Google Caps BigQuery AI Use at 200 Billion Daily Tokens per Project

Q: Why might BigQuery's built-in token counter dangerously underestimate your actual AI costs before execution?

BigQuery’s built-in counter can dangerously understate real AI cost because it only estimates input text tokens, while actual billing and quota use include several other token sources that can be substantial. Google’s AI.COUNT_TOKENS does not count output tokens at all, so any long model response is missing from the estimate. More importantly, BigQuery’s generative AI quotas also count reasoning or “thinking” tokens as output. Those internal reasoning tokens can materially raise usage, especially on more complex prompts or models configured for deeper thinking. The function also omits added instructions and system-side prompt content that BigQuery may send along with the user’s visible input. That means the model may process more text than the pre-check suggests. It further ignores non-text and multimodal inputs such as images, audio, video, and PDFs, even though BigQuery AI functions can analyze those formats and they can increase consumption. This creates a practical risk: a workload that looks safe based on AI.COUNT_TOKENS may still run into much higher real token use, higher Vertex AI charges, or even quota exhaustion during execution. Google itself warns that the function can underestimate real consumption and recommends testing a small sample job to measure actual usage before setting production quotas. In short, the token counter is a rough sizing tool, not a reliable cost forecast, because it sees only part of what the model will actually process and generate.

3 articles · Updated · InfoWorld · Oct 15

Google has generally released daily token quotas for BigQuery generative AI functions, letting enterprises cap both input and output usage by project and by individual user.
Default limits are 200 billion input tokens and 20 billion output tokens per project per day, plus 150 billion input and 15 billion output tokens per user; embedding functions are excluded because they run on cheaper models.
BigQuery rejects queries with a QuotaExceeded error when limits are already exhausted, and if quotas run out mid-query it can return partial AI.IF results or leave AI.GENERATE rows without output.
Google also added AI.COUNT_TOKENS to estimate usage before execution, but it covers only input tokens and can understate actual consumption because it excludes output, reasoning tokens and non-text inputs.
Analysts said the controls improve cost predictability and reduce runaway-bill risk, but warned enterprises still need FinOps, team-level budgeting and separate production versus experimental limits to avoid reliability hits.