What does Snowflake churn math look like under AI pressure?

Direct Answer

Snowflake's churn math has three distinct buckets that AI pressure hits asymmetrically: logo churn (low, ~3-5% annually for $1M+ accounts), downsell/optimization (the headwind that crushed NRR from ~131% in FY24 to ~126% in FY25 and continues bleeding through FY26), and consumption-shrink (the AI-specific risk where Cortex agents and Iceberg substitution erode existing warehouse compute). The countervailing tailwind is real but unevenly distributed: Cortex consumption per AI query is genuinely additive, and AI-driven new workloads are widening seat-equivalent consumption across customer orgs. Net-net, the FY26 CFO commentary frames this as a battle between optimization-driven shrink and AI-driven expansion, with the long-term NRR floor model settling somewhere in the 110-115% range by FY28 if AI substitution accelerates faster than Cortex monetization. Operators modeling this need to stop treating NRR as monolithic and decompose it into the three buckets — because the defense plays differ wildly per bucket.

The Three Churn Buckets

• Logo churn (gross): Historically low at Snowflake's enterprise tier — top-100 customers churn at <2% annually, mid-market closer to 5-7%. AI doesn't move this much in 2025-26 because switching costs (data gravity, query rewriting, BI tool integration) remain high. The risk is *future* logo churn from Iceberg-portable customers in FY27-28.
• Downsell / optimization: The dominant NRR drag in FY25-26. Customers running cost-optimization programs (warehouse right-sizing, query tuning, multi-cluster consolidation) are pulling 10-25% of consumption out per renewal cycle. Capital One, NYSE, and other named optimization stories from FY25 earnings calls validate this is structural, not cyclical.
• Consumption-shrink (AI-specific): New in 2025-26. When a Cortex agent or external AI co-pilot answers a question that previously required 50 ad-hoc warehouse queries, that's compute that doesn't get spent. Hard to measure cleanly because Cortex itself adds consumption — the question is the *ratio* of substitution-to-net-new.

What AI Pressure Adds in 2026-28 (Headwinds)

• Cortex agents replacing warehouse queries: An agent that pre-computes a daily exec dashboard answer eliminates the ad-hoc warehouse spin-up that previously generated billable seconds.
• Iceberg substitution risk: Open table format adoption means customers can keep storage in S3/ADLS and bring any compute engine — Databricks, DuckDB, ClickHouse — to query it. The lock-in erodes.
• Customer self-optimization tooling: AI-powered query optimizers (third-party and Snowflake's own) help customers cut waste. Good for the customer, bad for consumption revenue.
• AI-native competitor steal: Databricks' Mosaic + Lakehouse pitch and MotherDuck's serverless DuckDB are siphoning new AI workloads, not just replacing existing ones.
• Named optimization stories: Capital One, NYSE, and JPMorgan-class accounts have publicly discussed multi-quarter optimization programs — these set the template other CFOs will follow.
• Macro CFO pressure: 2026 cloud-cost scrutiny is structurally higher than 2022-23. Every consumption SaaS vendor is feeling this; Snowflake just has more surface area exposed.

What AI Pressure Subtracts (the Tailwind)

• Cortex consumption per query: LLM inference inside Snowflake credits is genuinely additive — every Cortex Analyst, Cortex Search, and Document AI call burns credits the warehouse never would have.
• AI-driven new workloads: RAG pipelines, vector search, embedding generation, and agentic workflows are net-new compute that didn't exist in the FY24 baseline.
• Cross-functional adoption widening: AI is pulling Snowflake into marketing, support, and product teams that previously had zero data-platform spend. Seat-equivalent consumption widens.
• Named Cortex wins: Snowflake has cited specific customer Cortex deployments in recent earnings calls — these are proof points that the consumption tailwind is materializing in real accounts, not just slideware.
• Iceberg-as-onramp: Counterintuitively, Iceberg support is bringing *new* data into Snowflake's orbit that was previously locked in lake-only architectures — some of that becomes Cortex-queryable.

The Math: 3 NRR Scenarios FY27-FY28

• Bear (~115% NRR): Cortex monetization disappoints, Iceberg substitution accelerates, top-100 optimization continues. AI mix of total revenue stays <15%. NRR drifts toward the long-term floor faster than guided.
• Base (~120% NRR): Cortex consumption offsets ~half of optimization drag. AI mix reaches 18-22% of total revenue by end FY28. NRR settles in the low 120s, consistent with CFO long-term framing.
• Bull (~125% NRR): Cortex agents become a true second product line, AI mix exceeds 25% of revenue, named-account optimization plateaus. NRR stabilizes above the 120s through FY28.

Operator Moves to Defend NRR

1. Consumption-tier guarantees: Offer customers a credit-rebate if their optimized spend drops below a floor — trade margin for retention math.
2. Cortex bundling discount: Price Cortex credits at a discount when bundled with multi-year warehouse commits — shifts mix toward AI without leaving warehouse on the table.
3. Multi-year commit pricing: Push 3-year deals with usage ramps that bake in expected optimization — locks in NRR floor regardless of per-quarter consumption variance.
4. Named-account swat teams: ServiceNow-style strategic-account program for top-50 customers — embedded SAs, quarterly business reviews, AI workload roadmapping.
5. Iceberg-defensive product moves: Make Snowflake the *best* Iceberg query engine, not just a participant — turn the substitution risk into a moat.
6. AI co-development credits: Give top accounts free Cortex credits to build production AI workloads — seed consumption that compounds.
7. Optimization-as-a-service: Pre-emptively offer Snowflake-led optimization (with floor commitments) before customers hire third parties to do it for them.

Customer Cohort Risk Matrix

Churn Driver Flow

Bottom Line

Snowflake's NRR isn't dying — it's *rebalancing*. The optimization headwind is real and structural; the Cortex tailwind is real but immature. By FY28, expect NRR to settle in a 115-125% band depending on how aggressively Cortex monetizes versus how fast Iceberg substitution erodes warehouse lock-in. The operator job is to decompose churn into the three buckets and defend each one with a different play — treating NRR as monolithic is how you miss the structural shift. (see also: q1572, q1587, q1594)