Korva is a scroll of the internet’s best ideas.
The model gap
The US is spending probably seven times as much as China this year on data centers, and Chinese models are only four to nine months behind. American labs have to charge enough to pay their bondholders and stockholders. A company that can get almost as good a search for half the price from DeepSeek, Moonshot, or Alibaba's Qwen has a reason to switch.
The car comparison
Cars show how it goes. China makes a pretty darn good electric vehicle in BYD, and it has taken over most of the world except for the US. Most of an economy never buys the frontier version of anything, whether a desk chair or an automobile. In New York City it can look as if everyone drives a BMW, yet that is not what the Uber drivers are driving.
Mixing models
Hundreds of companies in 2026 are already leaning into AI while managing what it costs. They use what technologists call orchestration, mixing models so no single vendor can lock them in or charge a lot more. Each time American companies raise prices, more US companies turn to the good enough Chinese models. China is also ahead on industrial robotics.
How did BYD grow so quickly?
BYD started in 1995 as a maker of rechargeable batteries in Shenzhen and entered cars in 2003 by buying a small automaker. Because it builds its own batteries, it controls one of the largest costs in an electric vehicle. Chinese national and local incentives for electric vehicles helped create a large home market through the 2010s. In late 2023, BYD sold more battery-electric vehicles in a quarter than Tesla for the first time.
How does model orchestration avoid vendor lock-in?
Orchestration software sits between an application and the AI models it calls, and routes each request to whichever model fits. Easy requests can go to a cheap model, and hard ones to an expensive one. Because many providers accept similar request formats, swapping one model for another often means changing a setting rather than rewriting the application. Open-weight models, whose files can run on a company's own servers, add an option that no outside vendor can reprice.
Why do cheaper models trail by only months?
Research methods in AI spread quickly through published papers and through engineers who move between labs. Smaller models can also be trained on the outputs of larger ones, a technique called distillation. Stanford's AI Index reported in 2025 that the gap between leading US and Chinese models on major benchmarks had narrowed sharply during 2024. Its report named MMLU and HumanEval among the benchmarks where the gap had nearly closed.
More from Freakonomics Radio Network
