The Under-Tracked Constraint: Why Networks Bound Compute
Everyone models power and land. Almost nobody models the fiber. That asymmetry is the opportunity. A gigawatt campus with no diverse long-haul path is a stranded asset the moment a backhoe cuts its single trunk. My read: over the next cycle, the binding constraint on where AI compute actually lands shifts from megawatts to routed, diverse, low-latency photons. Network architecture is the under-tracked constraint on AI compute delivery. Treat fiber as a first-class site variable, not a closing checklist item, and you see stranded-asset risk before the market prices it. The building runs on electrons; it matters because of photons.
Knowledge check
Why can a fully powered data center campus still fail to deliver usable AI compute?
Show the answer
Because compute must be reachable and interconnectable. Without diverse, high-capacity long-haul fiber and low-latency routes to users and to partner campuses, the site cannot move training and inference traffic. Power and land are necessary but not sufficient. A campus dependent on a single fiber trunk is one backhoe strike away from being severed from its customers, so its installed capacity is not the same as deliverable compute.