Korva is a scroll of the internet’s best ideas.
Cost that changes by person
In the old arrangement, a few people knew what everyone made, and the figure was readjusted on a quarterly basis, maybe after performance reviews. Now every employee runs up a different amount of inference, and the amount depends on how effectively they pair expensive and cheap models with their tasks. Many companies handle the routing themselves, and it is likely to get pushed down to the individual employee.
The grid of cost and output
Managers can keep assessing how productive people are and line that up with what each person costs. That produces a grid with a quadrant of celebration, where people do a good job at a cost-effective price, and a quadrant of concern, where people do a so-so job at a cost that is off the charts.
Every employee controls their own number and can influence it. Each person now gets to ask how good they are at the job and how efficient they are being while doing it.
Why does one employee's AI bill differ so much from another's?
AI usage is billed per token, and the price per token varies widely between small and frontier models, often by an order of magnitude or more. Long documents and agents that re-read their context at every step multiply the number of tokens a single task consumes. Two people finishing the same task can therefore generate very different bills.
How do systems route tasks to cheaper models?
A router is a small program or model that looks at each request and decides which model should answer it. The RouteLLM project from researchers at LMSYS and Berkeley, published in 2024, trained routers to send easy questions to a cheaper model and hard ones to a stronger model. Their paper reported large cost reductions on benchmarks while keeping most of the stronger model's quality.
How did cloud computing change budgeting before this?
Cloud services turned infrastructure from a fixed purchase into a bill that moves with usage. That shift gave rise to FinOps, a practice and professional community built around tracking and managing variable cloud spending. The FinOps Foundation, part of the Linux Foundation, now publishes frameworks for it.
More from 20VC with Harry Stebbings
