Maybe they want to build their own cloud infrastructure, host open source models themselves as an inference service provider?