost per training step or per inference, not just the hourly rate, to compare.

Why are idle TPUs expensive?

Because TPUs cost a lot per chip-hour, an idle TPU or a pod left running after a job burns money at that high rate for zero work. Shutting down idle TPUs immediately is the highest-impact saving, the same as with GPUs.

Does C3X estimate Cloud TPU cost?

Cloud TPU cost is driven by chip-hours, version, and pod size, which depend on the training workload. C3X prices the surrounding infrastructure, and you model TPU usage to estimate the accelerator cost.

What to do next

Estimate the infrastructure around your ML training before you run it. C3X reads your Terraform and prices your resources against a live catalog. Start with the quickstart.