Snowflake Gets Paid Only When You Run. That's the Genius — and the Time Bomb.

A data engineer at one of Snowflake's largest customers tunes a query overnight. The next morning the same report runs in eleven seconds instead of ninety, on a warehouse that wakes up, does its job, and goes back to sleep. The customer is delighted: the work got cheaper. And here is the strange part — Snowflake's revenue from that customer just went down. Not next quarter, not at renewal. That morning. Snowflake doesn't sell seats or licenses; it sells the meter. And the meter only spins when work is happening, which means every efficiency the customer finds is an efficiency Snowflake's income statement absorbs in real time.

The official story is that Snowflake is a subscription software company with the best net revenue retention in enterprise software — 158% at IPO¹, a flywheel of customers who spend more every year. That framing is a category error. Snowflake isn't a subscription. It's a utility, and a utility's revenue is hostage to how efficiently its customers use the thing they're metered on.

The meter that spins both ways

Here is the thesis a smart friend can repeat at dinner: Snowflake's consumption pricing is not the durable recurring-revenue engine the bull case priced — it's a real-time amplifier that turns every efficiency gain, every open-format migration, and every IT budget cut directly into a revenue decline. The same mechanic that made expansion look effortless on the way up makes contraction unstoppable on the way down. Snowflake gets paid when you compute. So the moment you compute less — or compute the same work for fewer credits — the model has no buffer. There's no annual license floor to fall back on. The meter is the whole business.

The mechanic is mundane and that's exactly why it's load-bearing. Compute is billed in credits, per second, with a 60-second minimum every time a warehouse starts or resumes, plus another minute's charge on the added compute whenever a warehouse is resized larger mid-run.¹² Storage is billed per compressed terabyte. Cross-cloud and cross-region egress carries a data-transfer charge. The cloud services layer — metadata, security, query planning — is free up to 10% of daily compute credits, and only billed above that.⁵ None of this is a subscription. Even the pre-purchased capacity contracts that carry most enterprise revenue aren't recognized as a flat fee; they're drawn down as customers consume, which is why deferred-revenue balances exist at all. Buying capacity up front doesn't change the physics. It just front-loads the commitment to a meter that still only pays out when work runs.

More work, less money: the divergence the bulls missed

For years the consumption model read like a one-way ratchet because volume only ever went up. Net revenue retention of 158% at IPO meant the existing customer base was spending 58% more in aggregate than it had a year prior, and a headline like that gets priced as if it's structural.¹ But the S-1 itself told a quieter story underneath the number: NRR had already fallen from 223% in FY2019 to 180% to 169% to 158% across successive periods.¹ It kept falling — 131% by FY2024², 126% by FY2025³, 125% by FY2026.⁹ A decade-long decay, dressed up as durability. The bull case kept quoting the level. The story was in the slope.

Why does it decay? Because in a credit-consumption model, the platform's own progress is deflationary to its revenue. Snowflake's Q1 FY2024 disclosure laid it bare, flagging slower-than-expected consumption and customers actively optimizing usage — including by shortening data-retention policies to reduce storage.⁶ On the Q4 FY2024 earnings call, CFO Mike Scarpelli put the divergence in numbers, noting that jobs run on the platform had grown 62% year-over-year against only a 33% rise in revenue — because Snowflake keeps getting cheaper for customers to run.¹¹ That isn't a slowdown in demand — customers were doing more than ever. It's that each job cost fewer credits as the engine got faster, and that data was starting to leave Snowflake's billed storage entirely. The work grew far faster than the bill.¹¹ That dynamic is the single most important thing to understand about this company, and it is the exact opposite of how recurring-revenue software is supposed to behave.

Then came the structural twist. Iceberg Tables let customers keep their data in an open format that sits in cloud storage they manage themselves — and Snowflake's documentation confirms it does not bill for storage on Iceberg tables that use a customer-managed external volume.¹³ Tiered storage pricing compounded the squeeze. Snowflake itself named both in its Q1 FY2025 risk factors: 'the impact of new or optimized product features and pricing strategies on consumption, including Iceberg tables and tiered storage pricing.'⁷ The company, in other words, built the features that let its own meter spin slower — because customers wanted them, and the alternative was losing the customer. That's the trap of the consumption model in one move: you cannot refuse the efficiency your customers demand, and every efficiency you ship is a haircut on your own revenue.

On February 28, 2024, the amplifier ran in reverse in front of the whole market — a session complicated by Snowflake also announcing that CEO Frank Slootman would retire, which weighed heavily on the stock alongside the weak outlook.¹⁴ Snowflake withdrew the $10 billion FY2029 product-revenue target — the long-range number the bull thesis had been built around — and cut FY2025 guidance to roughly $3.25 billion, about 22% growth against a consensus near 30%.⁸ The stock fell about 18% in a single session.¹⁰ A class-action suit covering June 2023 through that day alleges the company already knew Iceberg adoption and tiered storage pricing were creating material consumption headwinds before it withdrew the target.⁸ You can litigate the disclosure timing. You cannot litigate the physics: the model that had transmitted every gain into revenue was now transmitting every loss just as fast.

Isn't this still the model everyone now wants?

The honest objection is that consumption pricing is the winning structure of the era, and Snowflake's numbers still look formidable — $4.7 billion of revenue in FY2026, up 29%⁹, with 13,328 customers and 790 of the Forbes Global 2000 on board.⁴ All true, and the case for the model is real: it aligns price with value, it lowers the barrier to land a customer because nobody signs a big license to start, and it lets a champion inside the customer prove ROI before scaling. Remaining performance obligations of $6.9 billion show customers are still committing capacity years out.³ A consumption model that nobody has to pre-justify is genuinely easier to expand than a seat model gated by procurement.

But notice what the steelman quietly concedes. Every advantage of the model on the way up — friction-free expansion, value-aligned billing, a meter that grows with the customer — is a liability on the way down, because the same meter shrinks with the customer just as fast and with no contractual floor underneath. The bull priced the model as a recurring-revenue annuity. It is closer to a toll on a road whose traffic the toll operator keeps making faster and cheaper to drive. The 125% NRR⁹ isn't a bad number. It's a falling one, and in a consumption business the direction is the whole story. Durable expansion was the asset everyone bought. Real-time sensitivity to their customers' efficiency was the asset they actually got.

Snowflake built one of the cleanest pricing models in software: pay for exactly what you use, nothing wasted, value and price perfectly aligned. The genius and the curse are the same sentence. When the work grows, Snowflake grows with it — serenely, automatically, no renewal required. And when the work gets cheaper — because the engine improved, because the data moved to open formats, because a budget got cut — Snowflake shrinks with it just as serenely, just as automatically, the same morning the engineer tuned the query. The meter was never the recurring-revenue engine the market paid for. It was a mirror, reflecting back, in real time and without mercy, exactly how efficiently the rest of the world had learned to compute.