Korva is a scroll of the internet’s best ideas.
Only frontier models earn a premium
Two tiers
Lagging models still have a large market, but the weights can't be sold for anything. Money comes from the compute, the inference, and the consulting work to put it all together. Only a lab at the frontier can charge for the model layer itself.
A year ago, five major companies were in the hunt to lead the frontier. Today it comes down to Anthropic and OpenAI, with xAI and Google still in the race but able to do well on compute alone.
Growth shows the premium
Growth shows the premium. Anthropic's annual recurring revenue is over $80 billion, up from $10 billion at the start of the year. It had forecast $100 billion by year end, which most people called impossible, and estimates for year end now run $110 to $120 billion or higher. OpenAI is also accelerating.
The Apple parallel
Apple and Android show the same split: Android is open source and has more users worldwide, yet nearly all the monetization goes to Apple because people pay for the premium experience.
Why can't lagging models charge for their weights?
Weights are files that anyone can copy once they are released, so the model alone has no scarcity to price. Meta released its Llama 2 weights in 2023 under a license that allowed broad commercial use. When many companies offer the same open model, competition among hosting providers pushes prices toward the cost of electricity and hardware.
How does Apple earn most smartphone profits?
Android is open source and runs on far more phones worldwide, yet Apple sells fewer, higher-priced phones along with services. Counterpoint Research has estimated that Apple takes the large majority of global smartphone profits. Android was first released in 2008 on the HTC Dream, and it is built on the Android Open Source Project, which any manufacturer can use freely.
Why do customers pay more for the best model?
For hard tasks like writing code or analyzing contracts, a more accurate model saves human review time. A small gain in accuracy can decide whether a task can be handed to a machine at all. Developers measure this with benchmarks such as SWE-bench, introduced in 2023 by researchers at Princeton University and the University of Chicago, which tests whether models can resolve real GitHub issues.
More from All-In with Chamath, Jason, Sacks & Friedberg
