This week we talk about energy consumption, pollution, and bipartisan issues.

We also discuss local politics, data center costs, and the Magnificent 7 tech companies.

Recommended Book: Against the Machine by Paul Kingsnorth

Transcript

In 2024, the International Energy Agency estimated that data centers consumed about 1.5% of all electricity generated, globally, that year. It went on to project that energy consumption by data centers could double by 2030, though other estimates are higher, due to the ballooning of investment in AI-focused data centers by some of the world’s largest tech companies.

There are all sorts of data centers that serve all kinds of purposes, and they’ve been around since the mid-20th century, since the development of general purposes digital computers, like the 1945 Electronic Numerical Integrator and Computer, or ENIAC, which was programmable and reprogrammable, and used to study, among other things, the feasibility of thermonuclear weapons.

ENIAC was built on the campus of the University of Pennsylvania and cost just shy of $500,000, which in today’s money would be around $7 million. It was able to do calculators about a thousand times faster than other, electro-mechanical calculators that were available at the time, and was thus considered to be a pretty big deal, making some types of calculation that were previously not feasible, not only feasible, but casually accomplishable.

This general model of building big-old computers at a center location was the way of things, on a practical level, until the dawn of personal computers in the 1980s. The mainframe-terminal setup that dominated until then necessitated that the huge, cumbersome computing hardware was all located in a big room somewhere, and then the terminal devices were points of access that allowed people to tap into those centralized resources.

Microcomputers of the sort of a person might have in their home changed that dynamic, but the dawn of the internet reintroduced something similar, allowing folks to have a computer at home or at their desk, which has its own resources, but to then tap into other microcomputers, and to still other larger, more powerful computers across internet connections. Going on the web and visiting a website is basically just that: connecting to another computer somewhere, that distant device storing the website data on its hard drive and sending the results to your probably less-powerful device, at home or work.

In the late-90s and early 2000s, this dynamic evolved still further, those far-off machines doing more and more heavy-lifting to create more and more sophisticated online experiences. This manifested as websites that were malleable and editable by the end-user—part of the so-called Web 2.0 experience, which allowed for comments and chat rooms and the uploading of images to those sites, based at those far off machines—and then as streaming video and music, and proto-versions of social networks became a thing, these channels connecting personal devices to more powerful, far-off devices needed more bandwidth, because more and more work was being done by those powerful, centrally located computers, so that the results could be distributed via the internet to all those personal computers and, increasingly, other devices like phones and tablets.

Modern data centers do a lot of the same work as those earlier iterations, though increasingly they do a whole lot more heavy-lifting labor, as well. They’ve got hardware capable of, for instance, playing the most high-end video games at the highest settings, and then sending, frame by frame, the output of said video games to a weaker device, someone’s phone or comparably low-end computer, at home, allowing the user of those weaker devices to play those games, their keyboard or controller inputs sent to the data center fast enough that they can control what’s happening and see the result on their own screen in less than the blink of an eye.

This is also what allows folks to store backups on cloud servers, big hard drives located in such facilities, and it’s what allows the current AI boom to function—all the expensive computers and their high-end chips located at enormous data centers with sophisticated cooling systems and high-throughput cables that allow folks around the world to tap into their AI models, interact with them, have them do heavy-lifting for them, and then those computers at these data centers send all that information back out into the world, to their devices, even if those devices are underpowered and could never do that same kind of work on their own.

What I’d like to talk about today are data centers, the enormous boom in their construction, and how these things are becoming a surprise hot button political issue pretty much everywhere.

—

As of early 2024, the US was host to nearly 5,400 data centers sprawled across the country. That’s more than any other nation, and that number is growing quickly as those aforementioned enormous tech companies, including the Magnificent 7 tech companies, Nvidia, Apple, Alphabet, Microsoft, Amazon, Meta, and Tesla, which have a combined market cap of about $21.7 trillion as of mid-December 2025, which is about two-thirds of the US’s total GDP for the year, and which is more than the European Union’s total GDP, which weighs in at around $19.4 trillion, as of October 2025—as they splurge on more and more of them.

These aren’t the only companies building data centers at breakneck speed—there are quite a few competitors in China doing the same, for instance—but they’re putting up the lion’s share of resources for this sort of infrastructure right now, in part because they anticipate a whole lot of near-future demand for AI services, and those services require just a silly amount of processing power, which itself requires a silly amount of monetary investment and electricity, but also because, first, there aren’t a lot of moats, meaning protective, defensive assets in this industry, as is evidenced by their continual leapfrogging of each other, and the notion that a lot of what they’re doing, today, will probably become commodity services in not too long, rather than high-end services people and businesses will be inclined to pay big money for, and second, because there’s a suspicion, held by many in this industry, that there’s an AI shake-out coming, a bubble pop or bare-minimum a release of air from that bubble, which will probably kill off a huge chunk of the industry, leaving just the largest, too-big-to-fail players still intact, who can then gobble up the rest of the dying industry at a discount.

Those who have the infrastructure, who have invested the huge sums of money to build these data centers, basically, will be in a prime position to survive that extinction-level event, in other words. So they’re all scrambling to erect these things as quickly as possible, lest they be left behind.

That construction, though, is easier said than done.

The highest-end chips account for around 70-80% of a modern data center’s cost, as these GPUs, graphical processing units that are optimized for AI purposes, like Nvidia’s Blackwell chips, can cost tens of thousands of dollars apiece, and millions of dollars per rack. There are a lot of racks of such chips in these data centers, and the total cost of a large-scale AI-optimized data center is often somewhere between $35 and $60 billion.

A recent estimate by McKinsey suggests that by 2030, data center investment will need to be around $6.7 trillion a year just to keep up the pace and meet demand for compute power. That’s demand from these tech companies, I should say—there’s a big debate about where there’s sufficient demand from consumers of AI products, and whether these tech companies are trying to create such demand from whole cloth, to justify heightened valuations, and thus to continue goosing their market caps, which in turn enriches those at the top of these companies.

That said, it’s a fair bet that for at least a few more years this influx in investment will continue, and that means pumping out more of these data centers.

But building these sorts of facilities isn’t just expensive, it’s also regulatorily complex. There are smaller facilities, akin to ENIAC’s campus location, back in the day, but a lot of them—because of the economies of scale inherent in building a lot of this stuff all at once, all in the same place—are enormous, a single data center facility covering thousands of acres and consuming a whole lot of power to keep all of those computers with their high-end chips running 24/7.

Previous data centers from the pre-AI era tended to consume in the neighborhood of 30MW of energy, but the baseline now is closer to 200MW. The largest contemporary data centers consume 1GW of electricity, which is about the size of a small city’s power grid—that’s a city of maybe 500,000-750,000 people, though of course climate, industry, and other variables determine the exact energy requirements of a city—and they’re expected to just get larger and more resource-intensive from here.

This has resulted in panic and pullbacks in some areas. In Dublin, for instance, the government has stopped issuing new grid connections for data centers until 2028, as it’s estimated that data centers will account for 28% of Ireland’s power use by 2031, already.

Some of these big tech companies have read the writing on the wall, and are either making deals to reactivate aging power plants—nuclear, gas, coal, whatever they can get—or are saying they’ll build new ones to offset the impact on the local power grid.

And that impact can be significant. In addition to the health and pollution issues caused by some of the sites—in Memphis, for insta