Snowflake Bills Data Compute by the Second — and Wrote the Volatility Into Its Own Risk Factors

Spin up a Snowflake warehouse to run one query, and the meter starts the instant it wakes — a 60-second minimum charge, then billing by the second for as long as it runs.¹ Shut it down and the meter stops. There is no monthly seat, no ratable slice booked whether you touch the product or not. Snowflake gets paid when — and only when — a customer actually burns compute. It's the difference between renting an apartment and paying for a taxi: you can leave the apartment empty and the landlord still collects; the taxi earns nothing while it idles.

The official story is that Snowflake's genius was inventing per-second cloud billing. It wasn't. The per-second increment and the one-minute minimum are conventions Google's BigQuery already used.⁷ The real decision — the one that shaped the company's income statement and its stock — was choosing to recognize revenue on that consumption, and accepting that real-time demand swings would flow straight through to the top line instead of being smoothed away.

The choice was recognition, not the increment: billing by the second is a convention; letting consumption drive the reported number is a strategy

A subscription software company sells you a year and books that year in even monthly slivers, whether you log in daily or never. Revenue arrives smooth, predictable, forecastable — the whole point of the SaaS model was to make the next quarter legible. Snowflake threw that legibility away on purpose. Its product revenue is tied to consumption of compute, storage, and data transfer, which the company itself notes is variable at the customer's discretion and not tied to the size or duration of a contract.² That means demand isn't a leading indicator that Snowflake gets to watch and plan around. Demand is the revenue, recorded in the same period it happens. When customers lean in, the number rockets. When they pull back, there is no ratable buffer to hide behind — the drop shows up now.

For a few years the model was pure tailwind. Total revenue more than tripled from $592 million to $2.1 billion in two fiscal years.³ Consumption-based recognition amplified the good times: when customers' own businesses were growing fast, they ran more queries, and Snowflake booked that surge in real time rather than waiting for contracts to renew. But the same filing that showed the tripling also carried the warning — Snowflake disclosed, as a formal risk factor, that because consumption is discretionary and hard to model, its ability to forecast revenue is limited, and that it expected its revenue growth rate to decline.³ The company wrote the volatility into its own prospectus. It knew exactly what it had signed up for.

When a handful of customers turned the meter down: the slowdown wasn't a broad enterprise retreat, and the CFO named who it wasn't

The warning came due in 2022. Product revenue that had grown 70% in fiscal 2023 was guided to just 40% for fiscal 2024.⁵ The lazy read was 'macro headwinds hit Snowflake,' but that flattens what actually happened. On the Q1 fiscal 2023 call, CFO Mike Scarpelli was precise: the drag was concentrated in a subset of consumer-facing cloud customers who were consuming less than anticipated in a tougher environment — a mirror image of the prior year, when those same kinds of customers' own rapid growth had driven consumption far above expectations.⁴ When an analyst reached for the obvious digital-native names, Scarpelli said flatly that Meta, Netflix, and Peloton were not among those dragging results down.⁴ That specificity is the tell. A subscription business would have reported the same smooth line regardless of how much those customers actually used the product. Snowflake couldn't, because its revenue was their usage.

Isn't this just pay-as-you-go with a nicer story?: the pure-usage framing is wrong, and the volatility is a feature the company chose to keep

The fair objection is that this is just pay-as-you-go, and pay-as-you-go isn't exotic. Two things complicate that. First, Snowflake isn't purely on-demand: while its on-demand track carries no minimum, most large customers sign prepaid capacity-commitment contracts with a stated minimum around $25,000, drawing them down over time.⁶ The per-second granularity governs how a commitment is consumed, not whether one exists — so there is a floor beneath the volatility, not a free-fall. Second, and more honestly: the volatility genuinely hurt. A model that spikes on the way up will sag on the way down, and Snowflake's forecasting was so tied to customer behavior that its own guidance for fiscal 2024 was a moving target across a single year — a drag it tried to reframe by claiming its addressable market had grown from $81 billion at its 2020 IPO to $248 billion about two and a half years later.⁵ The steelman holds partway: consumption pricing is real utility, aligning what customers pay with what they use, which is why they tolerate it. But the company traded predictable revenue for honest revenue. When usage was up, it collected more than a subscription would have allowed. When usage fell, it had nowhere to hide. That is the deal, taken whole.

Snowflake didn't invent billing by the second; it inherited an increment BigQuery already used.⁷ What it chose was to let that meter tell the truth about demand in real time, and to book its revenue on the truth rather than on the comfortable fiction of a smoothed contract. That choice bought explosive early growth and a growth rate that fell exactly as promised. The company's most quietly radical decision wasn't in its pricing page. It was in its risk factors — where it agreed, in writing, that it could not fully predict its own next quarter, and shipped the model anyway.