Snorkel AI Raises $350M as Expert Costs Shape Growth

On September 22, 2026, Snorkel AI announced a $350M financing at a $3.5B valuation in a San Francisco-datelined release. Insight Partners and S32 co-led the round. Snorkel plans to use the money to expand its capacity to produce data and simulated environments for advanced AI systems.

TechCrunch’s Series E report puts Snorkel’s annualized revenue run rate at $375M and describes its shift from data-labeling software to completed datasets. It also reports that payments to the human experts who help make those products are recorded in cost of goods sold. The sales pace is substantial, but the disclosed figure does not show how much gross profit remains after delivery.

Snorkel now sells the finished data

Snorkel’s earlier software helped customers automate data labeling. With data as a service, the customer instead buys a completed dataset or a reinforcement-learning environment: a setting designed for an AI system to practice tasks or be evaluated. Snorkel must assemble and check that output before delivery. The distinction matters because the price of a finished product reflects production work as well as the technology used to perform it.

In co-founder and CEO Alex Ratner’s account, Snorkel grew more than eighteenfold after introducing the new offering nearly a year before the financing. He describes a production process that combines subject-matter experts with specialized AI models and agents. For coding-agent tasks, experts help define the work while automated systems expand initial inputs, direct review toward likely model errors and carry out quality checks.

That process could let each expert contribute to more finished work, but it does not remove expert judgment from the product. Complex tasks need suitable goals, grading criteria and review; otherwise a dataset can be large without being useful for training or evaluation. The business therefore has two linked questions: how much demand exists for these outputs, and how efficiently Snorkel can produce them at the quality customers require.

The run rate measures pace, not a completed year

An annualized revenue run rate extends a recent sales pace across a full year. It describes the current scale of a fast-growing business, not revenue already earned over the preceding year. If sales rose sharply after Snorkel changed its offering, the annualized figure could be far above its historical revenue. Keeping that pace would still require future orders and the capacity to deliver them.

The public disclosures do not separate recurring software revenue from sales of finished data products. That missing split matters because the two kinds of revenue may have different cost patterns. A software subscription can continue without rebuilding the product for each customer. A dataset or environment may require new expert input, tailored checks and revisions, even when Snorkel reuses its underlying tools.

Nor does a run rate establish how much work comes from repeat customers rather than new projects. A customer may buy further datasets, but the disclosed figure does not show the frequency, duration or economics of those purchases. Treating every dollar in the annualized pace as though it were a software subscription would assume a level of repeatability that Snorkel has not publicly quantified.

Expert payments reduce gross profit

Snorkel sells datasets and environments rather than simply matching customers with expert labor. Its reported revenue is the sales figure for those products; payments to experts sit in cost of goods sold. They are not money left over after the revenue figure is calculated. A reader therefore cannot treat the run rate as the amount available to cover research, sales staff and other operating expenses.

Gross profit is revenue after the direct costs of delivering what was sold. If more sales require a similar rise in paid expert work, revenue can grow rapidly without gross margin improving. If Snorkel’s automation allows experts to build and review more output for a smaller increase in cost, margins could improve. The public revenue figure cannot determine which pattern is emerging because Snorkel has not disclosed expert costs or gross margin.

This is also why a comparison with a software company needs care. Both businesses can report impressive revenue growth, yet the direct work required to serve another customer can differ. Snorkel’s hybrid method may make its data operation more scalable than a service built entirely on human labor. That is a plausible benefit of the method, not a measured financial result in the information released with the round.

The valuation leaves the margin question open

Dividing Snorkel’s valuation by its stated annualized revenue pace produces a figure of roughly 9.3 times. That is a valuation-to-run-rate comparison, not a multiple of completed annual revenue or profit. It says little about the cost of producing the datasets behind the pace. Without a revenue breakdown and gross-margin figure, the valuation cannot establish that the newer business has the economics of a software subscription.

The financing gives Snorkel money to expand production and research while it pursues demand for more specialized AI data. What remains undisclosed is the portion of sales that recurs, the direct cost of expert-led delivery and whether automation is improving margin as volume rises. Those figures would show more clearly how much of the revenue surge can translate into durable profit.