Korva is a scroll of the internet’s best ideas.
Free models sell the consulting
Two business models
In the revenue-share version, a lab releases an open model and lets inference providers such as Fireworks or Together deploy it. Once a provider passes a set amount of revenue, the lab asks for a share, which makes the lab a partner in the compute revenue.
The second version, closer to what Mistral and Thinking Machines do, treats the model as a lead generation tool. Companies build on it, then come back for help fine-tuning it or shaping their AI strategy, and the lab sends in forward-deployed engineers, who work directly inside the customer's business.
The consulting market
That market is enormous. One of the biggest markets of the next 10 years is AI modernization: going into every business in the world, restructuring its data, building models into workflows, and teaching employees to use them. A lab that captures it could grow into a company worth hundreds of billions of dollars or a trillion.
Frontier labs and Microsoft also run forward-deployed engineering. An open American model adds ownership to it. A business can keep the model completely, keep fine-tuning it in-house, and avoid both third-party costs and supply chain risk.
How do open licenses carry revenue terms?
Some open-weight licenses attach conditions to very large users. Meta's Llama 2 license, released in 2023, required companies with more than 700 million monthly active users to request a separate license from Meta. Free use below the threshold and negotiated terms above it let a lab keep the model open while still charging the biggest players. Each license sets its own threshold and terms in its text.
Why do software firms give products away?
Red Hat built a large business selling support, certification, and services around Linux, which anyone can download for free. The same pattern runs through MongoDB and Elastic, whose free versions pulled customers toward paid features and hosting. IBM agreed to acquire Red Hat in 2018 for about 34 billion dollars, and the deal closed in 2019.
Why do forward-deployed engineers exist?
Palantir popularized the role: engineers who work on site inside a customer's organization instead of from the vendor's office. Complex software often fails to take hold because it has to be wired into messy data and workflows, and someone has to do that wiring. An engineer on site sees the problem directly and can adapt the product on the spot. Other enterprise software and AI firms have since adopted the title.
More from 20VC with Harry Stebbings
