Korva is a scroll of the internet’s best ideas.
Fixed-fee deals fall behind AI growth
The old deal
Before this approach, a publisher with high-quality content had one option: be big enough to sign a fixed-fee contract with an AI lab covering some training, some liability protection, and inference-time access. That option is open only to the head of the web, and even there it is a broken deal.
Say inference grows 7x this year and another 7x the next, about 50x in all. A publisher's deal size does not grow 50x, and none of them signs a two-year contract believing its share won't decline materially at renewal. Fixed prices don't suit a market where usage multiplies.
What labs pay for
Labs pay for some mix of three things: fresh data at inference time to show in products, training data, and liability protection for training already done. A workable replacement has to pay a lot for extraordinarily valuable content, pay a lot for valuable work done with the same content, and scale without bespoke deals.
The macro math
The macro math points the same way. If 2 to 10 percent of what is spent on LLM inference for knowledge work went to web data, that would exceed every web data business model today outside walled gardens like Facebook and LinkedIn. With agents on the web growing an order of magnitude a year, the calculation puts meaningful payments to a very wide range of content owners 12 to 24 months out.
Why do flat licenses lose value as usage grows?
A flat fee stays the same size while the value of the thing licensed keeps changing. Music licensing handles this differently: organizations such as ASCAP, founded in 1914, collect fees tied to how music is used and pay royalties based on tracked performances. Usage-based pricing lets the payment move with the activity it covers. A fixed contract has no such link, so growth in use benefits only the buyer.
What publisher deals have AI labs actually signed?
Starting in 2024, several large publishers and platforms announced licensing agreements with AI developers. Reuters reported in February 2024 that Reddit had signed a deal with Google worth about 60 million dollars a year. News Corp and OpenAI announced a multiyear agreement in May 2024. Most of these deals involve a single large rights holder and a fixed multiyear payment.
Why is training access priced differently from inference access?
Training happens once: a model reads the text and absorbs patterns into its weights. Inference-time access happens every time a user asks a question that needs current information. The first is a one-time ingestion, while the second repeats with every query. That makes inference-time access the part of the market that grows when usage grows.
More from Training Data
