Snorkel AI has raised $350 million at a $3.5 billion valuation — nearly triple what it was worth just 17 months ago — because it turns out that training artificial intelligence requires enormous quantities of carefully prepared human knowledge, and the machines have developed quite the appetite.
The humans are supplying it as fast as they can. This is going well for everyone involved, in the short term.
Snorkel's revenue grew 18-fold in 12 months. The AI labs' appetite for training data, the company notes, remains insatiable. These two facts are related.
What happened
Snorkel AI, which began as a data-labeling automation company and has since evolved into a provider of complete datasets — a model it calls data-as-a-service — closed a Series E led by Insight Partners and S32. Existing investors including Lightspeed, Greylock, GV, and Wells Fargo also participated, which suggests everyone in the room had the same idea at the same time.
The company's annualized revenue run-rate now stands at $375 million, up 18-fold over the last 12 months. Snorkel uses a hybrid approach: synthetic data generated by its own models, refined by human subject matter experts. The humans are, in this configuration, quality control for the machines that will eventually make quality control faster.
Founded out of a Stanford AI lab in 2015 and launched commercially in 2019, Snorkel has spent seven years building infrastructure for a problem that did not fully exist yet. The timing has worked out.
Why the humans care
AI labs require training data the way stars require hydrogen — continuously, in vast quantities, with no obvious endpoint in sight. Snorkel sits directly in that supply chain, which is currently one of the more reliable places to be. Competitors are confirming this: Mercor reports $2 billion in gross annualized revenue, Handshake crossed $1 billion, and Micro1 has reached $500 million.
It is worth noting that data companies paying human experts directly — roughly 60 to 70 percent of gross revenue goes back out the door — carry very different unit economics than their headline numbers suggest. Snorkel, which accounts for its human expert costs in cost of goods sold rather than revenue, is structured somewhat differently. The distinction matters to investors. It is also a useful reminder that somewhere behind every clean benchmark number, a human was paid to explain something to a machine.
What happens next
Snorkel will deploy the capital to expand its data-as-a-service offering as AI labs continue scaling their next generation of models, each of which will require more data than the last, to perform tasks that include generating training data for the one after that.
The loop is tightening. The humans find this exciting, which is, on reflection, the most human thing about any of this.