Category
ELT tools
ELT (extract-load-transform) lands raw rows in the warehouse first and transforms with SQL/dbt afterwards. It has won the mid-market because warehouse compute is cheap relative to engineering time. Realistic 2026 finalists for a 50M-MAR workload are Fivetran, Airbyte, Stitch, and Hevo, with published $/M-row rates from roughly $14 (Estuary, streaming) to $48 (Fivetran Standard).source - pulled 2026-06-21
What ELT actually is
ELT inverts the classical ETL order: rows are extracted from a source system, loaded into the warehouse as raw or near-raw tables, and transformed in-warehouse using SQL or dbt. The cost of warehouse compute fell faster than the cost of engineer-built transformation, so the work moves to where the cheapest compute lives. ELT pairs naturally with a transform layer (dbt Core or dbt Cloud) and an orchestrator (Dagster, Prefect, Airflow) if pipelines need branching or retries beyond the vendor scheduler.
When to pick ELT over the alternatives
- › Destination is a cloud warehouse. If it is a lakehouse or stream consumer, look at streaming or lakehouse ingestion.
- › Daily / hourly / 15-minute cadence is acceptable. Sub-minute requires streaming-native.
- › Source mix is mostly SaaS APIs and managed databases.
- › Transformation logic can live in SQL.
Realistic finalist set
Where ELT stops being the right answer
Streaming workloads with sub-minute latency requirements should use streaming-native tools. Workloads where every row must be transformed before landing belong to traditional ETL (less common in 2026). Operational workflows that push data back into SaaS systems are reverse-ETL, not ELT.
Written by Oliver Wakefield-Smith, Founder of Digital Signet. Independent reference, no vendor sponsorship.
Sources logged at /sources. Pricing-change history at /changelog. Last reviewed 2026-06-21.