When Agents Got Chatty, This Startup's Margins Went to Minus 50 Percent
Bloomberg reports legal AI firm Harvey watched gross margins fall from about 50 percent to minus 50 by June as token use rose twentyfold. It recovered by building an in-house model on Moonshot's open Kimi K3.
Harvey sells AI tools to law firms and is valued at around 15.6 billion dollars. According to Bloomberg, its gross margin went from roughly 50 percent at the start of the year to minus 50 percent by June, meaning every dollar of revenue cost about a dollar fifty to serve. The trigger was a March update that turned its product into agents, software that works through a task over many steps rather than answering one question. Token usage jumped twentyfold. Because OpenAI and Anthropic charge per token, and Harvey charges its customers a flat subscription, all of that increase landed on Harvey.
The fix was to stop renting the engine. In August, Harvey launched its own in-house model, post-trained on Kimi K3, the open-weight model from Chinese lab Moonshot AI. Post-training means taking somebody else’s finished model and continuing to train it on your own data and tasks, which is possible only when the weights are downloadable. Margins went positive again. Bloomberg names Abridge in healthcare, Decagon in customer support and Ramp in finance as making similar moves, with backing from Sequoia Capital and General Catalyst.
What is behind this
This is the collision of two business models that were never designed to meet. Software companies have always sold seats: a fixed price per user per month, with costs that barely move when that user works harder. AI agents break that, because a diligent agent that checks its work is an agent that spends more of your money, and the better your product gets, the worse your unit economics look. That is an unusual and quite brutal shape for a business.
There are three ways out, and everyone is trying some mix of them. Charge customers by usage, which they hate and which makes budgets unpredictable. Wait for prices to fall, which they do, though demand has so far risen faster. Or move the workload to a model you control. The third option only became realistic because open-weight models from Chinese labs got good enough this year, which is why a New York legal startup ends up building on a model from Beijing. Worth being precise about the caveat: running your own model is not free, it trades a per-token bill for engineering staff, hardware and the risk of falling behind the frontier.
What this means for you: If you buy AI software, ask how your vendor pays for inference, because a supplier losing money on every heavy user will eventually raise prices, throttle usage, or quietly route you to a cheaper model. Watch for usage caps appearing in plans you already have. And if you are building something on top of a paid model, measure token use per completed task early, not after launch. The interesting signal for everyone else is that “open weights are for hobbyists” is no longer true. It is now a margin decision made by companies worth billions.
Sources
Alibaba's New Model Watches and Listens, and Charges Almost Nothing for It
Qwen3.8-Omni-Flash handles text, images, audio and video in one model with a million-token context. Alibaba reports hourly audio ingestion costs falling by more than 98 percent. No open weights at launch.
