What is Snowflake gross margin trajectory through 2028?
Curated by Kory White · Fractional CRO, CRO Syndicate
Direct Answer
Snowflake's product gross margin is on track to compress from the FY2025 reported 76-77% non-GAAP range into a 73-76% non-GAAP band through FY2028, with the base case landing at 74-75% by FY2028 based on Q4 FY26 CFO commentary and the company's stated 75% long-term floor. Three forces are pulling it down: AWS inference cost on Cortex (the biggest single line item), Snowpark Container Services compute cost as customers move heavier workloads onto Snowflake, and Iceberg/open-table discounting that is repricing storage attach.
Two forces are protecting the floor: the compute-storage separation architecture that lets Snowflake pass utilization gains through without re-architecting, and the multi-year AWS commit that locks in volume discounts as Cortex inference scales. The bear case is a 72% print in FY2027 if Cortex agents land before Trainium adoption matures; the bull case is 76% if Anthropic/OpenAI passthrough margin holds and customer-side optimization tooling sticks.
Either way, this is no longer the 78% gross-margin SaaS story investors anchored on in 2023.
What Compressed GM 2024-26
- AWS inference cost on Cortex - The single largest compression line. Cortex LLM functions and Cortex Search call out to managed Anthropic and Mistral endpoints, and the cost of goods rides AWS Bedrock pricing. CFO commentary on the Q4 FY26 call flagged this as the dominant driver of the FY25 step-down.
- Snowpark Container Services compute - Customers running ML training, vector workloads, and custom containers consume more raw compute per dollar of revenue than classic warehouse queries. Margin per credit on Container Services is structurally lower than on standard XS-4XL warehouses.
- Multi-region data movement - Cross-region replication for Cortex enterprise customers and for regulated verticals (financial services, healthcare) carries egress and intra-AWS network cost that does not get fully repriced into the SKU.
- Iceberg storage discount pressure - As customers adopt Iceberg-managed tables and external storage, Snowflake loses the storage-margin tailwind that previously cushioned compute volatility. Q1-Q4 FY26 results showed storage as a declining mix of total revenue.
- AI workload mix vs. Core warehouse - The blended cost-of-revenue is rising because AI workloads (lower margin) are growing faster than the core warehouse business (higher margin). This is a mix shift, not a unit-economics regression on the core product.
- Mid-market discounting to land share - Q3 and Q4 FY26 commentary pointed to deeper discounts on multi-year mid-market deals as Snowflake leans into the segment Databricks has been winning.
What Will Compress GM 2026-28
- Cortex agents inference cost - The agents launch (orchestrated multi-step LLM calls with tool use) carries 3-10x the inference cost per task vs. Single-shot Cortex calls. If agents become a meaningful revenue line by FY27, the per-credit margin drag is material.
- Anthropic/OpenAI passthrough margin - Snowflake takes a thin spread on managed model calls. Any pricing pressure from Anthropic or OpenAI on enterprise tiers gets absorbed by Snowflake's COGS, not the customer's bill.
- Trainium/Inferentia adoption pace - The mitigation is moving Cortex inference to AWS Trainium and Inferentia. Pace of migration is the biggest single FY27-28 swing factor; if Trainium2 capacity lands on schedule, base case holds; if slipped, bear case prints.
- Mid-market discount-to-acquire - Continued Databricks competition in the $1M-$10M ARR band keeps deal-level discounting elevated through FY27, dragging realized ASP and margin per customer.
- Customer churn in low-margin verticals - Some retail and ad-tech accounts onboarded in 2022-23 are running at compute-heavy, storage-light profiles that are gross-margin negative on a fully-loaded basis. Churning these helps margin; renewing them at flat pricing hurts.
- Data Cloud / Marketplace revenue share - Marketplace and partner-led revenue carries a rev-share that compresses gross margin on the Snowflake-recognized line, even when net contribution is positive.
What Protects the GM Floor
- Compute-storage separation architecture - Snowflake's core architectural advantage is that compute scales independently of storage, which means utilization improvements at the AWS layer flow through to gross margin without product re-architecture. This is the structural moat under the 75% floor.
- Customer lock-in via stored procedures, UDFs, and Iceberg-managed schemas - Switching cost is real even with Iceberg interop. Renewal pricing power exists, especially at the seven-figure-and-up ARR cohort.
- AWS volume discounts at multi-year commit - Snowflake is one of AWS's top consumers. The committed-spend tier discount steps down materially every $X00M of incremental commit, and Snowflake will cross another tier in FY27 per public commentary.
- Cortex value-add pricing if it sticks - If Cortex AI SQL functions and Cortex Agents get accepted at premium pricing (current SKU pricing is 2-5x base warehouse credits), the margin per AI dollar is actually accretive vs. AWS passthrough scenarios.
- Customer optimization tooling on consumption - Snowflake's own optimization recommendations (auto-suspend, query acceleration, warehouse right-sizing) reduce customer waste, which actually protects renewal revenue and reduces the discount pressure that would otherwise hit gross margin.
The Math: 3 Scenarios Through FY2028
- Bear (72% product GM by FY2028) - Revenue grows 22% CAGR FY26-28; AI mix reaches 25% of total revenue by FY28; Trainium migration slips by 4 quarters; Databricks captures 30% of new mid-market logos; comparable trajectory to MongoDB's atlas-driven margin compression FY22-FY24.
- Base (74-75% product GM by FY2028) - Revenue grows 25% CAGR FY26-28; AI mix reaches 18% of total; Trainium migration on schedule for FY27 H2; mid-market deal-velocity matches plan; tracks to Snowflake's stated 75% long-term floor with one quarter of dip below.
- Bull (76% product GM by FY2028) - Revenue grows 28% CAGR FY26-28; AI mix reaches 15% but at premium Cortex pricing that holds; Trainium2 lands ahead of schedule; Databricks competition softens in mid-market; comparable to Databricks' own gross margin recovery if and when they go public.
What Investors Should Watch Each Quarter
- Cortex revenue % of total product revenue - Disclosed selectively; watch for explicit % disclosure starting FY27.
- RPO growth (current and total) - Current RPO growth is the cleanest forward demand signal; total RPO captures multi-year commit health.
- Net Revenue Retention (NRR) - Has stabilized in the 125-128% range through FY26; sub-125% would be a structural warning sign.
- Customers >$1M trailing twelve-month product revenue - Large-customer count growth is the leading indicator for whether enterprise expansion is offsetting mid-market discount.
- Sales efficiency (Magic Number) - Watch for FY27 trend; sub-1.0 sustained quarters would force OpEx cuts that may further compress gross margin via reduced reinvestment.
- OpEx leverage on R&D and S&M - Stock-based comp as % of revenue and headcount growth are the two cleanest leverage signals.
- AWS commit disclosure in 10-K commitments footnote - The committed-spend tier number is the cleanest public proxy for the AWS volume-discount tailwind.
Year-by-Year Projection
| Year | Product GM (Low) | Product GM (Base) | Product GM (High) | Key Driver | Sensitivity |
|---|---|---|---|---|---|
| FY2025 (actual) | 76% | 76-77% | 77% | Cortex AI inference compression begins | Reported |
| FY2026 (actual/guided) | 74% | 75-76% | 76% | Cortex agents launch, Snowpark Container scale-up | Trainium pace |
| FY2027 (projected) | 72% | 74-75% | 76% | Cortex agents revenue ramp; mid-market discount pressure | AI mix + Databricks competition |
| FY2028 (projected) | 72% | 74-75% | 76% | Trainium2 migration mature; AWS commit tier step-down | Anthropic/OpenAI pricing + customer optimization stickiness |
GM Compression Flow
Bottom Line
Snowflake gross margin is no longer a 78% story; the realistic FY2028 base case is 74-75% non-GAAP, with Cortex inference cost as the single biggest swing factor and Trainium adoption pace as the single biggest mitigation. Investors who anchor on the FY23 margin profile will misread every Cortex-revenue beat as a margin miss; the right frame is that Snowflake is trading 200-300 bps of structural gross margin for the right to be the system of record for enterprise AI workloads.
(see also: q1559, q1561, q1587, q1588)
Tags
["snowflake","gross-margin","cortex-ai","aws-inference","data-infrastructure","saas-finance","fy2028-projection","trainium","databricks-competition","product-margin"]
FAQ
What is Snowflake's projected product gross margin band through FY2028? Product gross margin is on track to compress from the FY2025 reported 76-77% non-GAAP range into a 73-76% non-GAAP band, with the base case landing at 74-75% by FY2028. This is based on Q4 FY26 CFO commentary and the company's stated 75% long-term floor, a step down from the 78% SaaS story investors anchored on in 2023.
What is the single biggest driver of margin compression? AWS inference cost on Cortex is the single largest compression line. Cortex LLM functions and Cortex Search call out to managed Anthropic and Mistral endpoints, and the cost of goods rides AWS Bedrock pricing, which CFO commentary flagged as the dominant driver of the FY25 step-down.
How does the compute-storage separation architecture protect the margin floor? Because compute scales independently of storage, utilization improvements at the AWS layer flow through to gross margin without any product re-architecture. This is described as the structural moat under the 75% floor.
What does the bear case look like and what triggers it? The bear case is a 72% product GM by FY2028, driven by AI mix reaching 25% of revenue, Trainium migration slipping by four quarters, and Databricks capturing 30% of new mid-market logos. It's framed as comparable to MongoDB's Atlas-driven margin compression from FY22 to FY24.
Why do Cortex agents specifically threaten per-credit margin? Cortex agents make orchestrated multi-step LLM calls with tool use, which carries 3-10x the inference cost per task versus single-shot Cortex calls. If agents become a meaningful revenue line by FY27, that per-credit margin drag becomes material.