A typical hyperscale data center draws around 100 MW of continuous power — enough to serve a city of 300,000 people. The newest AI campuses are being designed for 500 MW or more: a city of 1.5 million. One building. Hundreds in the pipeline.

This is not a technology problem. It is a physics problem. And physics has a preferred solution — one that computing has always found when centralization hit its ceiling.

Every time a critical resource in computing became scarce, the industry distributed it. Compute went from mainframes to minicomputers to PCs to client-server architectures. Virtualization abstracted workloads from physical hardware running them across pools of resources spanning racks and regions. Cloud extended that logic at massive scale. Networking followed the same arc: backbone fiber gave way to metro densification, CDN edge nodes and 5G small cells. Each wave was triggered by the same dynamic — centralized resources hit a wall in cost, capacity or latency — and distribution unlocked the next order of magnitude. The pattern is consistent enough to be predictive.

AI inference is the wall we are hitting now. A single large language model query consumes roughly ten times the energy of a web search. At hundreds of millions of daily users, the power math becomes structurally difficult. But the deeper problem is not volume — it is geography.

The clusters powering AI today are concentrated in a handful of corridors: Ashburn in Virginia, the Alphaville corridor outside São Paulo, the Amsterdam-Frankfurt axis in Europe, Johannesburg’s Midrand strip, Singapore and Tokyo in Asia and Sydney in Oceania. These locations were chosen historically for tax incentives, land and fiber access — not for proximity to electrical grid capacity or to end users. Grid operators across all of these markets have already issued capacity warnings. Building new transmission infrastructure takes a decade or more in most jurisdictions. And the latency that physics imposes — light traveling roughly 200 kilometers per millisecond in fiber — is becoming a hard technical limit as AI moves into real-time applications: autonomous systems, industrial control, connected healthcare and agentic workflows that cannot absorb the round-trip to a distant campus.

You cannot engineer around geography. You can only reorganize around it.

The market is already moving. NVIDIA launched its EGX platform specifically to run AI inference closer to data sources rather than centralizing it. Jensen Huang has been direct: “Inference has become the most compute-intensive phase of AI — demanding real-time reasoning at planetary scale.” Akamai built its Inference Cloud on NVIDIA’s Blackwell infrastructure to place AI decision-making across thousands of distributed locations worldwide. The direction of the industry’s largest infrastructure bets is not toward bigger campuses — it is toward smarter distribution of where compute physically lives.

Three vectors of value are converging for operators who move early. Latency-sensitive AI workloads will pay a meaningful premium for sub-10-millisecond response times that only local infrastructure can guarantee — a pricing tier that did not exist in the cloud economics of the previous decade. Grid access at underserved nodes, where capacity exists but large-scale compute has not yet been built, will become a structural moat that centralized campuses cannot replicate. And data sovereignty legislation across Brazil, Europe, the Middle East and Southeast Asia is making distributed in-country infrastructure not a compliance cost but a product feature.

The scarcest asset in digital infrastructure over the next decade is not capital or land. It is reliable power interconnection close to where demand is growing.

The grid cannot scale to meet centralized demand. The infrastructure will distribute to meet the grid. Every previous wave of computing ended the same way — and the operators who saw it early captured disproportionate value. The ones who waited optimized for a model that was already changing.