When we published Snowflake — Initiation 2026, the open question on the name was whether a business that had decelerated to 30% growth could turn the AI story into revenue rather than slideware. On 2 September, Snowflake answered it. Total revenue grew 35% year-on-year — the third straight quarter of acceleration — and the company raised guidance while doing it.
The quarter, quickly. Product revenue of $1.49 billion, up 37% and the second consecutive quarter of record sequential dollar growth. Non-GAAP operating margin of 15%, up more than 400 basis points year-on-year. Net revenue retention of 126%. Remaining performance obligations of $9.0 billion, up 30%. The customer base reached 14,554, with 692 net new adds — a 32% increase in net additions year-on-year — and 65 customers now above $10 million of trailing-twelve-month product revenue. The company ended with $4.3 billion of cash and investments.
The reacceleration is the whole story
Snowflake exited last fiscal year growing 30%. It is now growing 37%. That is seven points of acceleration in two quarters, and the third quarter in a row that the growth rate went up rather than down.
Asked to split the acceleration between the new AI products and the core platform they run on, CEO Sridhar Ramaswamy would not let the answer be tidy: he called it “roughly even.”
AI products — Cortex Code (CoCo), CoWork, AI functions and the new Cortex AI Gateway — contributed about half of the acceleration. The other half came from the core doing what it always did, only faster: Notebooks, Streamlit and React applications deployed into Snowflake, and above all migrations.
Migrations are the quiet engine here, and AI has changed their clock speed. Ramaswamy’s example: a large network-equipment manufacturer is completing a Teradata migration in under three quarters — work that “would have taken probably 2 to 3 years in any previous time.”
A large Australian bank moved its financial-crime platform onto Snowflake, processing 17 billion transactions at 10x faster query performance, and is now using AI agents to migrate the rest of its estate. This is the mechanism by which AI pulls core consumption forward, not just AI-line revenue.
The trade nobody is hiding
Here is the number that deserves its own sentence. Snowflake cut its full-year non-GAAP product gross margin guide to 74% — and raised its full-year operating margin guide, from 13.5% to 14.5%, in the same breath.
Those two moves point in opposite directions on purpose. The gross-margin cut is the cost of the AI mix: fast-growing AI workloads “carry a lower contribution margin today,” and the more of them Snowflake sells, the more the blended gross margin gives way. CFO Brian Robins was blunt about the sequencing — build a great product, drive adoption through use cases, then work on the margin. Gross margin fell “because we’ve increased our guidance so much.”
So how does operating margin go up while gross margin goes down? The answer is on the headcount line, and it is the most striking single figure in the release.
| Year-to-date FY27 | Year-ago period | |
|---|---|---|
| Employees added | 334 | 935 |
| of which, from the Observe acquisition | 173 | — |
| Organic adds, implied | ~161 | 935 |
Snowflake is offsetting rising cloud costs with slowing headcount expense. It is buying acceleration with gross margin and paying for the profitability with restraint on people. That is a coherent trade — but it is a trade, and it puts the burden of the margin story on two levers that both have limits: cloud unit economics, and how long you can grow 37% while barely hiring.
Snowflake as its own case study
The headcount restraint is not austerity. It is the product. Snowflake is running CoCo and CoWork through its own business, and the internal numbers management disclosed are the most concrete evidence on the call that this technology removes labour rather than merely assisting it.
- Marketing. CoCo brought search optimisation in-house, eliminating $400,000 of annual agency spend. Keyword research fell from ~10 hours to 20 minutes; content production from an estimated 24 hours to 2.
- Finance. Long-range planning — once a three-person team and more than 50 spreadsheets — now runs with one analyst and a set of models. Adoption inside finance is near 100%, spanning deal desk, tax, accounting, internal audit, FP&A and treasury.
- Sales. Prospecting is automated across 125,000+ contacts and leads, and 70% of initial inbound-lead outreach emails are now generated automatically before an SDR is involved.
More than 150 internal “Snowflake on Snowflake” deployments now exist. Read that alongside the headcount table above and the operating-margin story stops being an accounting choice and becomes a demonstration.
The company selling AI efficiency is the clearest proof point for it — and it is using that credibility in the room. Robins says he now has three to five conversations a week with CFOs; the buyer for this is no longer only the data team.
The flywheel, in one metric
Management’s entire framing is a flywheel: AI brings new workloads onto the platform; those workloads need a governed data foundation; customers who adopt AI then consume more of the core. The claim is easy to assert. The call offered one number that makes it harder to dismiss.
Snowflake tracks “time to 80% of purchased consumption” for each new-logo cohort — what fraction of, say, January’s new customers are consuming more than 80% of their purchased capacity roughly three months after they signed. Ramaswamy said this metric has “very, very visibly improved” for the newest cohorts. Customers are ramping to their committed spend faster than they used to, which is the cleanest available signal that AI is compressing time-to-value rather than simply pulling demand forward.
The supporting evidence lines up. Net revenue retention held at 126%. Gross retention has been “relatively flat” across several quarters. Use cases deployed grew 89% year-on-year, and use cases won per account executive grew 43% — demand and sales productivity rising together, which is an unusual combination.
Where the AI revenue actually comes from
“AI revenue” at Snowflake is not one product. It is a portfolio, and the two flagship agentic products are scaling fast.
| Product | Accounts | Growth | What it is |
|---|---|---|---|
| CoCo (Cortex Code) | >9,100 | +2,000 net new in the quarter | The coding/agent-building surface. Makes the platform — and Snowflake’s own sales team — AI-native. |
| CoWork | 5,800 | ~+11% quarter-on-quarter | Conversational access to enterprise data for knowledge workers; the “agentic control plane.” |
The strategic bet underneath the product names is that Snowflake becomes the governed control plane for the agentic enterprise — the layer that gives an agent both the context to understand a business and the authority to act on its behalf, with governance intact.
Two launches make that concrete. Cortex AI Gateway dynamically routes each task to the right model under customer-defined cost and governance policies, and through its integration of Natoma lets users send emails, summarise Slack threads and open Jira tickets without leaving CoWork or CoCo — insight to action. Cortex Sense captures the business definitions and institutional knowledge an agent needs and serves that context at the moment it answers.
Model neutrality, the second time around
Asked whether being model-neutral is a competitive advantage, EVP of Product Christian Kleinerman answered “absolutely, it is.” His reasoning is worth quoting in substance: many customers made large commitments to one model company and later wished they hadn’t; a commitment to Snowflake gives them the flexibility to switch, plus automatic routing to the right model for each task.
Ramaswamy framed it as history rhyming — the cloud-neutrality Snowflake pioneered, now replayed at the model layer, “deja vu all over again.”
The old split, where Anthropic sat mostly on AWS and OpenAI mostly on Azure, has broken down as the labs deploy their own capacity across clouds. Three points from the exchange matter for economics:
- Open-source inference is a margin lever. Snowflake runs it itself, so “we have pretty different economics” on open models — “a lot of potential for future optimization.”
- “Auto” mode pairs each task with the most appropriate model, giving Snowflake room to optimise cost on the customer’s behalf.
- Frontier prices keep falling. Kleinerman noted the frontier labs revise prices down regularly and keep adding models, which “kept costs somewhat in check relative to the usage.”
Snowflake’s own Arctic models stay in their lane: not frontier-class, but tuned for narrow, constrained tasks — AI functions, document processing, embeddings — where in-house models buy accuracy and efficiency. The takeaway for the model: today’s dilutive AI gross margin is not fixed. Auto-routing, open-source inference and falling frontier prices are all levers Snowflake has named. They are levers, though, not yet realised margin — which is exactly why the gross-margin line is the one to watch.
The guide, raised
| Guidance | New | Note |
|---|---|---|
| FY27 product revenue | $6.07bn (+36%) | Raised; management framed it as >500bps higher than the prior guide. ~1pp from Observe. |
| Q3 product revenue | $1.588–1.593bn (+37–38%) | — |
| FY27 non-GAAP product gross margin | 74% | Lowered — richer mix of lower-contribution AI workloads. |
| FY27 non-GAAP operating margin | 14.5% | Raised from 13.5%. |
| Q3 non-GAAP operating margin | 15.5% | — |
| FY27 non-GAAP adj. free cash flow margin | 23% | Reiterated. |
| GAAP profitability | On track for Q4 FY28 | Not this year. |
One nuance on the guide worth carrying forward: the raise this quarter was broad-based. Last quarter, management said most of the increase came from CoCo. This quarter, asked the same question by JPMorgan, Robins described it as a reflection of both AI-product strength and underlying core strength — because AI adopters increase their core consumption too. The composition of the raise, not just its size, is the tell that the flywheel is turning.
What could actually slow this
The bull case is well told above. The honest risk register, mostly assembled from what management would not quantify:
- Gross margin is the price of the mix. The faster AI grows, the more the 74% gross margin gives way. The offsets are real but unproven; the whole margin story leans on headcount restraint, which cannot compound forever.
- The consumption uplift is asserted, not sized. Asked directly for the uplift AI adopters show versus non-adopters, and how to know it is durable rather than pull-forward, management declined to share the numbers and answered qualitatively. The cohort metric is encouraging; the magnitude is undisclosed.
- RPO grew slower than revenue. RPO rose 30% against 37% revenue growth. Management flagged an increasing Q4 weighting of bookings as customers favour Q4 renewals, so the headline may understate momentum — but it also loads the year’s proof onto a single quarter.
- International was a non-answer. Asked about Europe and Asia, where Snowflake is under-indexed, the reply was that performance is “not region-specific” and all regions are doing well. That is a stated opportunity with no disclosed plan attached.
- GAAP profitability is still a year and a half out — Q4 FY28. The 15% figure that anchors the story is non-GAAP.
What we are watching
- Whether the gross-margin guide stops falling. The single cleanest read on AI unit economics. The named levers — open-source inference, auto-routing, frontier price cuts — either show up here or they don’t.
- Whether headcount stays restrained without breaking growth. Roughly 161 organic adds year-to-date against 935 a year ago is the engine of the margin raise. That is a ratchet, not a cushion.
- A fourth straight quarter of acceleration. Three in a row is already rare. A fourth would make the reacceleration a trend rather than a rebound.
- The cohort ramp metric getting disclosed. Management leaned on “time to 80% of purchased consumption” without a number. If it is as strong as described, quantifying it would be the most persuasive slide they could publish.
- An acceleration in NRR. The new cohorts fast ramp and old cohorts accelerating migration and workloads we will follow if Snowflake NRR accelerate further than 126%.
- Hybrid Tables and Postgres. Ramaswamy pointed to operational and system-of-record workloads — not just analytics — as the foundation for a new generation of agent applications built on Snowflake. That is the expansion of the addressable surface to track next.
Thank you!
CoreValue Research
The full quarter-by-quarter financial history sits in our Data Hub, and the complete company work-up is in Snowflake — Initiation 2026, available to PRO subscribers.
Independent educational analysis by CoreValue Research. Not investment advice, and no recommendation to buy or sell any security. All company figures are as reported by Snowflake on its fiscal second quarter 2027 earnings call of 2 September 2026 and in the accompanying materials. Non-GAAP figures are management’s; definitions and GAAP reconciliations are in the company’s investor presentation.
