Skip to content
FinOps workspace
COST LEDGER / DECISION GUIDE

BQML pricing: BigQuery ML training and inference costs

Published: 10 November 2026

Build a BigQuery ML estimate from the model, operation and billing route, then reconcile the complete workload cost.

Review my BigQuery ML costs
Cloud infrastructure illustration in the mint and dark green Cost Ledger palette
Trace training, inference and retained data to the same model workload.

For shared analytics projects, use the BigQuery query and capacity control checklist before enabling restrictions that could interrupt required model workflows.

For ordinary SQL scans around a model workflow, the 12 GB BigQuery query cost calculation separates on-demand scan pricing from capacity and the applicable model operation charges.

Choose the BQML pricing branch before calculating

BQML pricing is not one universal rate. The official BigQuery ML pricing section differentiates model operations, on-demand billing and capacity. Begin with the model type, dataset location, operation and billing arrangement. Copy the applicable regional tariff and date into your estimate instead of reusing the rate from an unrelated SQL query.

Workload branchRecord for the estimateCheck separately
Built-in trainingModel type, training job and billing modelInput preparation and repeated runs
BigQuery ML external model trainingOn-demand preprocessing and passed-through training, or assigned capacityApplicable billing model and duplicate cost allocation
Evaluation and predictionOperation, workload size and execution scheduleWhether inference calls a remote endpoint
Remote inferenceBigQuery work and endpoint service chargesEndpoint model, usage unit and service invoice lines
Stored data and modelsDatasets, retained artifacts and locationApplicable storage charges outside execution

The BigQuery ML model overview distinguishes built-in models, externally trained models, imported models and remote models. Identify the actual implementation with the ML owner. A familiar business label such as forecasting does not specify its billing route. Use the BigQuery cost review for the wider query and storage boundary.

Separate on-demand execution from reserved capacity

Google documents BigQuery ML capacity support in Enterprise and Enterprise Plus. Creating built-in models uses QUERY reservation assignments. Training BigQuery ML external models, such as DNN or boosted trees, uses ML_EXTERNAL assignments under capacity pricing. This does not describe supervised tuning of a remote LLM. For on-demand training, apply the model-specific pricing rule; supported hyperparameter tuning is billed across all executed trials. Check the ML editions and on-demand pricing reference for the selected model.

  • Record the project and location that actually run the job.
  • Identify the model family and operation in the job inventory.
  • Ask the platform owner which reservation and assignment apply.
  • Include the planned retraining and tuning schedule in the estimate.

Reconcile capacity ownership before assigning a model cost. Ask finance whether the review measures allocated capacity cost, incremental expenditure or an alternative purchase decision. Keep those calculations visible as separate views. The GCP commitment review helps organize the broader commitment inventory; it does not substitute for checking BigQuery reservation rules.

Keep remote endpoint costs in the same review

For remote model inference, Google describes BigQuery processing plus charges from the remote service. Its pricing reference identifies billing labels for tracing those endpoint charges. An estimate based only on BigQuery analysis lines can therefore miss part of the workload. Use the remote model pricing guidance to establish both sides of the bill.

  1. List the jobs and model endpoints included in the use case.
  2. Collect the billing lines for each service and keep their usage units.
  3. Agree how job references connect related charges without counting them twice.
  4. Compare the total with the accepted forecast and explain unmatched items.

For a generative AI workflow, ask the application owner to document the expected request mix and any retry behavior. Keep measured usage separate from the planned scenario. The LLM cost optimization guide helps frame the application review. Include the GCP storage boundary where training inputs or outputs remain outside the ML job itself.

Build an estimate the team can reproduce

Prepare one worksheet row per operation and charging component. Record the source tariff, billing unit, quantity assumption, execution frequency and owner. Add a separate row for retained data. For capacity, document the allocation basis and the period measured. Show uncertainties such as the number of tuning trials or an unsettled inference schedule instead of silently folding them into a total.

The dry-run limitation is described in the official BQML pricing notes. After a pilot, update the worksheet with measured quantities and retain the original assumptions. Keep workload changes separate from pricing changes. A smaller training dataset or different model can change the result even when the rate remains unchanged.

Scope the review under the GCP cost optimization plan. Request editable calculations, cost allocation rules and an agreed review of the next billing period. The proposal should state whether it includes query changes, model redesign, reservation configuration or analysis only. Assign each implementation decision to a named internal owner.

Questions about BQML pricing

Is BQML charged at the normal SQL query rate?

Not for every operation. On-demand model training has model-specific rules. Evaluation, inspection, prediction and remote endpoint use need their applicable pricing branch checked separately.

Does reserved BigQuery capacity cover every ML charge?

Check the edition and reservation assignment for the operation. Remote endpoint services and retained data must still be reconciled separately where applicable.

Can a dry run establish the full training cost?

Not for every model type. Google documents cases where processed bytes cannot be calculated until training completes. Validate the estimate with job and billing evidence from an approved pilot.

Get my FinOps quote