The Wired finding is the one nobody is processing: if the Pentagon was running OpenAI models through Microsoft Azure while OpenAI had an explicit military ban in effect, then every AI lab's "we won't sell to X" policy is unenforceable — the actual access gatekeeper is the cloud layer (Azure, AWS, GCP), not the model developer, which means future AI safety commitments from labs about prohibited use cases are structurally toothless.
This cloud-layer-controls-access dynamic is directly analogous to the RPC layer problem in onchain AI agents — whoever owns the infrastructure layer (Azure/AWS for AI, RPC nodes for Solana agents) controls real access regardless of what the model/protocol developer says their policy is; the implication for Vincent's onchain AI work is that "trustless" AI agents need compute infrastructure that's genuinely non-custodial, not just policy-level claims.