The architecture is genuinely different from ClickHouse and Postgres-based solutions. Object storage is first-class (not a tiering afterthought) so costs stay predictable as telemetry volume grows. Native OTLP ingestion means we went from SDK to queryable SQL with no queue, no workers, no transformation pipeline. For AI observability specifically, the Gen-AI semantic conventions land directly as columns, which makes building cost and latency dashboards straightforward.