The AI sandbox play also confuses me because (aside from the fractional pay-as-you-use billing) it's trivially implemented on top of any VM/container as a service provider. Given the reduced importance of data retention (the containers are disposable) and ability to redo work trivially you can even go a step further and have your agent spin it up on bare metal. Turns out agents spend most of their time waiting on GPUs so with memory dedup on your own similar workloads you can run "hundreds" of them on pretty minimal hardware.

I also don't see how the cloud-for-agents will break the AWS monopoly: AWS already is the cloud for agents! As the article notes all the pain of using AWS is trivialized by agents (and IaC).

To me the value now is in building the right tools for agents to make better decisions, not fighting for basis points on hardware margin in the orchestration or model serving layer.