because claude wasn't using enough data center resources we need to give him additional servers in data centers on demand?

Fair, if you're running one-off prompts, you don't need this. The real use case is longer agent tasks where you don't want rm -rf, a runaway loop, or pnpm install on some random repo touching your laptop. VM spins up, agent works, VM gets destroyed. Same reason people already sandbox in Docker/Codespaces, this just collapses create → run → destroy into one command.