SnowPro Specialty: Gen AI Question 263
Single answerTracking model usage and consumptionA data platform team uses Snowflake Cortex AI functions across several internal applications. Finance wants a weekly report showing which teams are driving the highest LLM-related consumption so they can allocate costs more accurately. The team needs a solution that relies on Snowflake-native metadata rather than custom application logging. Which approach should the team use?
- A
Query account usage views that expose AI service consumption and join them with query history to attribute usage back to users, roles, warehouses, or applications over time.
- B
Review warehouse load history, because Cortex model usage is billed entirely through the virtual warehouse that executed the SQL statement.
- C
Inspect only ACCESS_HISTORY, because it directly records token counts and model-level AI billing details for every Cortex request.
- D
Enable Search Optimization Service on the application tables, because Snowflake records model token consumption only after search optimization is configured.
Show answer and explanation
Correct answer: A
Explanation
To track model usage and consumption in Snowflake, teams should use Snowflake-native metadata that exposes AI service consumption and combine it with operational metadata such as query history for attribution. This is the most practical approach for Finance and platform teams that need showback or chargeback reporting by team, user, role, or application. In practice, Snowflake usage/account usage views are the authoritative place to analyze billed consumption, while QUERY_HISTORY helps connect that consumption to specific workloads. Warehouse-centric views alone are insufficient because Cortex and other AI services are not tracked purely as warehouse compute. Governance views like ACCESS_HISTORY are valuable for auditing data access, but they are not the primary source for token or model consumption billing. This aligns with Snowflake best practices of using ACCOUNT_USAGE/organization billing metadata and query metadata for operational and financial reporting on AI workloads.
- A. Correct.
Correct. For Snowflake-native tracking of model usage and consumption, the right pattern is to use Snowflake-provided usage and billing metadata, including account usage views for AI service consumption, and correlate that with query metadata such as QUERY_HISTORY to understand who invoked Cortex functions and when. This supports chargeback/showback reporting without requiring custom logging in each application.
- B. Incorrect.
Incorrect. A common misconception is that all Snowflake SQL-related costs roll up to warehouse usage. However, Cortex AI functions are tracked as AI service consumption, not simply as warehouse runtime. Warehouse load history may help with overall SQL execution analysis, but it does not provide the authoritative source for LLM consumption attribution.
- C. Incorrect.
Incorrect. ACCESS_HISTORY is useful for governance and lineage-style analysis of what objects were accessed, but it is not the primary source for model token usage or AI billing metrics. Someone might choose this option because it sounds like the broadest activity log, but it does not directly replace usage/billing views for AI consumption reporting.
- D. Incorrect.
Incorrect. Search Optimization Service is unrelated to whether Cortex model usage is tracked. This distractor targets the misconception that special performance features must be enabled for observability. Snowflake records service consumption through its native metadata and billing systems independently of search optimization.