The forecast is four weeks long

Snowflake sells multi-year capacity contracts but earns its revenue only as customers consume, so its guidance is built from the last four weeks of daily usage. The contract tells you what will be billed; it doesn't tell you when any of it becomes revenue.

In May 2023 an analyst on Snowflake's earnings call asked the CFO, Mike Scarpelli, the question he said investors were asking him most: the forecast had come down a couple of times in recent quarters, so how could anyone be confident this cut was the last? Scarpelli's answer, in the corrected transcript, was that nothing about the method had changed: "we forecast looking literally at consumption trends on a daily basis, literally four weeks prior to the earnings through yesterday." What had changed was those four weeks. After Easter there had been, in his words, "four weeks in April where there was no week-over-week growth", which he put down partly to large customers deleting old data, one of them cutting its retention from five years to three. Full-year product revenue guidance went to about $2.6 billion, down from the $2.7 billion projected in March, and the stock had its worst day since listing. "In a consumption model, customers have the ability to dial it back," Scarpelli said, "and I can only guide based upon the data we have available to us."

The earlier posts in this series followed usage from the unit through the period close to the commitment drawdown on the invoice. This one is about what a company that sells that way can say about next year.

A contract that doesn't say when

Snowflake's contracts look predictable from the outside. Its 10-K for the year to January 2026 says that under capacity arrangements, "from which a majority of our revenue is derived, we typically bill our customers annually in advance of their consumption," and that it recognises revenue "as customers consume compute, storage, and data transfer resources." The billing follows the contract and the revenue follows the customer, so the invoice schedule can be known years ahead while the revenue line can't be known past the next few weeks.

The filing is direct about it. Remaining performance obligations were about $9.8 billion, of which Snowflake expects around 46% to become revenue in the next twelve months "based on historical customer consumption patterns," followed straight away by the caveat that those patterns "are not necessarily indicative of future results" and that RPO "does not account for the timing of customers' consumption or their consumption of more than their contracted capacity." Customers, it says elsewhere, "have flexibility in the timing of their consumption, which can exceed their contracted capacity or extend beyond the original contract term," and the company estimates future revenue "using predictive models based on customers' historical usage." Four years earlier, on the March 2022 call, the same figure was $2.6 billion with about 52% expected inside a year, so the share of signed value the company expects to see within twelve months has gone down as the pool has grown.

The vendor's own roadmap as a negative input

The less obvious problem is that a consumption vendor makes its own unit cheaper and has to forecast the effect. On that March 2022 call Scarpelli said Snowflake had "assumed an approximately $97 million revenue impact in our full year forecast" from platform improvements that in testing ran 10% to 20% faster, "but there is still uncertainty around the full impact." The detail came in the Q&A: a warehouse scheduling service, live for three weeks in January, had already taken $2 million out of the quarter, a gross impact of "more like $160 million" for the year was expected to be offset "by over $60 million in additional workloads," and the usual lag before cheaper compute pulled in new work was "about six months." The Register reported 30% wiped off the company's value after hours. By 2023 the estimate had become a standing rule: hardware gains from the cloud providers and Snowflake's own software work together came to, in Scarpelli's words, "a 5% headwind every year to our revenue."

The 10-K carries the same idea as a risk: better compression, faster processors and compute optimisation "may make our platform more efficient, enabling customers to consume fewer compute, storage, and data transfer resources to accomplish the same workloads," and if new workloads don't make up the difference, "we may experience lower revenue." A seat vendor that ships a faster product sells the same number of seats, while a consumption vendor has to put the release calendar into the revenue model with a minus sign, and guess how long customers take to spend the savings.

Modelling it, on paper

As before, this maps the published structure onto Revenue Cloud using public documentation mirrored in the SalesforceLabs revenue-cloud-foundations repository. It's a thought experiment, not a claim about how Snowflake runs its billing, and I haven't run any of it.

Revenue Cloud has a forecasting feature, and its scope is the useful part here. The Billing Forecast article for the Winter '27 release says it "estimates upcoming invoice charges before invoices are created," covering "new sales, amendments, renewals, and cancellations," and then: "Billing Forecast doesn't apply to milestone charges or usage charges." It works from billing schedules, and the worked example is all subscriptions: an annual platform fee, monthly licences raised by an amendment, a support package cancelled.

Put Snowflake's capacity contract through that and it splits in two. If the annual prepayment were modelled as a scheduled charge, it would appear in the billing forecast, dated and exact, because it is set by the contract. The consumption that turns the prepayment into revenue would not appear there at all. For that, the Usage Management overview points somewhere else: ad hoc generation of usage and ratable summaries "helps account executives generate reports to forecast consumption of a resource." The consumption forecast is built from the same UsageSummary records that feed the invoice, meaning from what customers have already done, which is Scarpelli's four weeks expressed as objects.

That split is the design decision the case forces. The billing forecast answers what the customer will be invoiced and is right as long as the contract holds; the consumption forecast answers when the prepaid balance will turn into revenue and is only as good as the history behind it. Neither one holds the two things that moved Snowflake's guidance: a vendor release that lowers consumption per query, and a customer changing its retention policy. A forecast built on summaries will see both of them only after they have shown up in the summaries, which is roughly what happened in April 2023. Snowflake's response then was to carry the weak weeks forward through the whole year "due to our lack of predictability and visibility" into customer behaviour, a choice about how far to extrapolate rather than a better model.

The next four weeks

The window works in both directions. In September Snowflake reported its second-quarter results and, as IT Brief reported, raised full-year product revenue guidance to $6.07 billion from $5.84 billion after a third consecutive quarter of faster growth. The next results are due on 2 December, and the guidance will again come from the weeks just before. A method that carried a four-week dip through a whole year in 2023 now has to decide what to do with four weeks of growth: does a consumption forecast treat a ramp as conservatively as it treated a slowdown, and should it?

No comments yet