Snorkel AI raised a $350M Series E in September 2026 at a $3.5B valuation, led by Insight Partners and S32. The valuation was approximately 9.3 times its $375M annualized revenue, reflecting investor demand for businesses supplying specialized training data to AI labs.

The financing followed a $100M Series D in May 2025 at a $1.3B valuation, led by Addition. Existing investors included Greylock and Lightspeed, while BNY and QBE Ventures participated as strategic investors, connecting Snorkel's enterprise data-development platform to buyers in banking and insurance.

Alex Ratner, Henry Ehrenberg, Braden Hancock, Paroma Varma, and Chris Ré founded Snorkel AI in 2019 out of research at the Stanford AI Lab. The initial business focused on programmatically generating training labels from expert knowledge rather than requiring people to annotate every example individually; that technology became the foundation for its broader data-development offering.

Snorkel AI develops specialized training datasets and evaluation environments for AI labs and enterprise teams. A customer begins with a capability gap—for example, an agent that fails on complex terminal tasks or a model that mishandles specialized financial reasoning—and works with Snorkel to define the tasks, acceptance criteria, and evaluation rubrics.

Snorkel assembles contributors with relevant professional or academic expertise to create examples, reasoning traces, preference judgments, or other training signals. Its software and models support synthetic generation and automated checks, while expert reviewers validate correctness, resolve disagreements, and check consistency against the task specification. The result is a dataset with traceable quality decisions, rather than an unstructured collection of contractor outputs.

Customers receive training data, evaluation tasks, or runnable environments that let them test performance across model versions. Snorkel also organizes task collections into Data Series covering areas such as agentic coding, multimodal reasoning, and computer-use workflows, with difficulty tiers and evaluation slices designed to expose specific model failures.

The approach builds on Snorkel Flow's programmatic data-development technology. In that software workflow, domain experts encode reusable labeling logic, combine imperfect supervision signals, and revise the data as model errors emerge, reducing reliance on example-by-example manual annotation.

Snorkel also builds custom AI agents for enterprise customers. It evaluates these agents against task-specific rubrics and programmatic checks, using the same expert-data development and adjudication methods employed in its frontier-model work. This offering carries the company beyond supplying model inputs into designing and testing systems that perform business workflows.

Snorkel sells completed datasets and reinforcement-learning environments rather than simply connecting AI labs with contractors. Buyers pay for deliverables built to their task specifications, and Snorkel accounts for payments to human experts in cost of goods sold. Its economic return depends on the difference between what customers pay and the costs of expert contribution, synthetic generation, research, quality assurance, and delivery.

Programmatic tools can improve those economics by reusing task specifications, automating checks, and directing expert attention toward difficult examples. The tradeoff is that specialized work still requires expensive domain expertise and reviewer calibration: a high-value dataset for complex reasoning can command a larger contract, but it also demands more work than commodity annotation.

The company has also sold enterprise data-development software through Snorkel Flow and now offers custom agent development. Software access and reusable tooling can support ongoing relationships, while expert-data and agent engagements add implementation and delivery work. These products have different cost structures, so the $375M annualized revenue figure should not be treated as a pure-software ARR metric.

Revenue can expand as a model customer moves from one capability gap into additional domains or new releases requiring different training signals. Reusing experts, task design, and validation systems across engagements could lower incremental delivery costs, although customer-specific tasks can require new research and review.

Snorkel AI competes with Scale AI and Surge AI for training-data and evaluation budgets, with expert-workforce providers such as Mercor for specialist contributors, and with enterprise data tools and internal teams for data development work. Its advantage depends on delivering useful training data and reliable environments with less customer coordination.

Scale AI and Surge AI compete for model-training and evaluation budgets, while Mercor and other expert-workforce providers compete for access to specialized human contributors. Snorkel's distinction is selling complete data and environment deliverables built through a combination of experts, synthetic generation, and programmatic quality control, rather than relying solely on the number of contractors it can supply.

Large data vendors can offer broad delivery capacity, while expert marketplaces can recruit across many disciplines. Snorkel must show that its task design, research methods, and validation produce better training outcomes or more reliable evaluations to justify a customer choosing it over a cheaper labor supply or an established annotation vendor.

Labelbox, SuperAnnotate, and open-source labeling tools such as Label Studio give enterprise teams ways to develop training data using their own workers and workflows. Cloud platforms also have a distribution advantage because customers already store data and run models in their infrastructure. Snorkel's reusable supervision and evaluation methods offer a way to reduce manual work, but buyers can combine other tools with in-house subject-matter experts.

Frontier labs have a further option: developing their own datasets, graders, and environments. Snorkel's opportunity is strongest where an external provider can assemble expertise and establish reliable evaluation criteria faster than an internal team, while its defensibility depends on continuing to improve that process as customers become more capable themselves.

Custom enterprise agents give Snorkel a path into operational automation and systems-integration budgets beyond the training-data spending controlled by AI research teams. The company already offers to build specialized agents and test them against business-specific tasks, providing an entry point to departments that want a working system rather than a dataset or development tool.

To capture those budgets, Snorkel must translate its expertise in task design and evaluation into reliable execution inside enterprise workflows. That could increase the value of each engagement, but deployment, integration, and ongoing maintenance also bring service obligations beyond producing training data.

Services delivery costs: Snorkel's completed datasets require credentialed contributors, quality review, and research work, with expert payments recorded in cost of goods sold. If contract prices do not keep pace with task complexity and delivery costs, revenue growth may produce less gross profit than the annualized topline suggests.

Frontier-lab dependence: Selling sophisticated training data to frontier AI labs exposes Snorkel to purchasing decisions from a relatively small pool of buyers that can also build internal data operations. Budget shifts, procurement pressure, or greater internal production could create volatility even if demand for AI training data continues to grow overall.

Evaluation reliability: Snorkel's differentiation rests on expert-created tasks, rubrics, and programmatic checks producing reliable measures of model capability. Systematic labeling errors or evaluations that fail to generalize to real tasks could weaken training outcomes, force costly rework, and damage customer trust.