While everyone watches to see who releases the next model, the money is quietly going to those selling the raw material. Snorkel AI has raised 350 million dollars (around 315 million euros) in a Series E round, at an estimated valuation of 3.5 billion dollars. Seventeen months ago the same company was worth 1.3 billion - so nearly a tripling, in under a year and a half.

The round is led by Insight Partners and S32, with Lightspeed, Greylock, GV and Wells Fargo also participating, alongside new investors. The firm has seven years behind it, but the business it sells today is newer than that.

Snorkel started as software for automatically labelling data. Last year it turned towards something more profitable: delivering ready-made datasets - a model they themselves call data-as-a-service. The method is mixed: their own software and models generate synthetic data, and domain experts refine it.

The figure that explains the round is the revenue. The annual revenue run rate is now 375 million dollars - eighteen times more than twelve months ago. Eighteen times in a year is not product growth, it is market hunger. The artificial intelligence labs have hit the limit of what can be gathered off the internet for free, so now they buy data the way a factory buys steel.

Snorkel is not alone. Mercor has reached 2 billion dollars in gross annual revenue, Handshake 1 billion this year, Micro1 500 million gross. But there is a footnote here that rarely gets read: those firms usually hand 60 to 70 percent of the revenue straight to the domain experts. So the number on the front page and the number that stays in the company are two different worlds.

Snorkel does it differently - it sells reinforcement learning environments and ready-made datasets rather than human hours, so it books payouts to experts as cost of goods sold instead of deducting them from the revenue it reports. Same thing, different column in the table. Who reads that table before investing?

What remains is the familiar picture: a gold rush in which the surest money is made by those selling shovels. The models change every three months, but somebody has to make the data they learn from - and that somebody is now worth 3.5 billion dollars.