An on-premises data warehouse runs on hardware the organization owns and operates, inside its own data center, rather than as a managed cloud service. For two decades this was the only way to do large-scale analytics, and a generation of warehousing practice - dimensional modeling, ETL, MPP query engines - grew up on these systems. They haven't disappeared: regulated industries, data-sovereignty requirements, and very predictable heavy workloads still keep plenty of warehouses on-prem.
The classic platforms
The on-prem warehouse market was defined by purpose-built appliances and massively-parallel systems. Teradata pioneered shared-nothing MPP and remained the enterprise standard for decades. Netezza (later IBM) and Oracle Exadata bundled tuned hardware and software into appliances. Greenplum and Vertica brought MPP analytics to commodity hardware. What they share is the architecture that still underpins cloud warehouses: data partitioned across many nodes, queries executed in parallel, and columnar storage for analytical scans.
Why some warehouses stay on-prem
The reasons are rarely about performance and usually about control. Data sovereignty and compliance can require that data never leaves a specific jurisdiction or facility. Predictable heavy workloads sometimes cost less on owned, fully-utilized hardware than on metered cloud compute. And existing investment - hardware, licenses, and years of tuned ETL - makes migration a multi-year project rather than a switch. The trade is the operational burden the cloud removes: capacity planning, patching, hardware refreshes, and sizing for worst-case load.
The shift toward the cloud - and the local revival
Most new warehousing has moved to cloud data warehouses, which separate storage from compute and remove the hardware burden. But the on-prem and local story has an interesting modern chapter: DuckDB brings warehouse-grade columnar, vectorized analytics to a single machine as an in-process engine - no server, no cluster - making "local analytics" genuinely fast for datasets that fit on one box. It's a different shape from a Teradata appliance, but it answers the same impulse: run the analytics where the data and the control already are.
Where this fits today
On-premises warehousing is now a deliberate choice rather than a default - taken for compliance, sovereignty, cost predictability at scale, or to extend existing investment - while greenfield projects largely start in the cloud or lakehouse. The architectural ideas the on-prem era produced, from MPP to columnar storage to dimensional modeling, are exactly what the cloud generation inherited and refined.