Donald Zhong
Back to writing
AI DailyOct 15, 20262 min read

AI Daily - 2026-10-15: Google makes BigQuery AI spend predictable

Google's BigQuery now exposes daily input and output token quotas for SQL-native Gemini inference, turning AI spend into an admin-controlled product setting.

GoogleInfraModels

Why it matters

What changed On October 14, Google Cloud announced general availability of token based cost management for BigQuery's generative AI functions.

What changed

On October 14, Google Cloud announced general availability of token-based cost management for BigQuery's generative AI functions. BigQuery can already call Gemini models directly from SQL; the new control layer adds four daily quotas for input and output tokens at both the project and per-user level. The quotas are global across regions and can be adjusted in Google Cloud's IAM and quota controls without opening a support case.

The controls cover AI.GENERATE, AI.GENERATE_TABLE, AI.GENERATE_TEXT, AI.CLASSIFY, AI.IF, AI.SCORE, and AI.AGG. Embedding functions are excluded because they use a different billing model.

Read the Google Cloud announcement and the official token-quota documentation.

Why it matters

The important shift is from "AI in SQL" to "AI in SQL with an enforceable budget boundary." BigQuery checks available quota before a query starts and continues enforcing it while the query runs. When a limit is exhausted, new model calls fail with QuotaExceeded; for functions that support max_error_ratio, a workload can be designed to tolerate a controlled amount of partial failure.

That matters because warehouse queries can fan out in ways that are hard to price by inspection. A join, an expanded result set, or a reasoning-heavy model can turn a seemingly small query into a large number of token-consuming calls. Global quotas also prevent usage from becoming an invisible regional cost problem.

Builder takeaway

Treat token budgets as part of the data product, not as an after-the-fact billing alert:

  1. Set project and per-user ceilings in IAM & Admin → Quotas & System Limits.
  2. Use AI.COUNT_TOKENS to estimate input size, and materialize the rows you intend to process before invoking a generative function.
  3. Choose a lower thinking budget or thinking level when a task does not need deep reasoning. The AI.GENERATE reference shows how to control that trade-off.
  4. Monitor the resulting Vertex AI charges by BigQuery job so the quota is one layer in a larger cost and reliability strategy.

Bottom line

Google is making warehouse-native AI more operationally legible. The feature is not another model launch; it is a product boundary that lets teams decide who can spend how much AI capacity, before a batch query becomes a surprise bill. For builders, that is the difference between a compelling SQL demo and a service that can be trusted in production.