“People say “No AI Datacenters” and I’m right there with them… But to say “All datacenters are bad” is a gross oversimplification of the issue and stands to limit actual, beneficial, economic growth. Datacenter is a catchall term for several different kinds of building that, yes, are all ultimately giant rooms stuffed to the brim with computers but have a myriad of purposes. Some of them, even, are good. And in this era of hot-button politics, I think we all have a duty to understand what exactly we’re screaming at…”

I’m not afraid to go on the record and say it: I think OpenAI and Anthropic have a colossal nothingburger of a product that they’ve used to con CEOs, venture capitalists, the media, and everybody else desperate to be part of the “Next Big Thing” out of hundreds of billions of dollars that could’ve gone to creating things that actually benefit society.

Large Language Models, LLMs for short, draw staggeringly large amounts of power and use mind-boggling amounts of water (most of it evaporated off or irrevocably poisoned and returned to the sewer system for reuse) to do the exact same thing that your phone’s text suggestions do. The primary difference is that, unlike your phone, they have enormous tranches of reference data to draw their predicted words from– primarily because the ones at the levers of power decided that copyright law doesn’t apply to corporations that violate it en masse.

Transforming models, which were introduced with GPT-1, are no better: they required yet another scrape of every image and video on the internet so they could take TV static and change the color of each individual pixel and create recognizable images out of it, all so Sam Altman and his ilk could suddenly pretend that they were, and I cannot use this term loosely enough, “artists”.

People say “No AI Datacenters” and I’m right there with them. I have no desire to share a town with a Godzilla-sized cube that makes a whistling scream 24/7 just so people can tell a computer to generate increasingly busty Garfields.

But to say “All datacenters are bad” is a gross oversimplification of the issue and stands to limit actual, beneficial, economic growth. Datacenter is a catchall term for several different kinds of building that, yes, are all ultimately giant rooms stuffed to the brim with computers but have a myriad of purposes. Some of them, even, are good. And in this era of hot-button politics, I think we all have a duty to understand what exactly we’re screaming at our politicians to get off their tails and do– otherwise, we’re ultimately buying into the same con that I just railed against: that LLMs aren’t a furnace powered by our retirement accounts and are instead the inevitable future of modern computing.

There are four major types of datacenters: enterprise datacenters, managed datacenters, colocation datacenters, and hyperscale datacenters.

Enterprise datacenters are, as the name implies, private facilities owned and operated by a single organization. They house all the servers, networking equipment, and general infrastructure required by that organization for its operations. The easiest way to think about these is actually to reach back to the early age of computing and treat them as one giant mainframe that the workstations in the organization then communicate back to.

Managed datacenters are datacenters that rent out their resources to other companies. They take care of all the headaches of getting power to server racks, populating those racks with servers, getting everything up and running, getting it connected, etc. They’re a turnkey solution that organizations pay a monthly fee to keep using. They’re also very hands-off: you’re paying the datacenter company to manage everything, but because it “just works” most of the time the renting company has no control over the datacenter or its operations.

Colocation datacenters are, in essence, the computing equivalent of apartment complexes. These datacenters rent out space in their server racks to other companies, who then bring their own hardware in. A colocation datacenter can have as many different clients as they have available space, each of them paying for the convenience of not having to worry about things like security, where the power’s coming from, etc, while still maintaining control over their individual hardware.

And then there’s Hyperscale datacenters. They’re built to house server racks with complex and demanding requirements for power and cooling, and typically have their own power plants onsite. These are the giant datacenters (a “small” 1 megaWatt facility in Raleigh is the size of two houses and draws the same amount of energy as 33 homes) that house AI infrastructure– and are also what people tend to think of nowadays when they hear the phrase “datacenter”.

The important consideration, though, is that only one of these is actually the giant screaming cube I talked about earlier. Colocation, Managed, and Enterprise datacenters all have their place in running the technology we use every day, and without them many businesses genuinely would suffer. To paraphrase Abigail Spanberger’s recent remarks when asked about datacenters, “everybody wants the iPhone apps and everybody wants the data that runs them, but nobody wants the datacenter those things need”. In my eyes, that’s not coming from some misguided NIMBYism– it’s coming from a lack of nuance.

Everybody, myself included, really *does* want the iPhone app and the data that drives it. And I think if most folks understood what was trying to be built they wouldn’t have a problem with it. What we don’t want is to have to house the ugly, dirty, screaming infrastructure to power a lie that nobody actually thinks is beneficial to society.

So at the next City Council meeting, my advice is this: think about what you’ve read here, and then ask them straight-up: “What kind of datacenter is coming to Salisbury?” If it’s not a hyperscaler, maybe stop and think about the benefits. But if it’s a hyperscaler? I’ll be right there next to you, screaming as loud as I can.