You’re mixing up things here…sub-5MW on existing interconnects is real and it's useful. But it's scavenging spare headroom, which is finite and thinning fast (+ will attract NIMBY^10). That fills the marginal demand. It isn't where the bulk of capacity comes from. This isn't a training vs inference split. The inference-first buyers in front of us right now are asking for hundreds of MW, not 5. Serving a frontier model isn't one chip answering one request - weights are sharded across 100s of accelerators that have to behave as a single tightly-coupled system, with all-to-all traffic between them on every token. If you break that across 40 sites you don't get 40x the throughput. You get worse utilization, worse batching, no prefix cache reuse and a materially worse cost per token. The companies chasing the lowest latency are the ones needing the densest, most tightly-coupled deployments. Not the most distributed ones. And 5ms isn't the constraint anyway. Time to first token is hundreds of ms, generation runs for seconds. Nobody's bottleneck is 5ms of transit. The demand we see is 300MW+, and almost nobody can deliver it on a timeline that matters. Also - as a power guy, you’re ignoring the opportunity of co-locating renewables power generation + BESS to be financed by fully-priced PPAs…which is really why I dont’t get why you’re on this “small metro area DC” crusade in the first place.