Korva is a scroll of the internet’s best ideas.
Sixty model builders became three
Three possible outcomes
In April 2023, an investment firm laid out the possibilities for the foundation model layer. It might be winner take all, or a total commodity with open-source players racing the price to zero, or an oligopoly with three or four leading players.
Who entered the race
Some of the largest companies in the world joined the field. Amazon never really showed up. Meta came in strong, its effort faltered, and it had to do a total reboot. Anthropic was the dark horse, a startup that focused purely on the enterprise, while OpenAI had won the consumer and Google could never be counted out.
Like the cloud market
The race came to look like the cloud market, where three companies underpin the entire SaaS world. Open-source models from China came close to the leading edge, but with less compute they couldn't leapfrog it, and then they faltered. When a model is already 80 percent of the way to the top of the benchmarks, going from 80 to 85 is a huge unlock.
Why did most model startups die?
Training a frontier model takes clusters of thousands of accelerators running for months, and the bill runs to hundreds of millions of dollars. Each new generation then needs a bigger run, so a lab must keep finding revenue or backers to fund the next one. Few startups could do that against labs attached to large companies or to major cloud contracts. Sam Altman said in 2023 that training GPT-4 cost more than $100 million.
Why do three companies dominate cloud?
Cloud infrastructure carries enormous fixed costs for data centers, networks, and custom chips, and every added customer lowers the cost per unit. That favors the biggest providers, which can also spend more on research and new services each year. Synergy Research Group's market tallies have long shown Amazon, Microsoft, and Google holding the majority of global cloud infrastructure spending.
How do benchmarks measure model progress?
A benchmark is a fixed set of questions with known answers, and a model's score is the share it gets right. MMLU, released in 2020 by Dan Hendrycks and colleagues, covers 57 subjects from law to physics. As top models approached its ceiling, researchers built harder tests. One of them, Humanity's Last Exam, was released in 2025 by the Center for AI Safety and Scale AI.
More from Invest Like the Best with Patrick O'Shaughnessy
