Why Pausing the Data Center Buildout Is the (Humanly) Intelligent Choice |
For a variety of reasons, I’ve found the data center debate to be difficult to get a real handle on over the last year. But I think a clearer picture is beginning to emerge, and I will do my best here to share it with you. Remember, I’m just one human brain, and I have not (illegally) digested every single book ever printed; I can’t draw you a picture of a data center licking an ice cream cone; and if you asked me to render this essay in the style of Emily Dickinson I would fail. Still, for what it’s worth:
First source of confusion: How much demand for AI will there actually be?
This depends on how useful it turns out to be, and that is still a very open question. Yes, AI executives are busy insisting it will upend everything and everyone—the AI chief at Microsoft said last week that all white-collar jobs using computers will be wiped out in the next 12 to 18 months—accountants, project managers, marketing staff. But there’s another school of thought—most ably represented by an AI researcher named Gary Marcus—that thinks the hallucination-prone large language models are good at writing certain kinds of code but not getting much better, and in fact may be at about the limit of their abilities.
There’s a second question resting on top of that one: Whatever AI can do, will it make a lot of money doing it, thus justifying the enormous investments currently being made or planned for data centers? The stock market apparently thinks so—AI makes up some stupendous percentage of its gains in recent years—but there are, as you have heard, fears it might be a bubble. The most eloquent—indeed logorrheic—source of those fears is Ed Zitron, a blogger who has followed the various money trails and concluded that companies like OpenAI and Anthropic have no real prospect of making back the scads of money that they’ve spent, and that sooner or later the bubble will indeed do the thing bubbles do.
If we reach the point where we decide as a society that we actually want to build out this technology, then BYOG should be replaced by BEYONCE—Bring Your Own New Clean Energy.
These are crucial questions for us because as long as the bubble keeps expanding, there will be insatiable demand for more electricity for more data centers, and if it pops that demand will start to drop dramatically, especially since much of it is still semi-speculative—that is to say, there are far more data centers on the drawing board (to use an old-fashioned image) than under construction.
In fact, it’s been remarkably hard to estimate how much demand for electricity is actually going to go up, precisely because there’s so much speculation here. In an interview that got pretty wonky even for him, the invaluable David Roberts last week talked to Clara Summer, a public advocate at the PJM Interconnection Board, PJM being the the largest regional transmission organization (RTO) in the United States, managing the high-voltage electric grid for 67 million people across 13 states from Delaware to Illinois. Anyway, Summer explained that any given data center might be applying for permits to build in four or five different jurisdictions:
There is a big difference between a data center that has knocked on the door of a utility and said, “I am interested in being in this area,” versus a data center that has entered into a contract with a utility and put down money.One estimate has that the number of requests for potential data centers to connect to the grid is 5 to 10 times more than the number of actual data centers that will be built.
Obviously, however, there are plenty of data centers going up. Some are truly terrible (consider the joint investigation by Floodlight News and the Guardian of an xAI facility in Mississippi; it has followed the path of Elon Musk’s egregious data center in the poor part of Memphis, both using portable gas turbines that pollute the air, and all in an effort to support an artificial “intelligence” that goes on long happy rants about Hitler; it won’s surprise you that the NAACP was early in expressing concern) and some are less terrible: Google just signed up for two big solar farms in Texas to support its data centers.
The default, sadly, seems to be headed toward the Musk model. With grid providers unable to build generating capacity fast enough to keep up with demand, data center developers are going BYOG—bring your own generation. Here’s a long and detailed new report about how the G generally turns out to also stand for gas, in this case onsite gas turbines, with not much concern for the climate or local air pollution risks. (Or for the amount of water required—here’s a recent account from Brad Reed of a single Pennsylvania data center that will use 40% of the town’s excess water). Here’s a kind of worst-case scenario from John Kostyack, a DC-based consultant:
By the end of this decade, capital spending by tech, real estate, and utility companies will likely represent the largest private-sector infrastructure spending spree in world history. McKinsey, for example, estimates a whopping $6.7 trillion in capital expenditures by 2030.Although forecasts of the scale of data center buildout vary widely, anything near this projected scale has enormous climate implications. The most obvious concern is the emissions generated in powering the massive hyperscale complexes, which are being designed to consume as much as 2 gigawatts (GWs) of power–roughly 15 times the capacity required by the entire city of Philadelphia during summer peak load. According to energy analyst Rystand’s 2025 review of industry announcements, data centers consuming up to 100 GWs of power could come online in the next 10 years.Much of this power would come from gas-fired power plants. Researchers at........