On September 9, Harvey said it had raised $550 million. The company’s own post put the mark at $15.5 billion. The Next Web, matching Bloomberg’s tape, printed $15.6 billion and said the number sat a touch above the terms Harvey had been shopping in August. Lightspeed Venture Partners and Diffusion co-led. Diffusion is a new firm from Kris Fredrickson, who had already backed Harvey out of Coatue. Sapphire Ventures and Whale Rock came in as new money. The existing holders who re-upped include Sequoia, Kleiner Perkins, a16z, Coatue, Conviction, Elad Gil, Evantic, GIC, Goldman Sachs Alternatives, Verified Capital, and WNDR.

Harvey has now taken more than $1.5 billion since 2022. Annual recurring revenue has passed $400 million. The customer count is above 3,000 organisations, up from about 1,300 in March, when a $200 million round marked the company at $11 billion. That is a 42 percent step-up in six months while revenue roughly doubled. At $400 million of ARR, $15.6 billion is about 39 times revenue. Ordinary software does not clear that multiple. Vertical AI in 2026 does, if the buyer believes legal work can sit next to coding as an addressable market. Lightspeed’s Sebastian Duesterhoeft, who will sit on Harvey’s board as an observer, said legal is the second-largest AI market after coding, and that in-house teams are the path to the whole legal-services spend.

I read this as a stack-control round first. Harvey grew up on OpenAI. The new money is earmarked for models Harvey trains itself. Tenet, the first post-trained legal model, sits on Kimi K3, an open-weight checkpoint from Moonshot AI. The Next Web says Harvey already routes a task to the cheapest model that is good enough, and fine-tunes on historical documents the firms provide. Winston Weinberg told reporters that every software company needs to become an AI company, and that post-training is part of that job. A legal vendor that can park a Chinese open-weight base under a named product is telling you the frontier API is a cost line.

The Guardrails AI purchase is the other half of that job. Harvey called it the fourth acquisition of 2026. Terms were not disclosed. Co-founders Shreya Rajpal and Zayd Simjee and their team join product and engineering. Guardrails built an open-source validator layer that Harvey says is downloaded more than 250,000 times a month, plus Snowglobe, a simulator that generates synthetic users and stress-tests agent behaviour before a live client ever sees it. Weinberg said more buys are likely and that Harvey is shopping teams more than products. The label on the box is reliability. The cash is paying for people who can keep an agent inside a policy while it works a matter for hours.

The steelman is sitting in the customer list. Harvey says 80 percent of the Am Law 100 now uses the product, along with five of the Fortune 10, Latham & Watkins, and Microsoft’s in-house legal team. Law firms are slow software buyers. They also bill by the hour. A tool that compresses a junior’s night on a diligence memo can shrink the hours the firm has to invoice. Duesterhoeft’s answer is to go after corporate legal departments, where cutting time is the point of the budget. I accept that split as a sales story. I do not have a published mix of firm seats versus in-house seats inside the $400 million. Without that mix, 39 times revenue is a claim about where the next dollar of ARR comes from.

Competition is arriving from the model layer Harvey used to rent. Anthropic has legal plug-ins for Claude. OpenAI is customising ChatGPT with law firms. On the application side, Sweden’s Legora hit $100 million of revenue in 18 months. The Next Web puts Harvey at roughly four times Legora’s revenue and about three times its valuation. That is a lead. It will not stay one if the labs sit inside the same workflow and if Legora keeps compounding in Europe.

Harvey LAB, the legal-agent benchmark the company launched with Tenet, is how Harvey wants that fight scored. A vendor that owns the eval and the watchdog can tell a general counsel that unsupervised hours on a matter were simulated first. I want the split of the $400 million between firms still selling hours and in-house teams buying those hours back. The $550 million pays for weights Harvey can move. Guardrails is the kit for letting those weights run longer than a prompt.