A user argues that the opportunity around AI accelerators extends beyond running models, much as cloud providers built products around data, networking, security and storage. They highlight training, routing requests to the right model, security monitoring and tools for managing AI agents as areas to build. The post rejects a winner-take-all outlook, arguing that customers want choice, flexibility and building blocks for their own systems. It also predicts that constrained GPU supply will prolong competition: customers need both compute capacity and software, and will go elsewhere when capacity is unavailable.