Snorkel's $3.5bn price rests on an unaudited run-rate that grew 17.5 times in a year

What a frontier lab buys here is a task with a grader attached. Snorkel says its platform pairs human specialists with thousands of narrower AI models and agents, and that experts in fields including coding, law and medicine write the scenarios and the evaluation standards while automated systems generate, check and refine the resulting data ^[8][9]. Building a reinforcement-learning environment for a coding agent takes knowing the behavior the lab wants, the failure modes it needs to catch and the signals used to reward the model ^[10]. Snorkel says coding is among its largest areas of demand ^[19].

Anyone weighing a bought environment against an in-house harness is buying someone else's definition of done. When the pass condition in a vendor's grader differs from the one your own users would accept, the score it produces is a claim about the vendor's workload. Snorkel's stated customers are frontier labs, hyperscalers, enterprises and the US federal government ^[18], and the account of the round gives the $350 million run-rate as one figure, with no product-line split ^[20].

On the stated figures the price is about ten times run-rate ^[12]. Snorkel says annualized run-rate has passed $350 million, up from roughly $20 million a year earlier, a 17.5-fold increase ^[6]. That figure is a company-reported annualized snapshot. No audit stands behind it ^[7]. The valuation went up 2.69 times from the $1.3 billion set in May 2025 ^[4][1]. Run-rate grew 17.5 times across twelve months while price grew 2.69 times across sixteen, so each dollar of reported run-rate costs a new investor roughly 6.5 times less than it did in May ^[2]. The two windows are different lengths, so treat 6.5 as an approximation.

The raise and the stated run-rate are both $350 million. That is a coincidence ^[3]. Snorkel sells the datasets and environments its specialists help produce, billing for those outputs instead of units of human labor ^[13], and draws on a network of tens of thousands of specialists, according to Ratner ^[14]. A conventional annotation provider adds revenue by recruiting and managing more workers ^[15]. Snorkel's position is that software absorbs quality assurance and the other work that would otherwise grow headcount alongside revenue ^[16].

Ratner told Reuters that valuable training data will keep requiring human input, and that synthetic and automated methods will be needed to produce it at sufficient scale ^[17]. Working inside a lab's training loop is a deeper relationship than a software license, and it exposes Snorkel to the spending cycles of frontier labs ^[11]. Ratner plans to spend the round on researchers and engineers, enterprise and government operations, and support for third-party model evaluations ^[22].