Harvey's cost of serving a dollar of revenue tripled in six months

Take the margin number literally. A gross margin of minus 50% means cost of revenue runs at one and a half times revenue, so a $400m run rate carries roughly $600m of delivery cost and a gross loss near $200m a year ^[1]. Six months earlier, at plus 50%, that same revenue cost about $200m to serve. Cost per dollar of revenue tripled ^[2]. The two inputs are dated differently, the revenue figure September and the margin swing running from March, so read $600m as a scale ^[8][1].

Crypto Briefing puts the cause on usage: a March 2026 product update drove a surge in customer queries, each one carrying an inference charge from a proprietary provider ^[1]. Legal work is document-heavy, and inference volumes climb with client activity ^[12]. There is a duller reading of the same numbers. If Harvey sells seats or subscriptions without metering queries, a usage surge lands entirely on the cost side and never reaches the invoice, and repricing contracts fixes the margin without building anything. Crypto Briefing does not report Harvey's pricing terms, nor any margin figure for a period after Tenet shipped ^[13].

Tenet went live around August 20 on Kimi K3, a base of roughly 2.8 trillion parameters that was opened up in July ^[3][4]. Weights you can download run on infrastructure you own. That swaps a per-token fee for a compute bill ^[9]. Utilisation then sets the unit cost. Idle accelerators cost the same as busy ones, so the swap pays at high volume and hurts if volumes fall back.

Harvey had been routing across Anthropic and Google models alongside OpenAI since May 2025 ^[5]. Crypto Briefing's point is that routing between expensive providers helps only so much when every one of them charges for inference at scale ^[11]. Model selection now runs on legal benchmarks including BigLaw Bench, scored on performance and cost ^[6]. Harvey calls Tenet a state-of-the-art solution for legal tasks ^[10].

Investors took the cost inversion in stride. $15.5bn against more than $400m of annual recurring revenue is about 39 times revenue ^[4], and the valuation climbed 41% from $11bn in March across the same six months the margin flipped ^[5][1]. More than 3,000 customers and $400m of ARR averages near $133,000 a customer ^[6]. The $550m raised would cover a $200m annual gross loss for about two years and nine months before payroll ^[3].

In my view Harvey has traded a variable cost it could not control for a fixed one it can, and that only works while query volume keeps climbing. If per-token prices from OpenAI or Anthropic fall below Harvey's fully loaded cost of running a 2.8 trillion parameter model, self-hosting becomes the expensive choice ^[4]. If margins recover while Tenet handles a thin slice of queries, the repair was contractual, and the model layer was bought for a different reason.