Korva is a scroll of the internet’s best ideas.
Token spend depends on the use case
Growth-stage logic
For growth-stage companies, spending is worth it as long as the unit economics are fine. Coding agent spend for software engineers isn't cost of goods, and even when it's very high it can still give compounding gains.
The bad case
Picture a customer service agent with massive token spend, serving customers whose revenue is way lower than that spend: that company is definitely in a bad position.
Spend by team
Large companies need different spend profiles for different teams. Solutions engineering and forward-deployed teams have to think in unit economics, while R&D teams can tolerate more spend. Salesforce's reported 300 million dollars a year with Anthropic works out to about 3.8 percent of developer salaries, and that share is expected to rise past 3 percent over time.
Why do R&D costs differ from cost of goods?
Cost of goods sold covers the direct costs of delivering what a company sells, so it comes straight out of gross margin. Research and development is recorded as an operating expense, and under US accounting rules it is generally expensed as it is incurred. Investors watch gross margin to judge whether each sale is profitable. A coding tool used by engineers falls in the second category, while an AI agent serving customers sits in the first.
Why do AI products squeeze margins?
Subscription software traditionally carries high gross margins because serving one more customer costs little. AI features change that, since each request runs a model on computing hardware and costs money every time. Inference, the term for running a trained model to produce an output, is often billed per token.
Why can tool spending compound into gains?
Douglas Engelbart, who also developed the computer mouse, described the idea in the 1960s as bootstrapping: using tools to improve the process of building tools. A coding agent that speeds up the construction of the next tool feeds that loop. Engelbart's 1962 report Augmenting Human Intellect laid out the framework.
More from 20VC with Harry Stebbings
