Snowflake Sells "Pay Only For What You Use." Most of Its Money Is Prepaid.
Snowflake's pitch is elastic, on-demand metering. But per its own 10-K, the majority of product revenue comes from capacity contracts billed a year in advance - a credit-burn model closer to SaaS than to the meter. And the meter is slowing: net revenue retention fell from 158% to 125%.
Comes with a free Profit-Engine Map template.
Picture the meter on the side of a house, spinning a little faster every time someone runs the dishwasher. That is the picture Snowflake wants you to hold: a data platform you pay for only as you use it, the dial ticking up with each query, idle when you sleep. It is a beautiful story, because it lines you up perfectly with the customer - you both win when usage grows. But open the 10-K and the meter turns out to be mostly a formality. The majority of Snowflake's product revenue arrives the way a gym membership does: prepaid, a year in advance, against a number you committed to before you ran a single query.1
The official story is that Snowflake charges only for what you use, in real time. The truer story is that most customers buy a bucket of credits up front, billed annually, and then spend the year burning it down. On-demand metering - the elastic, pay-as-you-go dial in every pitch deck - is the minority path, the one smaller and lower-usage customers take.1 The difference is not cosmetic. It changes what kind of company this is.
“Product revenue is recognized based on the consumption of compute, storage, and data transfer resources... Our customers have the flexibility to consume more than their contracted capacity and may also roll over any unused capacity.”3
The credit bucket, not the spinning dial
Here is the thesis a smart friend could repeat at dinner: Snowflake isn't a pure usage meter - it's a prepaid credit-burn business that recognizes revenue as the credits drain. A customer signs a capacity contract, pre-commits to a fixed dollar amount of consumption, and gets billed for it annually in advance. Those rates run 20-to-30% below the on-demand list price, which is the whole inducement - commit big, pay less per credit.7 The cash arrives up front; the revenue is recognized as queries burn the balance down.13 That structure does two quiet things. It locks the customer to a number they already paid for, and it gives Snowflake the predictability investors usually associate with subscriptions - while letting it keep telling the elastic, consumption-aligned story that sounds so much better than 'enterprise contract.'
And there is real machinery underneath. When the bucket drains, you can keep going - consume past your committed capacity, or roll the unused remainder forward.3 So the model is genuinely consumption-flexible at the edges. But the center of gravity, the part that pays the bills, is the prepaid commitment. The dial spins; the cash was already collected.
Five meters, and most teams budgeted for one
What the customer is actually burning credits on is more layered than 'I ran a query.' The platform meters compute as virtual warehouse credits, charged per-second with a 60-second minimum every time a warehouse spins back up - and every step up in warehouse size doubles both the compute power and the credit burn rate.6 Storage is metered separately, a flat rate per terabyte of compressed data per month. Cloud services - the coordination layer - is free up to 10% of your daily warehouse credits, then billed on the excess.6 On top of that sit serverless features, container compute via Snowpark Container Services, and a newer layer of AI credits at per-token rates.9 The trap is structural: a team budgets for the warehouse they can see and gets surprised by the four meters they couldn't. Doubling a warehouse to make a query finish faster doesn't save money - it finishes in half the time at twice the burn, and the bill barely moves.
| The consumption story | The capacity reality | |
|---|---|---|
| How it's billed | Metered, in arrears | Prepaid, annually in advance |
| Share of product revenue | Minority (on-demand) | Majority |
| Per-credit rate | List price | 20-30% lower |
| What it creates | Alignment with usage | Switching costs + revenue predictability |
| Looks like | A power meter | A SaaS contract wearing a meter's clothes |
The number that is quietly deflating the whole model
A consumption business lives or dies on one figure: net revenue retention - how much more an existing cohort of customers spends this year versus last. For Snowflake it has done something no growth narrative wants to admit. It fell from 158% at the start of 2023 to 151%, then 142%, 135%, 131%, 128%, 127%, and on down to 125% by January 2026.452 That is not a bad quarter. It is a line sloping the same direction for three straight years - a structural compression of the rate at which customers grow their spend.
The mechanism is the cruel inverse of the alignment Snowflake brags about. Because you pay by consumption, anything that makes a workload cheaper to run cuts the bill: a better-tuned query, a right-sized warehouse, an engineer who finally turns off the warehouse that was idling overnight. Customer efficiency is Snowflake's revenue leak. And AI threatens to widen it - as models and optimizers compress what a given analysis costs in credits, the same business need burns fewer of them. To keep net retention above 100%, Snowflake has to keep finding new workload surface area - new meters, new credit types, new reasons for the bucket to drain - faster than its existing customers learn to drain it slower. That race is the entire investment case.
Isn't 125% still excellent - and isn't the business still huge?
The fair objection is that any number above 100% means customers are still spending more every year, and 125% is a figure most software companies would trade their roadmap for. True - and the absolute scale backs it up: $4.7 billion in total revenue in FY2026, 13,328 customers, and 790 of the Forbes Global 2000 contributing roughly 43% of revenue.2 The base is real, deep, and enterprise-anchored, and the prepaid capacity structure makes that base genuinely sticky. The honest counter, though, is about direction and second derivative. A consumption model is valued not on the level of expansion but on its persistence, and a metric that has fallen in nearly every quarter for three years tells you the easy expansion - lift-and-shift workloads from legacy warehouses - is largely spent. Slootman's $10 billion FY2029 ambition was floated at the 158% peak and formally withdrawn on the day his successor took the helm — when Snowflake simultaneously disclosed that product efficiency gains were eroding revenue at a rate of roughly 6% per year.811 The remaining question is not whether Snowflake is large. It is whether it can manufacture new credit-burning workloads faster than its customers and its AI tooling learn to burn fewer.
Consumption pricing sounds like perfect alignment - you only win when the customer uses more. But it hides an inversion most operators miss: every gain in your customer's efficiency is a cut to your revenue. The better they get at running on you, the less they pay. That works beautifully while a market is migrating onto your platform for the first time, because raw new usage swamps any optimization. It turns against you the moment migration slows and your installed base starts tuning its bills down. So if you sell a meter, watch the retention rate's slope, not its level - and have the next category of workload ready before the current one gets cheap. The prepaid capacity contract buys you predictability and switching costs in the meantime, which is exactly why the 'pure consumption' story is worth quietly hedging against your own customers' competence.
Snowflake built something genuinely hard - a place where the world's enterprises now keep their data and pay to ask it questions. But the genius investors fell for, the friction-free meter that spins up only when value is created, is mostly a story told over a SaaS contract. The cash comes prepaid; the meter is the marketing. And the deepest risk isn't a competitor or a price war. It's that the company's interests and its customers' interests point in opposite directions on the one thing every customer eventually gets good at: spending less. A money machine that runs on consumption is only as durable as the world's appetite to consume more - and Snowflake's own retention curve is the sound of that appetite, slowly, getting full.
Profit-Engine Map
A one-page map that pulls a business apart into the hook that gets the customer in the door and the engine that quietly earns the margin. Use it to see where the real profit lives, how the two halves are wired together, and what breaks if the link is cut. Blank to dissect your own P&L; filled as the worked example of a business whose advertised product is not where it makes its money.
The worked example unlocks with a subscription. See plans →
Sources
Where this comes from — the filings, records, and reporting behind it.
- 1Snowflake provides its platform through a consumption-based business model; customers choose between capacity arrangements (pre-committed, majority of revenue, billed annually in advance) or on-demand arrangements (charged monthly in arrears). Product revenue for FY2024 was $2,666.8M; FY2023 was $1,938.8M; FY2022 was $1,140.5M.
- 2Snowflake's FY2026 (ended Jan 31, 2026) total revenue was $4.7 billion; net loss was $1.3 billion; net revenue retention rate was 125%; customers totaled 13,328 including 790 Forbes Global 2000 companies contributing ~43% of revenue; customers with >$1M trailing product revenue grew from 576 to 733 year-over-year.
- 3Product revenue is primarily derived from the consumption of compute, storage, and data transfer resources; recognized based on platform consumption (variable at customers' discretion), not contract term duration; customers may consume more than contracted capacity and roll over unused capacity. This model distinguishes Snowflake from subscription-based SaaS that recognizes revenue ratably.
- 4NRR declined from 158% (Q4 FY2023, January 31 2023) to 151% (Q1 FY2024), 142% (Q2 FY2024), 135% (Q3 FY2024), 131% (Q4 FY2024) — a continuous multi-quarter compression documented in sequential SEC 8-K filings.
- 5NRR continued declining: 128% (Q1 FY2025, April 2024), 127% (Q2 FY2025, July 2024), 127% (Q3 FY2025, October 2024), reaching 125% as of January 31, 2026 per the FY2026 annual 10-K.
- 6Snowflake bills three primary meters: compute (virtual warehouse credits, per-second with 60-second minimum each time warehouse resumes), storage (flat rate per TB/month on compressed data), and cloud services (free up to 10% of daily warehouse credit consumption, then billed on excess). Each warehouse size doubles compute power and credit consumption rate vs. the next smaller size.
- 7On-demand credit costs range from $2.00 (Standard, US AWS) to ~$6.00 (Virtual Private Snowflake, US AWS); non-US regions add ~50%; capacity contracts reduce rates by 20–30%. AWS US East on-demand storage is now listed at $23/TB/month compressed (not $40/TB, which is a stale figure no longer in Snowflake's published pricing).
- 8The $10 billion product revenue target for fiscal 2029 was stated by then-CEO Frank Slootman in the Q4 FY2023 earnings press release (filed as 8-K exhibit, March 1, 2023). Slootman also stated FY2023 product revenue of $1.9 billion with 70% YoY growth.
- 9In addition to virtual warehouse compute, storage, and cloud services, Snowflake's billing architecture includes serverless compute (features like Snowpipe and Automatic Clustering that run on Snowflake-managed resources), Snowpark Container Services compute pools, and AI Services such as Cortex LLM Functions billed at per-token rates.
- 10Effective February 27/28, 2024, Frank Slootman retired as CEO of Snowflake and Sridhar Ramaswamy was named CEO and member of the Board of Directors, immediately.
- 11The $10 billion FY2029 product revenue target, reaffirmed as recently as Snowflake's June 2023 Investor Day, was withdrawn on February 28, 2024 — the same day Sridhar Ramaswamy was named CEO — when the company disclosed product efficiency gains were creating a 6.2%–6.3% annual revenue headwind and guided FY2025 growth well below the pace required to hit the target.