Snorkel books a $375m run rate a year after switching from tooling to finished datasets

Two human reviewers score the same model answer differently. In Snorkel AI's account of its reinforcement learning work, that gap usually means the written evaluation criteria are inconsistent, so the criteria get rewritten ^[8]. The rubrics can run several pages. For a programming task, the guidance has to cover every cybersecurity and performance requirement the generated code must meet ^[7].

That document is now part of what the customer buys. Snorkel's first product was Snorkel Flow, software that cut the labeling work in supervised learning projects using statistical methods its founders developed at the Stanford AI Lab ^[4]. Last year the company began selling ready-to-use datasets instead, alongside the rubrics and the virtual environments a model trains in, such as a simulated developer workstation for a code generation model ^[5][9]. Tens of thousands of human experts generate the reinforcement learning tasks ^[6]. TechCrunch reports the approach is hybrid, with Snorkel's own software and models producing data synthetically next to the subject matter experts ^[15].

Investors put $3.5b on a $375m annualized run rate, about 9.3 times revenue ^[1][1]. "Since launching our new data-as-a-service offering nearly a year ago, we've grown over 18 times, and this week crossed an annualized revenue run rate of $375 million," co-founder and chief executive Alex Ratner said in a blog post ^[10]. Seventeen months ago, a $100m Series D valued the company at $1.3b ^[11].

Anyone pricing this category has to ask where the labor shows up. Mercor's gross annualized revenue has reached $2b, Handshake passed $1b earlier this year, and TechCrunch reported Micro1 at $500m ^[12]. Those companies hand roughly 60% to 70% of top-line income straight to the domain specialists. Net revenue lands well below the headline ^[13]. Run that range against $375m and $225m to $263m would go out to experts, leaving $113m to $150m ^[2]. Snorkel told TechCrunch that it sells environments and completed datasets, and that it does not sell human labor, so expert payments sit in cost of goods sold instead of coming off the reported run rate ^[14]. TechCrunch's report does not include the size of those payments ^[17].

Teams buying this tell themselves they are buying data, and data is inputs. They are also moving the definition of a correct answer to a supplier. Does the evaluation rubric leave with the vendor when the contract ends, and can your own staff reproduce a score on a sample the vendor already graded? If the answer to both is no, the rubric is the product you rented, and next year's renewal is priced against a standard you cannot rebuild in-house.

Snorkel says the new money goes to hiring engineers, AI safety initiatives and support for open-source model evaluation benchmarks ^[16].