Snowflake's Moat Is Real. It's Also Quietly Draining, One Quarter at a Time.

Here is the single most expensive sentence in the Snowflake bull case: 'It pioneered the separation of compute and storage.' It did build that architecture, elegantly, and for a few years it was the reason customers switched. Then BigQuery did it, and Redshift did it, and Databricks did it, and now virtually every modern data platform does it. The thing that was supposed to be the moat became the price of admission.⁸ So the real question isn't whether Snowflake is a good product. It plainly is. The question is what — if anything — still stops a customer from leaving.

The official story is that Snowflake is a fortress: sticky, expanding, protected by a network effect nobody can copy. The financials tell a quieter, harder story. The most important number Snowflake reports is net revenue retention — how much more last year's customers spend this year — and that number has done something a fortress is not supposed to do.

The one number that tells the whole story

Net revenue retention is the moat's heartbeat. When existing customers keep spending more — pouring in new data, running more queries, attaching more teams — the number climbs above 100% and the business compounds without selling anyone new. Snowflake's hit 158% in early 2023, one of the highest figures any software company has ever posted.¹ It looked like proof of an unbreakable hold. Then it started falling, and it has not stopped. 151%, 142%, 135%, 131%, 128%, 127%, 126%, and 125% by January 2026 — eight straight readings down, every one of them lifted directly from an SEC filing.¹ That is not noise. That is a two-year, unbroken line, and what it measures is the rate at which the moat refills. The water is still coming in. It is just coming in slower every single quarter.

None of this means the company is failing. Revenue still grew 29% in FY2026 to $4.7 billion, with 13,328 customers and 790 of the Forbes Global 2000 on board.⁵ A draining moat and a growing castle can coexist for a long time. But the bull narrative isn't 'great company.' It's 'protected company.' And protection is exactly what that falling line calls into question.

What actually holds a Snowflake customer in place

Strip away the architecture marketing and two real defenses remain. The first is data gravity: once a company has loaded years of business-critical data into Snowflake, built its dashboards on it, trained its analysts on it, and wired its pipelines into it, moving out is a migration project measured in quarters and fear, not an afternoon. The second is the Snowflake Marketplace, where one customer can share live data with another without copying it — a setup that, in theory, gets more valuable the more participants join. That's a genuine network effect, the kind that rivals can't simply clone with a feature.

Notice what's happened to the second row. The pioneer myth doesn't survive contact with the calendar: AWS Redshift launched in 2012, the very year Snowflake was founded by Benoît Dageville, Thierry Cruanes, and the often-forgotten third co-founder Marcin Żukowski — and Snowflake didn't even leave stealth until October 2014, with 80 customers.⁷ Snowflake didn't invent the category. It executed inside one that already existed. Architecturally and commercially brilliant, yes. First, no.

Why Snowflake is loosening its own grip

The strangest part of the story is that Snowflake is dismantling one of its own load-bearing walls on purpose. Its strongest hard lock-in was the proprietary storage format — your data lived in Snowflake's house, in Snowflake's filing system, and leaving meant rebuilding. Then Snowflake adopted Apache Iceberg, the open table format, letting customers keep their data outside Snowflake entirely and even process it on a competitor's engine like Databricks.⁹ Read that again. The company deliberately handed customers the exit door it had spent years welding shut. Analysts called it exactly what it is: a move that erodes the moat.

Why do it? Because the alternative was worse. Customers were going to demand portability with or without Snowflake's blessing, and a vendor that fights open standards in a world moving toward them ends up looking like the thing you migrate away from. The bet is subtle: trade the hard lock — 'you physically can't leave' — for a soft one — 'you could leave, but our compute and ecosystem are good enough that you won't want to.' It's a more honest moat. It's also a thinner one, resting now on preference rather than captivity. And preference is exactly what a falling retention number measures.

The fair case that the moat is widening, not draining

The honest objection is that I'm reading decline into maturation. Of course retention falls from 158% — a company can't sustain that rate as it scales, because the base it's expanding against gets vast. The same FY2026 with 125% retention also showed RPO climbing and the count of million-dollar customers reaching 733, with the Global 2000 share of business growing.⁵⁴ A platform with 13,000 customers and 790 of the world's biggest companies, deepening its hold on its largest accounts, doesn't look like a leaking ship. And the Marketplace network effect, whatever its exact scale, is the one defense that genuinely gets harder to copy the longer it runs.

All true. But two things cut against it. First, the Marketplace's strength is asserted, not proven — the impressive listing counts come from third-party narratives, not from a hard metric in Snowflake's own filings about cross-customer sharing revenue or lock-in. A network effect you can't measure is a network effect you can't bank on. Second, the structural problem doesn't soften with scale: Snowflake runs on AWS, Azure, and GCP — and those three are simultaneously its landlords and its competitors, selling Redshift, Fabric, and BigQuery against it.⁸ You don't get a durable moat by paying rent to your rivals. The compression is the symptom; the co-opetition trap is the disease.

There's a final irony worth sitting with. Frank Slootman set a target of $10 billion in product revenue by fiscal 2029, back when retention was still near its peak.¹⁰ He's gone now, replaced as CEO, and the goal has never been formally reaffirmed by his successor.¹⁰ Snowflake may well hit it — 29% growth off a $4.7 billion base gets you most of the way there.²⁵ But that number was set in the world of 158% retention, a world that no longer exists. The destination might survive. The math that justified it is draining a point or two every quarter.

So what actually protects Snowflake? Not the architecture — that's commodity now. Not the first-mover crown — it never wore one. Not the storage lock — it gave that away. What's left is data gravity that weakens the moment you open the gates, and a Marketplace network effect that may be everything the bulls claim or merely a good story nobody has put a number on. The moat is real. It is also the only kind of moat that gets harder to defend the better and more open your product becomes. Snowflake didn't lose its wall. It chose to lower it — and now we get to watch, one filing at a time, whether the water it spent a decade collecting was ever really its own.