AfterQuery asks 10.7 times its last round's price for two to three times the revenue

The 10.7x is the easy number ^[18]; the split inside it is the informative one. The Series A's $300m against the $100m run rate Mateega cited in April works out to 3x ^[2][7][20]. The new round sets $3.2bn against a run rate the founder now places in the hundreds of millions, which taken at $200m to $300m is 10.7x to 16x ^[1][21]. So the price of a dollar of AfterQuery revenue rose somewhere between 3.6 and 5.3 times while the revenue itself doubled or tripled ^[22][26]. Most of the move is the market's view of the category rather than the company's execution, though the execution is what makes the view credible.

What the category sells is a payroll. AfterQuery pays doctors, lawyers, engineers and financial analysts to write out, step by step, how they reason through a hard professional call ^[10], because the labs have already picked over the usable text on the open web and synthetic material carries them only so far ^[11]. That is a cost of goods with a billing rate attached, valued here on software comparables. One of Forbes' sources says the company is already profitable ^[5], and that single unverified word is doing more work inside the $3.2bn than the growth rate is.

The comparables are moving too. Meta paid $14.3bn for 49% of Scale AI ^[16], which marks the whole of it near $29.2bn ^[23], and Mercor, at $10bn last October, is talking with Nvidia about $20bn ^[15], or 6.25 times AfterQuery's number ^[24], leaving AfterQuery at roughly 11% of Scale's implied whole ^[25]. Nvidia sits on both sides of that, consuming AfterQuery's output for its open-source Nemotron models ^[12] while negotiating the competitor's equity. Customers who fund their suppliers get good prices on the supply, right up to the point they decide to make it in-house.

This is probably wrong, but the scarcity being priced is real and dated. Maybe the labs just keep buying judgment, and 16x looks cheap in a year. Maybe they internalise expert networks instead, having learned from these invoices exactly what to pay for. Or maybe reinforcement-learning environments and generated data close enough of the gap that paid specialist hours become a premium tier rather than the input. Note what AfterQuery gave up to get here, since it wanted to build finance agents and stopped ^[13], so the money and the founders' attention now go into screening submissions for the right difficulty and pre-training on the data to prove it moves a model ^[14]. (The source has them founding in February 2025 and also founding 18 months after joining the Winter 2025 batch ^[8][9], which cannot both be true, so read the fastest-ever record as a claim rather than a measurement.) If a Nemotron-class model trains as well on generated environments as on credentialed hours, the 16x is a staffing multiple, and the arithmetic runs backwards from there.