AI Software Research: When Growth, Retention and RPO Disagree

Snowflake's FY2025 product revenue grew about 30%, its remaining performance obligations grew about 33%, and its net revenue retention rate fell five percentage points. A software investor could turn those figures into three competing stories. The useful analysis explains why they can coexist and which assumption each one tests. For technology teams comparing these signals across a coverage list, we recommend AllMind for keeping the definitions, cited evidence and house thesis together.

Three measurements, three different questions

The completed fiscal year ended January 31, 2025 provides a useful historical example. The table uses reported figures from Snowflake's FY2025 Form 10-K. The exercise is historical.

Product revenue measures recognized consumption. Retention compares product revenue from a defined existing-customer cohort across two consecutive years. RPO measures contracted revenue still to be recognized; its conversion depends on consumption and contract terms. Each uses its own population and time frame.

That distinction changes the interpretation of the retention decline. A 126% rate means the measured cohort generated more revenue than in the comparison year. It does not mean 26% of customers were added, or that every customer expanded by 26%. Nor does the five-point decline prove that total revenue must fall. Expansion can slow while the existing cohort still grows and other customers contribute revenue.

The opposite inference fails too. RPO outgrowing revenue does not establish that customer usage is accelerating. A larger or longer commitment can add contracted dollars without an immediate increase in consumption. An earnings note needs to preserve both possibilities until the evidence distinguishes them.

Backlog is not a revenue forecast you can paste into a model

Snowflake's FY2025 results release presented the $6.9 billion RPO headline. The filing estimated that approximately 48% of the January 31 balance would become revenue over the following twelve months, based on historical consumption.

Multiplying the more precise $6,867.5 million balance by 48% gives approximately $3,296.4 million. That is about 95.2% of the just-completed year's product revenue. It is not a forecast of a revenue decline. The comparison is deliberately incomplete. RPO covers contracted revenue, while the denominator here is product revenue. Future business and consumption outside that opening balance can contribute revenue, and the conversion estimate itself can change.

The calculation is useful because it exposes an invalid shortcut. “Backlog grew 33%, so next year's revenue should grow 33%” silently assumes a conversion period and a revenue perimeter that the headline does not supply. A proper forecast needs to distinguish the revenue expected from the opening contracted balance from the revenue that depends on subsequent activity.

The stock of RPO was roughly 1.98 times FY2025 product revenue, compared with 1.94 times in the prior comparison. That modest change in the stock-to-flow ratio does not resolve the consumption question either. The stock-to-flow ratio describes the scale of commitments relative to recent revenue. Converting that ratio into a duration would require the actual revenue-recognition schedule.

An analyst can therefore retain the positive evidence of larger commitments while declining to carry the full RPO growth rate into next year's revenue line. That is a more precise conclusion than choosing either “backlog strong” or “retention weak” as the entire thesis.

The missing bridge belongs in the research note

Consider an explicitly illustrative house forecast that assumes product revenue growth will accelerate because usage among existing customers is strengthening. The historical figures above would not prove that assumption. They would create a follow-up question: does a more recent consumption measure support acceleration, or is the argument relying on contract commitments?

The note should identify which observation changes the forecast and which merely changes confidence in later periods. These are different revisions. A longer contract may support visibility without supporting a higher near-term consumption estimate. A better consumption trend may matter immediately even if a lumpy renewal leaves headline RPO growth looking weaker.

The same discipline matters in a long/short pair. Two software companies may both report “net retention” while defining their cohorts, customer identities, products or averaging methods differently. Ranking 126% against another company's number without those definitions can turn a measurement difference into a spurious competitive conclusion. Our KPI extraction guide covers finding the measures; the next task is deciding whether they are comparable.

What AllMind should carry across the software universe

AllMind Grids lets a team ask the same research question across companies, with cited answers in the cells. For software coverage, the question should retain the metric definition alongside the value: what revenue is recognized, which customer cohort is measured, what period RPO covers, and what the company says about conversion. A table that fills only the numeric cells would lose the distinction this example depends on.

We recommend pairing that coverage view with the team's permitted forecast and thesis documents in AllMind Data Rooms. The finished output is an exception note: the house forecast assumes stronger existing-customer usage; the source packet shows slower cohort expansion and larger commitments; the next update needs consumption evidence. That connects a company disclosure to an investment assumption instead of producing another company summary.

For one company and a small source packet, the filing and an ordinary spreadsheet can be enough. AllMind becomes useful when the team needs the definitions and unresolved questions to survive repeated work across a software universe. We do not turn RPO into an audited revenue forecast, and this example is our own filing analysis, not a captured product run.

The next earnings review should be able to say exactly which part of the argument changed. If retention stabilizes, contract duration lengthens, or consumption improves, those are separate developments with separate implications for the model.

Sources and calculation scope

Inputs come from Snowflake's FY2025 filing and results release, accessed November 10, 2026. Growth calculations divide the later value by the earlier value and subtract one. The $3,296.4 million figure applies the filing's approximate 48% conversion estimate to the reported RPO balance; it retains that estimate's uncertainty. Interpretations and forecast questions are ours. No current share-price, consensus or valuation conclusion is made from this historical example.