Snowflake Got Paid Only When You Used It. That Was Its Genius — and Its Trap.

Run a five-second query in Snowflake and you are billed for sixty. The warehouse spins up, the meter starts at a mandatory one-minute minimum, and only after that does the famous per-second pricing kick in.⁹ It is a small, almost invisible detail — and it tells you everything about the bet Snowflake made. While the rest of enterprise software was busy locking customers into fixed annual subscriptions, Snowflake chose to get paid only when you actually used the thing. The meter was the product strategy. It was also, eventually, the problem.

The official story is that Snowflake is a SaaS company. It is not — or at least not in the way the label implies. A SaaS company sells you a seat and recognizes that revenue smoothly across twelve months whether you log in daily or never. Snowflake recognizes revenue on consumption, not on the calendar.¹ The difference sounds like accounting trivia. It is the whole company.

Why Snowflake refused to sell seats

From its first commercial release in 2014, Snowflake billed for compute, storage, and data transfer that customers actually consumed, with revenue recognized on usage rather than ratably over a subscription term.¹ Founded in 2012 and selling two years later, it grew product revenue from roughly $95.7 million in the fiscal year ending January 2019 to $252.2 million the following year — a near-tripling that happened largely because existing customers kept consuming more.⁴ That is the elegance of the design. A data team that finds Snowflake useful runs more queries, loads more data, and spins up more warehouses — and every one of those actions is a sale. There is no upsell call, no renewal negotiation to expand the contract. The product grows the account on its own. When the IPO came in September 2020, the market clearly liked the math: the shares priced at $120, well above the $75–$85 range the company had estimated only days earlier.²³

The single number that captured this was net revenue retention — how much an average cohort of customers spends this year versus last. For the fiscal year ending January 2020, Snowflake's NRR hit 169%, meaning the typical customer spent nearly seventy percent more than the year before without a single new logo being added.¹ On a small revenue base, that is the most beautiful figure in software: growth that arrives whether or not the sales team closes anyone new. The consumption meter, ticking upward across thousands of accounts, was doing the selling.

The meter runs both ways

Here is the part the bull case quietly skipped. A meter that captures every extra query on the way up captures every cancelled one on the way down — instantly, with no annual contract to cushion the fall. Snowflake's own SEC filings said so in plain language: because revenue is recognized on consumption rather than contract duration, there was a real risk that customers would consume the platform more slowly than expected, and that slowdown would hit recognized revenue directly.⁸ A subscription business with a customer trimming usage still books the full seat fee until renewal. A consumption business books the lower number this quarter. When the macro climate turned and finance teams started hunting for cloud savings, Snowflake's billing model handed them the knife. Every query a customer optimized away was revenue Snowflake never recognized.

And the descent was not a single bad quarter — it was a steady, multi-year compression. Net revenue retention fell from 142% in July 2023 to 135% that October, 131% by January 2024, 128% in April, 127% across the middle of the year,⁶ and 126% by January 2025.⁷ Each step was small. The trajectory was unmistakable. Snowflake was still growing — product revenue reached roughly $2.67 billion for the year ending January 2024, up 38%, and $943.3 million in the final quarter of the next fiscal year, up 28%.⁵⁷ But the rate of deceleration tracked the NRR line almost exactly, because they are the same phenomenon: customers consuming more carefully.

But isn't aligned pricing supposed to be the good kind?

The fair objection is that this is exactly what good pricing should do. Consumption billing aligns Snowflake's revenue with the value customers receive — they pay for what they use, nothing more, and nobody resents a bill that tracks their own behavior. That alignment is real, and it is a genuine moat: a customer who has built their data estate on Snowflake's meter is deeply embedded, and the same flexibility that lets them cut also makes the platform easy to adopt without a large upfront commitment. The honest counter is that alignment cuts symmetrically. The pricing model is not failing — it is working precisely as designed, transmitting customer reality straight to the income statement in both directions. Snowflake even names the threat itself in its own risk disclosures: the extent to which customers keep optimizing consumption, the effect of cheaper tiered and Iceberg-table storage on usage, and customers shortening contract durations to manage cash.⁸ That is not a company surprised by its own pricing. It is a company that built the most honest revenue model in enterprise software — and is now living with the volatility honesty costs.

Snowflake made a wager most of its industry refused to make: that the fairest way to charge is to charge for use, and that a product good enough to be used more would never need a sales team to grow the account. On the way up, that wager produced one of the cleanest growth stories in software. On the way down, it produced one of the clearest. The meter was never the bug and never the feature — it was simply the truth, ticking. Snowflake built a business that gets paid exactly what its customers think it's worth, and then discovered the inconvenient corollary: when they think it's worth a little less, you find out the same afternoon.