That is where the really big systems get developed, where the data gets gathered and cleaned, where new approaches get tried and where the master models get trained. Then smaller versions can be produced for laptops, phones, cars, hospitals and factories.
In other words, the expensive part may increasingly happen in one place while much of the everyday use happens somewhere else.
That doesn’t mean the data-center boom ends. This is where it gets a little more complicated, because whenever computing gets cheaper, we seem to come up with more things to do with it.
Suppose better chips and better models eventually cut the electricity needed for an AI task by 90 percent. You might expect power consumption to collapse. But if we respond by doing one hundred times as many AI tasks, total consumption still rises tenfold.
The data center can lose work and still grow.
Today I might ask an AI to summarize one document. Tomorrow I may have several agents continuously reading messages, monitoring accounts, checking equipment, writing software and talking to other agents. Scientists may throw enormous amounts of computing power at new materials, medicines and energy technologies simply because they finally can afford to.
The International Energy Agency expects global data-center electricity consumption to rise from about 485 terawatt-hours in 2025 to roughly 950 terawatt-hours in 2030. Consumption by AI-focused centers is expected to triple.⁹ So I’m not predicting rows of abandoned data centers. I’m saying their job may change.
The biggest centers may increasingly be where frontier intelligence is created, models are improved, scientific simulations are run, and where genuinely difficult problems are solved.
They may lose the email and gain the laboratory.
Which brings me back to Georgia.
The processors going into today’s data centers are moving on a technological clock measured in years. The buildings, substations and transmission lines around them are expected to last for decades. Microsoft recently said roughly two-thirds of one quarter’s capital spending went into relatively short-lived assets, primarily CPUs and GPUs, while it has also extended the expected useful life of its data-center and office buildings from 15 years to 25 years.¹⁰,¹¹
The chips will be replaced over and over. The substation will still be there.
That is where this stops being just a Silicon Valley investment question. Federal energy regulators are already looking at how utilities should protect ordinary ratepayers when transmission and other infrastructure are built for giant data-center loads that are later delayed, reduced or never materialize.¹² OpenAI says it will pay the full cost of the infrastructure needed for Project Camellia, so I’m not suggesting Georgia customers are going to be left holding that particular bag.¹
But we ought to pay attention to the assumptions being made before all this gets built. We are putting in infrastructure that may last 30, 40 or 50 years to serve a computing architecture that is changing almost annually.
OpenAI may need every one of those 3.2 gigawatts in 2036. The Georgia complex could be running flat out, training better models, designing medicines, running scientific simulations and supplying specialized intelligence to billions of smaller machines. I can easily imagine that.
I just don’t think it will need to read all our email.
My bet is that by the early 2030s, a substantial share of the routine AI used by individuals and businesses will run on personal devices, vehicles and private organizational systems. The largest centers will still be indispensable, but more of their work will be the stuff that genuinely needs enormous concentrations of computing power.