How Snowflake Forecasts Consumption Based Revenue

True story - I remember the first time a VC asked me that, and I was like, well that’s a lazy question… ‘what’s YOUR AI strategy? I sell car parts online, my mans’

Didn’t go over that well…

So if that boardroom question makes you sweat, you're not alone. Many of us CFOs are still sorting through where AI actually fits into our finance orgs in a way that is practical, not just performative.

That’s why Brex created The CFO Guide to AI Strategy. It’s a straightforward look at how finance teams are using AI today to cut manual work, move faster, and make better decisions, without overcomplicating things. Get the guide:

The #1 question I got from people during annual planning season was:

“How the hell do I plan for customer success?”

So I figured screw it, let’s solve it together.

If you take this quick survey, I’ll write about the results at the end of the month. Please take it. It’s really easy (and anonymous, I don’t even ask for your email or anything).

Snowflake is THE canonical example of a consumption based model that’s thriving at scale. The company is firmly in the usage / commitment / consumption category, and Dikembe Mutumbo’s any reference to SaaS / ARR / subscription revenue.

In this post, we’ll cover:

Let’s ride.

While Snowflake was always usage-based, they weren’t the first to do it. That title goes to the cloud vendors (AWS, Azure, and GCP) who tied pricing to the consumption of resources like compute and storage time.

Snowflake adopted a similar approach but wrapped it in a system of credits, which now function as tokens. One credit typically equals an hour of compute time on a baseline warehouse. As Snowflake expanded its product suite, the pricing model matured into something more nuanced, not just compute, but also:

So yea, a lot going on. But it all ties back to the concept of a credit for some sort of resource usage.

To handle this complexity, Snowflake introduced a “consumption table.” Customers commit to a spend level, and a single SKU reflects that capacity. And on the backend, the consumption table tracks how various workloads translate to credit usage (e.g., this workload = 1.1x the standard rate). Kind of like a massive Vlookup and multiplication problem.

While commitments are captured as a “booking”, that’s actually just the start (or in some cases the end) of the revenue cycle.

Let me show you what I mean.

In a typical SaaS business, bookings come first.

You land a $1M seat based deal, lock the customer into a contract, and then recognize that revenue over time. Finance can model it cleanly. Sales rings the damn bell. And finance high fives if they were able to collect all the cash up front (F’yea).

But with Snowflake, that sequence is flipped.

Here, usage comes first, and only after a customer proves they’re consuming at a certain level does a booking happen.

Think of it like this:

A customer starts using Snowflake running small queries, testing performance, maybe even doing it with minimal committed spend. Then they ramp. Their usage grows.

Only once that usage crosses a meaningful threshold will the Sales team come in to negotiate a formal commitment.

So the booking (e.g., “$2M capacity agreement”) is just catching up to what’s already happening on the platform.

The implication is you can’t use bookings to predict revenue.

That’s why bookings - and by extension metrics like Remaining Performance Obligation (RPO) - are largely useless as forward-looking indicators from an operational standpoint. They reflect past behavior, not future intent.

Traditional SaaS forecasting says: Booking → Revenue

Snowflake says: Usage → Revenue → Bigger Booking (maybe)

The booking represents a customer committing to more because they were using consistently (and accruing revenue). Typically Renewals at Snowflake occur ahead of the scheduled date if everything is going well, unlike SaaS where you have a hard date that serves as a binary go or no go event. That’s why the reps have to be so involved with their target accounts, as they may increase commitments at any time.

On the finance team, instead of just modeling pipeline conversion rates on net new customers, it’s much more important to model consumption trends of existing accounts:

And yes, it’s a DAILY forecasting exercise, which we’ll talk about next.

Forecasting usage is a data science exercise.

Initially, Snowflake looked to Twilio, but Twilio’s hybrid model didn’t translate well (the larger customers looked more like SaaS). So they went upstream and talked to the cloud vendors who originally pioneered the model.

However, AWS and friends spent most of their time forecasting existing user behavior. Their bases were so large, new customers didn’t move the needle in the short term, as they still needed to ramp. Anything they got from new customers was just upside.

Snowflake’s challenge was tougher: how do you forecast a brand new customer with no usage history?

They approached it like a time series problem, where historical consumption is king. If there’s no history? Find a proxy, typically based on employee count.

A 10,000-person company behaves very differently from a 500-person one. Their consumption velocity, workload adoption, and seasonality patterns all diverge.

Still, every customer is unique. Some massive logos would load data, spike usage, and then… do nothing. Others grew steadily over time. Snowflake needed machine learning models to capture these behaviors and predict future consumption with daily granularity.

They’ve built their own in-house forecasting engine that:

Yes, they forecast 10 years out.

One model runs the daily forecast for now through year two, and another runs the model for years two through ten.

There are tons of seasonality quirks in consumption based models. They are dependent on customers using the platform and what’s going on within their business during a calendar year. If you think about a standard work week, employees are in the office Monday through Friday, which drives more consumption than on the weekends.

So if you’re forecasting a month you need to count how many work days there are. Some months there are 20, others 23. And then you have to calculate how much a holiday is worth (is it less than half of a weekday?). Internally they built a machine learning model to forecast customers and digest local holidays. If the customer has a ship to address in Japan and this month it’s Golden Week, you need to take that into account.

They also had to retrain models during COVID, when usage patterns completely changed. Employees were working around the clock outside of normal working hours.

At this point, Snowflake is basically a macro measure of enterprise productivity. If no one’s working, no one’s querying data.

Because most sales people coming in are from a typical software background they have to mentally reset what success looks like. Their key metric in terms of their effectiveness, success, and productivity is no longer defined based on them completing this booking event. Yes, I signed this contract with a customer… But that’s really just a bi product of usage.

That means they need to think about how to expand within the customer account from a use case perspective - finding workloads within each department where they can upload data.

One of the sneaky challenges in a usage-based world? It's harder for sales reps to blow past their number.

You have to change the culture around paying people in a consumption model. It’s harder to overachieve because the finance org is so good at forecasting users based on historical data. To be a million dollar rep you can’t just do 110% of your number.

In traditional SaaS, overachievement often comes from:

There’s a lever-pulling dynamic where a crafty AE can find a way to stuff more revenue into the quarter. Whether or not the customer actually uses the seats they bought? Not the rep’s problem.

But in Snowflake’s world, none of that works. Consumption models are hard to game. You’re not just selling a sku; you’re teaching higher adoption, which is a different skillset, and arguably a harder job.

Snowflake learned this the hard way.

At one point they had reps on 100% consumption-based comp, but reps struggled… not just with delayed payouts, but with the mental model. If your comp is based on how much a customer uses, and usage is predictable, how do you outperform?

Over time, they rebalanced:

It’s a spectrum; not one-size-fits-all. But the cultural shift is the same across the board:

Sales isn’t about the close. It’s about the adoption curve. Everyone is laser focused on adoption, just at varying levels linked to their ability to impact the result.

You could argue Snowflake sells the promise of compute, not software. Forecasting revenue in that world is about predicting behavior, rather than bookings. And that requires not just data science, but a complete cultural shift in how you think about pricing, forecasting, and paying your team.

If you’re a company making the shift from a traditional seat or licensed based model to consumption, the change permeates through all parts of your org. And you need everyone on board for the benefits to compound. If finance isn’t instrumented to forecast usage patterns on a daily basis, they will be late to the party when it comes to telling reps that a customer is either ready to convert to a longer term plan, or missing their expected bookings entirely (which will result in a revenue hole).

It also creates an emphasis on customer success being an org wide function, not a separate department. Snowflake is famous for not having CS because it creates a layer of abstraction between the rep hunting for new workloads and the customer. You gotta be there.

Usage based is in many ways as old as time - pay for what you use. But getting it right is part data science, part incentive structures, and part gut feel.

This is a special IN PERSON episode I recorded at Campfire’s annual finance conference.

I speak with Daniel Kang, CFO of Mercury, Brad Floering, SVP of Finance at Snowflake, and Mitzi Yue SVP of Finance at Boulder Care.

Today’s piece leans heavily on my discussion with Brad, who was the first FP&A hire at the commitment based juggernaut.

Here’s a guide to crafting a world class investor update. It covers both the monthly “flash” report as well as the more wholesome quarterly update.

It comes with a template you can steal (for all those who haven’t sent their board an update on year end 2025 yet!!!!!).

Hoping customers ramp into their commitments,

CJ