Ask a general-purpose data platform a question, and you get billed for it three times before you see an answer.
The dominant architecture for enterprise data platforms today separates storage, compute, and increasingly AI into independently metered layers. The pitch is flexibility: scale each layer on its own, pay only for what you use. In practice, a single natural-language question against your data quietly opens three running meters at once — a compute cluster spinning up and staying warm, storage billed by the byte on its own contract, and now an AI layer charging per token or per unit of model usage on top of both.
None of these meters know about each other. They just run, concurrently, for as long as the session stays open — whether or not the question that triggered them turned out to be useful.
Vendors selling this architecture will point to committed-use discounts, auto-suspend settings, and reserved capacity as the fix. Those help at scale. They don't fix the shape of the problem, which is structural: the cost of asking a casual, exploratory question is dominated by keeping infrastructure warm, not by the question itself.
Run the math on a warehouse that only needs to be up a few hours a day for ad hoc analysis, and compute alone can clear a couple hundred dollars a month — before a single AI token is billed, and before storage. That's not a pricing-tier problem. It's what happens when an architecture built for petabyte-scale pipelines gets asked to also be the interface for "what were our top five accounts last quarter."
Separating storage from compute was supposed to make cost proportional to value. Stacking a metered AI layer on top of that just proves the opposite — the more independently-billed layers a question has to pass through, the less proportional the bill is to the value of the answer.
The lakehouse model makes sense when the workload is the pipeline — petabytes, governance, ML training at scale. It makes much less sense when the workload is a person asking a question about data they already have.
A purpose-built analysis platform doesn't need to meter storage, compute, and reasoning as three separate products, because it isn't trying to be the warehouse too. It can treat the AI layer as the only thing it sells, and leave storage and compute to whatever the user already has — a local Postgres instance, a SQL Server sitting on hardware down the hall, or a cloud contract they've already negotiated on its own merits.
That's the model VerbaGPT is built on. Taurus, our local agentic mode, connects directly to databases you already run — on your own machine or over the network — and does the analysis there. There's no warehouse to spin up, no second meter for compute, no markup riding on storage you're already paying for. The subscription buys the reasoning. Everything underneath it is yours.
For most teams asking most questions, that's not a discount. It's a different bill entirely.