Will Increasing Efficiency of AI Computing Render Huge Data Centers Superfluous?

Comparing the DoD's first major digital computer to the iPhone and wondering if similar size and resource efficiency is in the cards for AI computing

The first major digital computer used for a large-scale U.S. defense project, the ENIAC (funded by the U.S. Army Ordnance Department), occupied 1,800 square feet of floor space and weighed 30 tons. A modern iPhone is hundreds of thousands to millions of times more powerful than this room-sized machine.

As AI computing becomes more efficient, requiring smaller machines, energy, and cooling capacity, will the massive data centers being built today be rendered superfluous?

The largest AI data center complexes under development require 11 gigawatts (11,000 megawatts) of total power capacity.

To understand the scale of 11 gigawatts of power being delivered to a gigantic machine — as distinct from other forms of human activity and labor — consider that a single gigawatt of continuous power running 24/7 is equivalent to the entire electricity consumption of a major metropolitan area like Denver, Colorado (roughly 850,000 to 1 million homes).

What exactly is going to be done with all of this machine intelligence, apart from rendering most human intellectual labor obsolete?

While AI is becoming marvelously capable at synthesizing enormous amounts of data, analysis, and concepts with lightening speed, it strikes me as odd that it consumes so much more power than the human brain in a state of intense concentration, which uses about 20 watts of power, roughly 25% of the body's total energy, or about 400 to 500 calories per day.

I used to date a girl who could simultaneously speed through London traffic, smoke a cigarette, drink a Corona, answer her cell phone for a quick chat, put on lipstick, and have an intelligent conversation with me without having to think about any of these tasks and without causing an accident.

How much processing power is required to do all of these actions at once, and how much energy would it require if a robot were trained to do all of them at once?

I wonder if the AI nerds are studying the human brain in the way that aviation guys (finally) got around to paying close attention to birds’ wings (such as that of the mallard duck) which feature the sort of aerodynamic twist that was ultimately built into the Boeing 787 wing, thereby greatly increasing its efficiency—that is, reduction of induced drag.