Agents don't go home at 6pm

A human analyst runs 10-20 queries to answer a business question. An AI agent runs 500-2,000. A fleet of agents operating across an enterprise can generate 10,000-100,000 queries per hour without breaking a sweat.

This is not a hypothetical. It's the operational reality for any company deploying AI agents against production data. And it exposes a structural flaw in how the data infrastructure industry prices its products.

Per-query pricing — the dominant model across Snowflake, Databricks, and BigQuery — was designed for a world where humans were the bottleneck. Queries were relatively infrequent, individually expensive to compute, and easy to budget for. The pricing made sense when your data warehouse handled a few hundred queries a day from analysts who went home at 6pm.

AI agents don't go home at 6pm. They don't take lunch breaks. And they are fundamentally incapable of being "careful" with query volume the way a human is. Their value comes precisely from their ability to explore data exhaustively — testing every hypothesis, checking every correlation, running every permutation that a human would never have time for.

$27 billion to explain your database bill

The FinOps market — the industry that exists solely to help companies manage and reduce their cloud and data infrastructure spending — is projected to reach $27 billion. Think about what that means. We've built a multi-billion dollar industry whose entire purpose is to prevent people from using the infrastructure they're already paying for.

FinOps teams spend their days writing query governors, setting budget alerts, throttling warehouse sizes, and negotiating committed-use discounts. Every one of these activities is a direct tax on data access. They don't create value. They manage the cost of extracting value.

When you add AI agents to this equation, FinOps goes from annoying to existential. A single poorly scoped AI workflow can burn through a quarterly data budget in hours. The response from most organizations is predictable: limit what the AI can do. Cap query counts. Restrict it to pre-aggregated views. Make it dumber so it stays affordable.

BCG reported that 74% of companies struggle to scale AI value within their organizations. Infrastructure cost is not the only reason, but it is a consistent and underappreciated one. When every query has a marginal cost, curiosity has a price — and organizations learn to suppress it.

One query. $5,269. No, really.

Here's a concrete example. A Snowflake user ran a single Cortex AI query — one query — against a dataset of 1.18 billion records. The bill: $5,269. Not for a pipeline. Not for a month of analytics. For one question answered once.

Now imagine an AI agent that needs to run similar analytical queries across that dataset as part of its reasoning loop. Even at a fraction of that cost per query, 50 queries a day puts you at over $250,000 per day. No business model survives that math. So the agent never gets built, the insight never gets surfaced, and the data sits there being expensive to look at.

This is the core problem with per-query pricing at AI scale. It doesn't just make things expensive — it makes entire categories of applications economically impossible.

Pay for the pipe, not the water

The alternative is capacity pricing: pay a fixed cost for infrastructure, then run unlimited queries against it. This is not a new concept. It's how you pay for compute (EC2 instances), storage (S3), and bandwidth (CDN). You provision capacity and use it. The marginal cost of the next request is zero.

Data infrastructure has been the conspicuous exception to this model, largely because the dominant vendors built their businesses around per-query revenue. It's an extremely profitable model — for the vendor. For the customer, it creates a permanent tension between data access and cost control.

Under capacity pricing, the economics of AI agents change fundamentally. An agent that runs 10,000 queries costs the same as an agent that runs 10. There's no query governor, no budget alert, no FinOps review. The agent explores the data as thoroughly as its reasoning requires, and the infrastructure cost is a known, fixed line item.

This doesn't just save money. It changes what's buildable. Products that were economically impossible under per-query pricing — continuous monitoring agents, exhaustive anomaly detection, real-time personalization across millions of users — become viable when query volume is decoupled from cost.

The companies that get this will win

Companies building AI-first products face a choice that most don't realize they're making. The data infrastructure they select determines not just their current costs, but the upper bound of what their AI can do.

Per-query infrastructure puts a ceiling on intelligence. Every query has a cost, so every agent has a budget, so every answer is a compromise between thoroughness and affordability. This ceiling gets lower as data volumes grow — precisely the opposite of what you want.

Capacity infrastructure removes that ceiling. The constraint shifts from "how many queries can we afford" to "how much data can we store and index." That's a fundamentally different engineering problem, and one that scales in the right direction.

The companies that figure this out early will build AI products that are qualitatively different from their competitors — not because their models are better, but because their models are allowed to actually use the data. The companies that don't will spend the next several years wondering why their AI initiatives keep stalling at pilot stage, joining the 74% that struggle to scale AI value.

The bottleneck was never the model. It's the meter running on every question the model asks.