Cost
Why cloud warehouses get expensive as soon as they succeed
A cloud warehouse does not get expensive because it runs badly. It gets expensive because it runs well and people start using it.
Consumption billing is a fair model for variable load. For data work it has a property you only notice in year two: it couples the invoice to success. The more departments use the warehouse, the more dashboards appear, the more questions get asked, the higher the cost.
The obvious commercial reaction is control: cap usage, review queries, hand out budgets per department. That turns a tool meant to answer questions into a tool whose use has to be approved.
A warehouse whose usage you cap in order to control cost has missed its purpose — no matter how good the technology is.
What the figures support
Many numbers circulate on cloud cost and few of them hold up. Three sources I consider citable, and for each I say what it does not prove.
Flexera: 29 percent considered wasted
Flexera's State of the Cloud Report 2026 puts wasted cloud spend at 29 percent — rising again for the first time in five years. 85 percent of respondents name managing spend as their biggest challenge. The basis is over 750 cloud decision-makers surveyed.
What it does not prove: the figure covers cloud spend overall, not data warehouses specifically. And “wasted” is a self-assessment by respondents, not a measured quantity.
a16z: repatriation halves total cost
In 2021 Andreessen Horowitz estimated in “The Cost of Cloud, a Trillion Dollar Paradox” that repatriating cloud workloads roughly halves total cost of ownership, and extrapolated the effect to market valuations.
What it does not prove: it is an essay by a venture capital firm, not an empirical study. The 50 percent comes from a cited expert and is professionally disputed. The calculation was done for software companies at scale, not for warehouses in industrial companies.
37signals: one case with real invoices
37signals published a breakdown of its cloud exit: roughly $3.2 million in annual spend, then about $2 million less per year after a $700K hardware purchase. For the storage portion alone the bill fell from around $1.5 million to under $200K.
What it does not prove: it is a single case, and it is application hosting, not a data warehouse. Its value lies in being traceable, not in being transferable.
I have found no defensible study specifically on the total cost of data warehouses. Anyone quoting you a percentage for exactly that case should be able to say where it comes from.
Three cost drivers rarely named in a quote
| Driver | Why it grows | Visible in |
|---|---|---|
| Warehouses that never suspend | auto-suspend set too high or a floor above zero | the invoice, not the dashboard |
| Egress and data sharing | every handover to another system or site | a separate line, often overlooked |
| Repeated full scans | dashboards without materialization, refreshed hourly | credits consumed, not user count |
The third is the most awkward, because it grows with the number of dashboards rather than the number of people. Ten dashboards scanning a fact table hourly cost more than a hundred analysts asking a question twice a day.
What that means for your own numbers
Before comparing models, it pays to analyze your own query history. Three figures are enough:
- Ratio of peak load to median load. Below roughly three, provisioned capacity is usually cheaper; above ten, elasticity pays.
- Share of cost attributable to scheduled refreshes rather than human queries. If it is high, materialization is a cheaper lever than changing vendors.
- Share of egress in the total bill. It shows how much you pay for data having to leave the system.
All three are available in any warehouse and are rarely collected. They are more defensible than any study, because they describe your case rather than the average.
Sources
Every figure in this article is sourced. Where no defensible source exists, no figure is given.