Harvey's $550M round prices legal AI at 39 times the ARR it just crossed

Divide $400m of annual recurring revenue by the more than 3,000 organizations now using Harvey and the average customer is paying somewhere near $133,000 a year ^[2][2], a real line item in a law firm's technology budget and a long way from the roughly $5.2m of valuation each of those logos is carrying ^[3]. Eighty of the Am Law 100 are in the base already ^[2][4], so the twenty that are not are the thinnest growth pool the company has.

The multiple is where the cash question sits. Techtimes puts the price at about 39 times ARR ^[7]; at a conventional ten times, the same $15.6bn mark corresponds to roughly $1.56bn of ARR, which is 3.9 times where Harvey is now ^[7]. The six months since the $11bn round produced $4.6bn of new mark against $550m of new cash, about $8.40 of paper value per dollar in ^[1][5], and the more than $1.5bn Harvey has raised since 2022 is now 9.6% of what the company is said to be worth ^[5][6].

The round has a job beyond the mark. Harvey Tenet downloads open-weight parameters, trains them on Harvey's own infrastructure, and runs the result inside an architecture that never sends client content to an outside provider ^[12], starting from Kimi K3, the model Moonshot AI released in July 2026 ^[13]. K3 is a 2.8-trillion-parameter mixture-of-experts that activates 16 of its 896 experts per token, 1.8% of them ^[14][8], which is what keeps the inference bill from tracking the parameter count. So Harvey is buying GPUs and a training team rather than API capacity from OpenAI, Anthropic or Google, and techtimes argues the reason is doctrinal rather than economic: privilege can be waived when privileged material reaches an unauthorized third party, and a model API provider is a candidate for that role ^[9], with courts treating self-hosted deployment as the strongest protection available ^[10]. Harvey's promise not to train on customer data ^[11] narrows the exposure without removing the third party from the chain.

What the evidence does not establish is that 39 times is now the reference price for vertical AI generally. The source offers no comparable multiple from any other vertical AI round, only the assertion that few enterprise software companies have compounded value this fast, and a ladder from $3bn in February 2025 through $5bn, $8bn and $11bn to here, a 5.2-fold move ^[8][6][9]. That is a claim about scarcity of comparables, not evidence of a benchmark.

The risk I would underwrite against is not the moat failing but the moat holding while growth changes shape: with 80% of the Am Law 100 signed and 20% of the Fortune 500 ^[2], the next tranche of revenue comes from in-house legal departments and from selling more seats and workloads into firms Harvey already serves, and seat expansion inside partnerships whose compensation still runs on billable hours is slower arithmetic than new-logo land. Winston Weinberg, a former O'Melveny litigator, presumably knows that better than his investors do ^[15]. Priced here, Harvey has taken on the obligation to add about $1.16bn of ARR before its own entry price reads as ordinary software ^[10].