The Idea:
In the mid-2010s, Alex Ratner was a computer science PhD student at Stanford, working under Chris Ré, a professor who had already sold an earlier AI company, Lattice Data, to Apple. The pair noticed that even Google and Microsoft were struggling with the mountains of hand-labelled training data their models needed. Doctors were being asked to tag X-rays one at a time. Ratner thought making experts label 10,000 data points by hand was a ridiculous waste of what they knew.
Their fix was simple. Get experts to write down their rules of thumb, then let software apply them at scale. In a trial with Stanford Hospital, labelling that had taken person-years was done in hours. The Snorkel research project launched at the Stanford AI Lab in 2015. After four years deploying it with Google, Intel, Apple and DARPA, Ratner, Ré and fellow researchers Paroma Varma, Braden Hancock and Henry Ehrenberg spun it out. The pitch was blunt: replace the armies of human labellers.
This was the beginning of Snorkel AI.
The Execution:
- July 2020: Snorkel emerged from stealth with $15M from Greylock, GV and In-Q-Tel, launching its platform Snorkel Flow with paying customers including top US banks and government agencies.
- April 2021: Lightspeed led a $35M Series B, with GV, Greylock, In-Q-Tel and BlackRock joining.
- August 2021: Four months later, Addition and BlackRock co-led an $85M Series C at a $1B valuation, taking total funding to $135M and making Snorkel a unicorn in two years.
- May 2025: Then it stalled. Nearly four years on, Snorkel raised a $100M Series D at $1.3B and unveiled Expert Data-as-a-Service. A 30% bump in four years, in the middle of an AI boom.
- June 2025: The market it was pivoting into got a price tag when Meta bought 49% of Scale AI for $14.3B.
- September 2025: The pivot got painful. Snorkel cut 13% of its staff, 31 of 240 people, mostly software engineers, as it went all in on selling data.
- 2025 to 2026: The company built to replace human labellers now runs a network of tens of thousands of specialists in coding, law and medicine, selling finished datasets rather than billing for their hours.
- September 22, 2026: Snorkel raised a $350M Series E at $3.5B, co-led by Insight Partners and S32. Its data business had grown over 18x in a year to a $375M run-rate, up from roughly $20M twelve months earlier.
- Today: Snorkel works with frontier labs, hyperscalers and US government agencies, and says it expects to turn profitable in 2026.
The lesson?
Snorkel spent six years building software to take humans out of data labelling. The breakout came when it put the experts back in and charged for the finished product, not the hours. The thesis that AI is a data problem never changed. The business model did, and it cost 31 jobs and four flat years to find the right one. Hold your thesis tightly and your product loosely.