Computing did not begin with datacenters. Ada Lovelace wrote the first algorithm in 1843 for Charles Babbage’s Analytical Engine, extending ideas first explored decades before digital logic existed. From those humble origins, computing advanced through mechanical machines, electromechanical and hybrid digital systems, to commercial computers like UNIVAC I and IBM 701. And in that era, Howard Aiken famously argued that the world would only ever need “six… or even two or three” computers.
The same pattern is repeating with AI.
From symbolic reasoning and expert systems, through statistical learning and neural networks, to deep learning and generative models… The milestones might be different. The journey is the same. Each implementation started centralized. Each implementation ended distributed. Intelligence always moves outward — from the center to the edge, from the lab to the device, from the few to the many.
The entropy flattens the distribution curve.
Frontier models have their place. I am not arguing against that. They have real value for specific problem domains. They are also expensive, power‑hungry, and structurally dependent on hyper‑scale infrastructure. The latest industry change is per-token billing, ending the days when a flat monthly subscription covered customers’ needs. Today, users are billed on a per-token rate. Metered inference makes frontier models predictable for vendors but volatile for users. There are no profitable AI companies. They must raise rates. Per-token billing is the only way they can do it.
This change creates a forcing function.
Local agents, by contrast, align with the constraints of real systems. They operate where computation actually happens — on the computers individuals and businesses already own, inside the workflows people already use, and within the budgets real people can afford. These are not implementations that require billions in capital expenditures. They do not require water, power, and enough hardware to create global shortages of critical components. New tools, techniques, and optimizations emerge every day. More will appear by the time you finish this article. These innovations point us in a new direction.
We have reached an inflection point.
The future of AI is distributed. Frontier models will continue to push the boundaries of science, reasoning, and creative problem solving, but the systems people actually use will be local. Per-token pricing makes this shift inevitable. Latency, privacy, and context reinforce it. Intelligence is leaving the datacenter and moving into every device, every workflow, and every system.
The Datacenter won’t win.
Originally published on LinkedIn on August 1, 2026.