Build-vs-Rent AI: Harvey's $550M Round as a Test

If your product, pricing, or customer trust rides on a rented model API, Harvey's new round is a stress test. The legal AI firm closed a $550 million round at a $15.5 billion valuation, with Diffusion and Lightspeed Venture Partners co-leading. Its annual recurring revenue has passed $400 million, and it now serves more than 3,000 customers. That works out to roughly 39 times revenue. The multiple matters because it prices in a future where the model layer stops being a commodity input.

Harvey's initial product relied on OpenAI technology. That was the right move when the model was a component, not the product. The customer was buying the workflow, the interface, the trust, and the speed. Renting the model is cheaper than training it, and it lets the product ship before the model choice is settled.

The five questions that force a decision

• Workflow centrality: Is the model the reason the customer stays, or a feature behind the reason?
• Switching cost: How much retraining, revalidation, and customer reassurance would a model change require?
• Margin exposure: Does token cost move with usage, and can you pass that through without losing the deal?
• Data advantage: Does your workflow produce feedback that improves a model you control?
• Model ownership risk: Can the supplier raise prices, change terms, or sell the same capability to your buyer?

Most early AI products should answer no to the first two and yes to the last one in a mild way. Rent. The moment the model becomes the workflow, the question changes from cost to control.

Use the answers to set the posture. If workflow centrality and switching cost are low, keep renting and put the savings into the customer workflow. If margin exposure and model ownership risk are high, but data advantage is still thin, add a post-trained hedge. If the model is the product, switching cost is high, and your workflow produces data that only you can use, commit to in-house training.

The hedge is optionality, not supremacy

Harvey introduced Tenet, an in-house post-trained open-weight model, and Harvey LAB, a benchmark for legal AI agents. Tenet is a fine-tuned product based on Moonshot AI's open-weight Kimi K3 foundation, an open-source LLM with 2.8 trillion parameters. Harvey built one of its products on a Chinese open-weight model because it was sufficiently capable and cheaper than an American frontier-model API. The in-house model does not need to beat every frontier model. It needs to keep the workflow running if the rented layer gets expensive, slow, or politically inconvenient.

Harvey assigns each task to the lowest-cost model that can perform it adequately, instead of relying on a single model for all requests. That routing is the real product. It turns model choice into an operational decision, not a brand loyalty test. A company that can swap models by task can keep the expensive frontier model for the few steps where quality matters, while using cheaper models for the routine steps.

The routing table changes the build-versus-rent calculation. When a frontier API raises prices, the founder can move a task, absorb the change, or substitute a cheaper model. The customer sees the same workflow. The supplier sees a smaller share of the revenue. The rent becomes a variable cost, not a fixed dependency. The founder still pays for the frontier, but the product does not break when the price changes.

When the supplier becomes the competitor

The strongest objection to building is simple: frontier APIs improve fast, and a small team cannot match them. Fair enough. The mistake is treating the API as a permanent utility. Harvey's model suppliers are turning into competitors, as Anthropic shipped legal plug-ins for Claude and OpenAI worked with law firms to customise ChatGPT. If your buyer can get a decent version of your workflow from the model vendor, your advantage has to be the workflow, the data, and the trust.

The same vendor that sells tokens can sell the workflow to your buyer. Speed matters when the API is a component. It matters less when the workflow, data, and trust are the product.

Winston Weinberg, Harvey's chief executive, said post-training models are part of software companies becoming AI companies and that model training is a competitive capability. Harvey is expanding its team and will spend heavily on its harness and its own model training. That is the expensive part of the hedge. It is also the part that makes the hedge real. If the five questions point to control, the hedge is worth buying even when the rented model still wins on raw quality.

This article is general information, not investment or legal advice.