Tech • AI • Robotics • Game
Andreessen Horowitz Launches Horowitz Andreessen, New Models, Human Powered Agents, McLaren Rebrand
7/10TL;DR
Harvey briefly fell to negative 50% gross margin as AI usage surged, then returned to positive margins within a quarter, highlighting the growing pains facing enterprise AI companies built on expensive frontier models.
KEY POINTS
Harvey’s margin swing
Harvey, an AI legal software company, saw gross margin drop from about 50% to negative 50% in June as customer use of agentic and reasoning-heavy features sharply increased. The shift meant the company was effectively selling AI output below cost for a period, a rare but increasingly visible problem for software firms relying on rented models from providers such as OpenAI and Anthropic.
Why costs spiked
The core issue was not simply more customers, but much heavier token consumption per customer. Legal users may have issued similar prompts as before, but newer reasoning models and multi-step agents generated far more compute-intensive workflows, with token usage reported up 20-fold this year. That created a mismatch between traditional seat-based pricing and rapidly rising inference costs.
Back to positive margins
The company said it restored gross margins to positive territory within a single quarter even as usage kept rising. The recovery reportedly came through model routing, infrastructure changes, spend controls, and post-training an open-weight model branded Harvey Tenant. The episode suggests that AI application companies can improve economics quickly, but only if they actively optimize model mix and customer pricing.
Scale softens the blow
At an annualized revenue run rate of about $400 million, Harvey’s monthly revenue is roughly $33 million. A negative 50% gross margin for one month would imply an operating hit of roughly $16 million before other expenses, significant but not existential for a company that has raised around $500 million. That context has tempered concerns that the June result signaled a deeper crisis.
Investor confidence appears intact
Harvey was recently valued at $15.5 billion in a deal associated with Sequoia, indicating investors were likely aware of at least some near-term margin pressure. The willingness to fund the company at that level suggests backers are treating the episode as a transitional problem tied to product adoption rather than a structural failure of demand.
A broader AI pricing problem
The incident underscores a wider challenge across the AI application layer: companies sold subscriptions before the true cost profile of reasoning agents was clear. As model capabilities improve, customers use products more deeply, but that can erase software-style margins unless providers reprice, fine-tune cheaper models, or sharply improve routing. Many AI startups may face similar moments as usage shifts from chat to autonomous workflows.
Benchmarks for legal software remain high
More mature legal technology companies offer a stark contrast. DISCO, an e-discovery software provider, has posted gross margins around 75%, while Thomson Reuters has reported legal software segment profitability near 50% adjusted EBITDA. That gap shows how far AI-native legal tools still need to go if they want to match the economics long associated with enterprise software and the highly profitable legal industry itself.
Cheaper models may help
New model releases are steadily lowering inference costs, giving companies like Harvey more room to defend margins without degrading product quality. One benchmark that drew attention recently was a legal task benchmark emphasizing cost efficiency, where newer models appeared especially inexpensive for law-focused workloads. If that trend continues, legal AI vendors may be able to keep premium features while reducing dependence on the most expensive frontier systems.
CONCLUSION
Harvey’s June margin collapse exposed how quickly AI product success can become a cost problem when usage outpaces pricing. Its rebound shows that demand remains strong, but sustainable profits in legal AI will depend on cheaper models, better routing, and business models built for the economics of agents rather than seats.
Ask a question