In 1965, Gordon Moore predicted that computer chips would keep doubling in power every two years while getting cheaper. He was right. That’s how we went from room‑sized machines to smartphones in our pockets.
Logic says infrastructure should have shrunk too. Instead, we now see mega‑data centers sprawling across the globe: Microsoft’s 600‑acre campus in Arizona, Google’s 23 giant facilities, Meta’s $800 million site in Illinois, and Amazon’s 125+ centers worldwide. These aren’t just buildings — they’re small cities, consuming as much electricity as entire countries.
AI broke the equation.
Moore’s Law vs. AI’s Appetite
Modern chips are incredibly efficient. But artificial intelligence demands far more than efficiency — it demands scale.
Training today’s frontier models costs tens or even hundreds of millions of dollars and uses enough electricity to power thousands of homes. Running them daily consumes energy on the scale of entire towns.
Moore’s Law promised we’d need fewer machines over time. AI flipped the script: bigger models demand exponentially more machines, housed in ever‑larger facilities.
The Scale of the Build‑Out
The AI market has exploded from $25 billion in 2013 to over $200 billion today, with projections of $400 billion by 2030.
Data centers already consume more electricity than Argentina, and by 2030 could use nearly 1 in 10 watts of global power.
Some facilities use billions of gallons of water a year for cooling, often in drought‑prone regions.
This isn’t just growth. It’s a reshaping of global infrastructure.
Four Walls Closing In
Physics: Chips are nearing atomic limits. Shrinking them further may take decades.
Monopoly: Only a handful of tech giants can afford the billions needed to train frontier AI. Startups are locked out.
Environment: Carbon emissions, water use, and energy strain are mounting. “Carbon neutral” claims often mask the reality.
Pushback: Communities from Ireland to Singapore are blocking new data centers over grid strain, water use, and minimal local benefits.
We’ve Seen This Movie Before
In the 1990s, telecom companies spent over $100 billion laying fiber‑optic cables, betting on endless internet growth. When the dot‑com bubble burst, much of that fiber sat unused, and companies went bankrupt.
Today’s AI boom shows similar signs: sky‑high valuations, massive infrastructure spending, and every company rushing to add “AI” to its products. If the hype slows, data centers could sit half‑empty, GPUs sold for pennies, and billions written off.
Three Possible Futures
AI Delivers (30%)
Real productivity gains, new breakthroughs, and energy solutions.Bubble Pops (40%)
Growth slows, facilities underused, valuations collapse.Hard Stop (30%)
Energy caps, water limits, or public resistance force a halt.
Bottom Line
Moore’s Law promised efficiency. AI demands scale at any cost.
But energy is strained, water is scarce, carbon targets are breaking, and five companies dominate the field. We’re building as if exponential growth will last forever. History says it won’t.
The question isn’t whether we can build larger data centers.
It’s what happens when we realize we shouldn’t have.
Are we building the future — or repeating history?
Top comments (1)
I liked the article and thinking behind it, including possible AI futures. With my added mainframe perspective, I think data centers could end up mostly empty. Here’s why.
I reflect back to the mid-70s when IBM moved me to Texas as a Systems Engineer. At that time, the System/370 Model 158 was often used as a benchmark, because it could execute 1 million instructions per second (MIPS). But it was expensive (~$3.5 million), large, and not only required AC but liquid cooling, because the processor chip ran as hot as an iron. It was approaching limits that sound familiar.
For general purpose applications, The iPhone 17's A19 CPU executes roughly 25 billion instructions per second (25 GIPS) using its standard processing cores. However, for specialized machine learning or AI tasks, its dedicated Neural Engine can process over 35 trillion operations per second.
By that measure, the iPhone is 25,000 to 35 million times faster, but there’s more. It’s smaller, battery operated, affordable, and personal, no longer shared like the expensive mainframes were. Could a similar paradigm shift occur in AI?
While Elon Musk talks about putting data centers in space to address the cooling problem, that introduces latency problems. With LEO satellites millions of miles away, the speed of light causes transmission delays of half a second each way, and that delay won’t work for applications like self-driving cars. So, at least some of the AI function will need to be closer, on the network edge, in the device it self, of on the sensors.
Looking back on the 60+ years of tech innovation and trends I experienced, I’m not convinced investing in big data centers is wise longterm. I remember one of my banking customers asking me to help design their new data centers, and their surprise when I recommended making it smaller, but with a much larger room for the check reader/sorter.