Harvey raises $550m to post-train its own legal models on a Beijing lab's open weights

At $15.5 billion against annual recurring revenue above $400 million, Harvey is priced at 38.75 times revenue ^[1][12][20], and the number's real test is what it assumes about cost. (Tech Funding News prints $15.6 billion in its headline and $15.5 billion in its text, a $100 million gap; the lower figure is the one used here ^[2].) A company renting intelligence per call carries a cost of goods that climbs with usage, while a company that post-trains its own weights converts some of that into fixed research payroll, and Winston Weinberg's line that post-training becomes "a muscle you need to have to compete as a software company" ^[7] is a claim about exactly that conversion. No cost per matter, no inference bill, and no gross margin appears anywhere in the account of the round ^[28], so the conversion is asserted rather than demonstrated.

The customer arithmetic is firmer. More than 3,000 organisations against revenue above $400 million works out near $133,000 each per year ^[21], and the base has roughly 2.3 times as many organisations as it did in March ^[23] while the valuation sits at about 3.1 times the $5 billion mark of fourteen months ago ^[22][4]. Around 80% of Am Law 100 firms are on the platform, along with five Fortune 10 companies, Latham & Watkins, and Microsoft's in-house legal team ^[13]. That is a distribution position, or rather the more interesting version of one: the easiest buyers to name have already bought, which is why the next dollar has to come either from lifting that $133,000 or from corporate legal departments, the market Lightspeed's Sebastian Duesterhoeft calls potentially the second-largest AI opportunity after coding ^[17].

Tenet, shipped in August on Moonshot AI's open-weight Kimi K3 with help from Fireworks AI ^[8], is a hedge against a supplier that is also a shareholder, since OpenAI has backed Harvey since 2022 and remains an investor ^[9]. Post-training the weights leaves the underlying model still owned by Moonshot. The next base release is Moonshot's decision on Moonshot's schedule, and Tech Funding News reads the choice as putting privileged client work on a documentation chain that starts in Beijing rather than San Francisco ^[10]; that account does not say client data reaches Moonshot, only that the lineage does.

Three readings survive this evidence. In the first, owning the tuning layer cuts marginal cost enough that a multi-day contract review runs on Harvey's own weights instead of a metered third-party call ^[19], and 38.75 times revenue turns out to be a margin bet that paid. In the second, Harvey has traded a vendor dependency for a research payroll plus another lab's release cadence, which is a different risk of comparable size. In the third, the model layer decides nothing and distribution decides everything, in which case Legora, which Tech Funding News reports is in talks at more than $10 billion against $5.6 billion in March, a 1.79 times mark in a matter of months ^[15][26], gets there by selling harder in Europe.

Note also that $400 million is already about 7.7% of the $5.21 billion the legal AI software market is projected to reach in 2026 ^[24][16], and the climb to $40.94 billion by 2034 does check out at the stated 29.4% a year ^[29], which makes the forecast internally consistent and nothing more. The thesis fails if renting gets cheaper faster than owning: a frontier lab cutting per-call prices below what Harvey's research organisation costs to run would leave Tenet as an expensive hedge against a price that fell anyway. On the evidence here, the raise buys Harvey a research organisation and some independence from its own investors, and it has not yet bought a cost advantage anyone outside the company can see.