Snowflake Didn't Invent Splitting Storage From Compute. It Did Something Harder.

Every data warehouse ever shipped before 2015 made one quiet, expensive assumption: that the disks holding your data and the processors crunching it had to live in the same box. Buy more storage, you bought more compute you didn't need. Need more compute for one heavy query, you paid to scale the storage too. The two were welded together, and the weld was the business model. Snowflake's founders looked at that weld and decided not to cut it. They decided to build a system that had never had one.

The story you'll hear is that Snowflake invented the separation of storage and compute. That's the part that's wrong. Snowflake's own academic paper — the one it presented at the field's most prestigious conference — frames the idea as something it found lying in the public cloud, not something it dreamed up. The genius wasn't the concept. The genius was the willingness to start from zero.

The idea was free. Almost nobody could afford to use it.

When Amazon's S3 and EC2 went mainstream, they handed the industry a gift few recognized: storage you could rent by the gigabyte and compute you could spin up and kill by the hour, as two completely independent meters. The architectural insight that you could disaggregate the two was, by then, not new — researchers had been circling it for years. Snowflake's SIGMOD 2016 paper, 'The Snowflake Elastic Data Warehouse,' is admirably honest about this: it describes building a new warehouse design on top of already-available public cloud resources, not on a breakthrough the company alone possessed.¹⁸ The platform that resulted had been available since June 2015.¹

So if the idea was sitting in plain sight, why did three founders building from a blank editor get a decade's head start? Because the incumbents could not pick the idea up without dropping the product they were already selling. That is the whole game.

Why an incumbent can't simply un-weld its own product

Here is the mechanism, worked all the way down. A traditional warehouse — Redshift, Oracle, Teradata — keeps data on the same nodes that run the queries. That coupling isn't a bug to those systems; it's the foundation every other assumption rests on. The query planner assumes the data is local. The pricing assumes you scale both together. The codebase, written across years, is shot through with the premise. To separate storage from compute, an incumbent doesn't tune a setting — it rewrites the engine, and in doing so admits the architecture it has been selling is the problem. Snowflake had no such admission to make. It had no installed base to protect, no on-premises customers to keep happy, no legacy planner to honor. It could build a three-layer system — storage on commodity object storage, independent compute clusters, and a separate cloud-services layer holding the metadata — because it had nothing to preserve.¹ The blank page wasn't a disadvantage. It was the moat.

The founders weren't generalists who stumbled onto cloud. Benoit Dageville and Thierry Cruanes came out of the deep end of database engineering, and the third founder, Marcin Zukowski, brought something the Oracle veterans didn't have: he came not from a database vendor but from CWI, the Dutch national research institute, where the X100 project spun off into VectorWise — a company eventually acquired by Ingres, later rebranded Actian.⁴⁵ Zukowski left in 2012 to join Dageville and Cruanes, carrying vectorized query execution and lightweight compression into the new system.⁴ The popular telling that all three were Oracle alumni misses the point: the engine got its speed from a research lineage that had nothing to unlearn about the cloud.

Isn't this just a clever repackaging of S3?

The fair objection is sharp: if the idea was free and the cloud did the heavy lifting, isn't Snowflake just a thin wrapper around Amazon's storage — a middleman taking credit for plumbing it didn't build? It's the strongest case against the company, and it deserves a real answer. The answer is that the hard part was never the disaggregation; it was making the disaggregated parts behave like one coherent, fully managed system. Anyone can store files in object storage. What's hard is a multi-tenant SaaS warehouse that resizes compute on demand, holds a consistent view of metadata in a separate services layer, and lets dozens of independent compute clusters read the same data without stepping on each other — all while a customer just runs SQL and never touches a knob.⁸ That coordination is the product. The honest counter to the 'thin wrapper' charge is that wrappers don't get written up at SIGMOD, and they don't take a decade to be convincingly copied. Snowflake's later filings make the same point in reverse: the platform now reads data whether it sits inside Snowflake or in external open formats like Apache Iceberg, because the value lives in the coordinating layer, not in owning the bytes.³

The market eventually agreed, loudly. When Snowflake went public in September 2020, it priced at $120 a share, opened at $245 — more than double — and closed up roughly 112% on the day, the largest software IPO on record at the time.⁶ Berkshire Hathaway and Salesforce each agreed to buy $250 million of stock at the IPO price in a concurrent private placement, and Berkshire took on millions more shares besides.⁶⁷ Warren Buffett, who famously avoids the technology he doesn't understand, wrote a check for an architecture diagram. That's how clearly the bet had paid off.

Snowflake's whole rebellion fits in a single sentence: it refused to inherit the weld. The incumbents weren't blind to the separation of storage and compute — they were imprisoned by the products that couldn't survive it. So the company built the one thing a market leader structurally cannot build: the system that assumes its own architecture from the very first line, with no past to apologize for. The idea was free. Starting over was the price. Snowflake was the only one willing to pay it.