Datacenters

Why Do We Need Datacenters?

The internet today runs on datacenters. And it's not just for AI, though the recent tsunami of datacenter construction was kicked-off largely due to AI. As alluded-to above, this is when many people first became familiar with the idea of a datacenter and may unfairly attribute them as serving only AI expansion.

But datacenters have been around for decades.

I myself worked in numerous datacenters back in my corporate I.T. life. These ranged from being as small as a typical home (what we called a computer room) to being as large as a Walmart (what we'd call a proper datacenter).

Your email account, your bank, all your social media sites, your phone and its apps, your doorbell camera, and very nearly everything else, are all powered by datacenters that are connected to the internet via highspeed fiber optic lines.

What Exactly is a Datacenter?

This section goes into some detail on what constitutes a data center. These details will help you understand why new datacenter construction is so controversial. After reading this, you'll have a better understanding than most.

A datacenter campus contains most or all of the following major components:

• There's the datacenter building itself that houses all the servers, networking equipment, and some limited office space.
• Power distribution frames that take in utility-grade power, steps down voltages, then distributes power to the various loads inside the datacenter.
• Chillers pump cold water into heat exchanges inside the datacenter, which chills the interior air (removing heat) to keep all those servers from overheating.
• Outside cooling towers that dissipate heat via evaporation (more on all this below).
• Some campuses, especially hyperscalers, have their own power generation as well.

Datacenter Building

The datacenter building itself is a cavernously large, low-rise structure. There's typically some limited office space, restrooms, and a breakroom with maybe a kitchenette.

The vast majority of the interior space is segmented by function. The largest functional space is the "data hall" where all the servers live. Other areas include mechanical, electrical, storage, docks, etc. The data hall is filled with hundreds or sometimes thousands of server cabinets.

A server cabinet, or "cab" in the biz, is more or less the size of a refrigerator. That is, typically 24 to 30 inches wide, 36 to 48 inches deep, and perhaps 7 feet tall.

A modern datacenter may have dozens of long rows, each comprised of dozens of cabs. They are arranged alternating front to front and back to back. The cabs are packed in side-by-side with no space between them.

The purpose of this alternating front/back arrangement is for air flow management. We want to minimize cool and warm air from mixing. The front of the cabs face the "cool side" aisle while the back of the cabs face the "hot side" aisle.

Cool forced air comes down into the cool side from elevated ducting, enters the front of the cab, draws heat away from the equipment inside, then exits the back of the cab into the hot side and from there is forced back up into the elevated return ducting. Some extraordinarily high powered servers that get really hot may even have on- chip liquid cooling.

The width of the aisles can vary, but generally the cool side is wider, 5-6 feet or so, to allow more room for the tech to move heavy equipment into and out of the cabs. The "hot side" may be a foot or two narrower because equipment is not generally installed or removed from the hot side and the techs don't spend as much time there.

This equipment is heavy! A single cab full of equipment can easily weigh upward 2-3 tons (more than some cars!) depending on exactly what it's holding. Techs will use a rolling powered hoist to help install and remove equipment safely.

Here's a simple birds-eye diagram showing 12 cabs per row. In reality, a row could have many dozens of cabs, depending on space, air handling, and other factors.

F=Front of cab, B=Back of cab.

----cool side/aisle----

F F F F F F F F F F F F ←Front

B B B B B B B B B B B B ←Back

----hot side/aisle-----

B B B B B B B B B B B B

F F F F F F F F F F F F

----cool side/aisle----

F F F F F F F F F F F F

B B B B B B B B B B B B

----hot side/aisle-----

B B B B B B B B B B B B

F F F F F F F F F F F F

----cool side/aisle----

Some datacenters take this a step further with cool side containment or hot side containment to further isolate cool and warm air. Air flow management is a major part of datacenter design and operation.

These cabs are arranged in rows, not unlike the aisles in a Home Depot or Lowe's, but generally much longer and sometimes narrower.

The server cabs and the aisles might run "north and south" for hundreds of feet, but every 25-50 feet or so, there may be an "east-west" cross aisle to make it easier for a server technician to access the back side of the cabs where power cabling is often routed and also for faster emergency egress. We don't want personnel trapped in a 300 foot long aisle without a shortcut for escape.

All those server cabs consume a lot of floor space. So these datacenters tend to be pretty huge. Hyperscale datacenters, the really big ones, could be hundreds of thousands or even a million-plus square feet, dozens of acres -- the size of an Amazon.com regional distribution hub.

Electricity

All datacenters, small to huge, are pretty power hungry relative to their size compared to other industrial users. Here's how much power datacenters of various sizes might draw.

• Small datacenters, typically upward 20,000 sq/ft, consume upward 8 MW* of power.
• Medium-sized datacenters, typically upward 100,000 sq/ft, consume upward 30 MW of power.
• Large datacenters, typically upward 250,000 sq/ft, consume upward 75 MW of power.
• Hyperscale datacenters, 250,000 to over 1,000,000 sq/ft, can consume 75 to 600+ MW of power, nearly the entire output of a medium-sized power plant.

* MW=MegaWatts. So 8 MW = 8,000,000 Watts.

Larger and hyperscale datacenters, the kind often built today, are among the highest single consumers of electricity on the planet especially so when considering their size and copious numbers. For datacenters that specialize in AI computing, the power required is even higher -- far higher -- upward 10 gigawatts.

Explainer

All those figures just above can be pretty nebulous so here's some examples to help make sense of all that.

A typical single family home (SFH) draws around 30 kWh* per day including intermittent loads like the air conditioner and large electrical kitchen appliances. That works out to 1.25 kW levelized power draw (1.25 kW x 24 hrs = 30 kW. Keep your eye on that number, 1.25 kW.

* A kWh (kilo Watt hour) is a measure of power over time. No sweat if you aren't familiar with this term, it'll all come together in a minute.

Here's a list comparing the power draw of the various data center sizes with how many SFHs that would equal. These aren't exact but they are definitely in the ballpark.

• Small datacenter ≈ 6,400 SFHs
• Medium datacenter ≈ 24,000 SFHs
• Large datacenter ≈ 60,000 SFHs
• Hyperscale datacenter ≈ upward 750,000 SFHs
• Stratos Project* datacenter ≈ upward 7,200,000 SFHs

* A massive AI-specific datacenter being planned for Box Elder county, Utah, has a planned capacity of 9 gigawatts. That's about twice the existing power consumption for the entire state.

That should give you an idea of how much electricity various sized datacenters consume. And because of the law of conservation, all that electricity is ultimately converted to heat.

That's a lot of heat to remove and we'll discuss that in the next section, "Cooling".

These numbers are all fairly rough as there are a number of individual variables. But these are fair analogies nonetheless.

Did that explainer help?

So it stands to reason that a datacenter can't just roll into town and "plug into the grid" -- it forces the grid to be reshaped around it. Utilities and regional operators have to plan and build out substations, transmission lines, supporting infrastructure, and even new generating capacity, often at very high cost. All that takes time -- years.

Developers will say they’re paying their share and they usually do cover the direct hookup costs and actual consumption. But a big chunk of the bigger, more expensive upgrades, the ones that make the whole system capable of handling that new load in the first place, are frequently pushed on to all customers. They are effectively underwriting part of the expansion through higher monthly electricity bills.

The risk doesn’t go away after construction, either. These projects are often justified on upward 25 year demand projections. If the datacenter scales back, under performs, or shuts down early, the grid infrastructure is still there. That cost doesn’t just disappear, it gets spread even more so across everyone else's electric bill.

Even in the success case, there’s a catch: Even when accounted for, a massive new load tightens regional supply margins, reducing grid flexibility. That can push up wholesale electricity prices, which then flow through to retail rates. So ratepayers may still end up paying more even if the datacenter meets all its commitments.

The sales pitch is long-term economic growth. The underlying reality is that a large, private customer is being supported by long-lived public infrastructure with much of the downside risk sitting with the public.

Performance or completion bonds and other financial instruments can help mitigate that risk. But municipalities courting a datacenter will almost certainly not have an equivalent level of negotiating skill and savvy to ensure their existing rate payers are sufficiently protected. Datacenter companies have armies of high-dollar lawyers to negotiate on their behalf. Some podunk rural town or county commish almost certainly will not.

Here is an eye-opening 32-minute podcast on NPR's Planet Money that takes a deep dive into all the power grid dynamics needed to support a datacenter. The podcast is not overly technical and is quite easy to understand. It addresses the question of how and why local electricity bills increase due to new datacenters coming online.

The phrase, "socialize the costs and privatize the profits", a foundational credo for many companies with significant negative externalities*, fits nicely here.

* Externalities: The effects, good and bad, that outside parties, like you and me, experience due to the actions of the first party (a company). e.g. Pollution from a factory is a negative externality.

BYOE

Bring Your Own Electricity. That’s my term, not an official one, for when a datacenter plans to introduce new power on their own rather than relying exclusively on existing grid capacity or expecting the local utility to ramp up production through expansion projects.

Given that datacenters need so much power, having to build on-site or near-site power generation is becoming more important. But how are these power plants fueled?

Building a new coal-fired plant is a non-starter for many reasons so that's not going to happen.

What's left?

▶︎ Natural Gas. The most common choice. It's relatively cheap and quick to build. But the idea of producing that much local pollutants (Carbon Dioxide, Nitrogen Oxide, Methane, etc.) doesn't sit well with local communities and for good reason. These are large power plants that will spew out a lot of pollutants.

▶︎ Grid + Local Renewables. This is a traditional grid-tied datacenter with some power coming from wind, solar, and possibly hydro. While these green energy sources are good to see, the main problem is the huge power needs of datacenters (compared to other uses) which is hard to practically meet with renewable sources.

▶︎ SMR (Small Modular Reactor). This is a small purpose-built, onsite nuclear reactor similar to that used in a nuclear sub. These nukes are built elsewhere and trucked to the site. Nuclear is still the clear winner for energy density. Inputs are plentiful, emits zero greenhouse gases, and is incredibly energy dense. A single onsite SMR or two could easily power a large datacenter.

How energy dense? Consider that a single 7 gram uranium fuel pellet the size of your fingertip contains as much energy as...

• 17,000 cubic feet of natural gas
• A ton of coal
• 150 gallons (about 3.5 barrels) of oil.

A pellet of uranium has more than 100,000 times the energy density of coal.

SMRs are still quite expensive, on par or even exceeding large scale nuclear on a per GW basis. We're still in the early phase of designing ideal purpose-driven SMRs. Unfortunately, this solution will take quite a long while to develop, indeed, if it ever does.

And just to say, nuclear doesn't have to be unsafe. There are plant designs that can "fail safe" if the need arose. Alas, we've long had a difficult relationship with nuclear power for many reasons -- technical, yes, but also to a great extent political.

As ever more power-hungry datacenter construction projects are launched, we’re seeing more of this BYOE methodology. Is it enough? Probably not.

Cooling

Having many tens of thousands of servers crammed in a building all busily answering billions of google searches, serving videos, processing business data, booking your flights, handling your email, a million other things, and increasingly, AI processes, will consume a lot of power which creates a lot of heat.

As you might remember from your middle school science class, energy is not created or destroyed -- it only changes form. This is the law of conservation I mentioned above. All that electricity ultimately turns into heat that must be removed.

Let's use an analogy to more easily comprehend just how much heat a datacenter produces.

We'll use space heaters in our analogy -- an appliance that everyone is familiar with. A typical household space heater draws about 1,500 Watts* on high and will maintain comfortable warmth in a typical 150 sq/ft bedroom (assuming wintertime). That's about 10 Watts per square foot to keep that room comfortable. Easy math, yes?

* It's actually closer to 1,400 Watts for electrical code reasons. But 1,500 is a round number making the example math easier.

Using the figures above, a large 250,000 sq/ft datacenter consuming 75 MW (75,000,000 Watts), would equal around 50,000 space heaters. On a per sq/ft basis, that's around 300 Watts of heat per sq/ft (in the data hall), compared to only 10 Watts per sq/ft for the bedroom. 30x more heat per sq/ft. Imagine how hot that bedroom would get if there were 30 space heaters all running 24/7 assuming the home could even support that much electrical draw (it cannot). Mighty toasty.

And that's just for a regular datacenter. Datacenters that focus on AI can produce 10x more heat per sq/ft -- about 300 space heaters in the bedroom analogy. That's not toasty, that's a raging bonfire.

There are two (main) ways heat is removed from a datacenter.

Evaporative cooling

Historically, the most common cooling method has been open-loop evaporative cooling -- a "swamp cooler". Treated (and sometimes potable) water from a local water treatment plant is pumped through a heat exchanger that sucks heat out of the air inside the datacenter, thus heating the water, which is then pumped back outside into cooling towers where it evaporates as steam thus releasing heat into the atmosphere. This method is sacrificial -- meaning water is not returned to the system. It evaporates into thin air.

Now that doesn't mean the water stops being water! That water (as steam) is absorbed by the atmosphere where it continues on to the next step of the circular hydrological cycle -- much like how lake water evaporates then falls as rain, etc.

But the incoming water used to cool the datacenter has to come from somewhere. And that "somewhere" often includes municipal water providers. That's because the cooling systems can't deal with dirty or contaminated water. Like anything that relies on water for life, both flora and fauna, datacenter cooling systems work better with clean water.

Evaporative cooling is more energy efficient and performs better in hotter climates as long as the humidity isn't too high, but at the cost of requiring a continuous, non-stop flow of vast quantities of clean water. That's the big problem right there.

Even a mid-sized datacenter can consume between 250 and 500 thousand gallons of water per day. Such a datacenter would be far and away the single largest user of water for any city under 1/4 million population.

The short and long term impact on the local watershed and water purification plants must be considered.

Non-evaporative cooling

In closed-loop non-evaporative cooling, far less water is lost from the system. It's not zero because of the initial fill and any system flushes for necessary maintenance. But even then, losses are still a tiny fraction of what an open-loop evaporative system would lose.

Most new datacenter projects today are designed with non-evaporative cooling and it's a good thing to see. There's still a lot of legacy datacenters using evaporative cooling, however, but there's a push to convert them to non-evaporative cooling especially where the local watershed and water authorities are under stress.

Closed-loop systems can be less energy efficient, require larger heat exchangers, and don't operate as well in higher temperature regions, thus cost more for the same amount of cooling, But they collectively save literally millions of tons of water.

Site Selection

Datacenters have unique requirements and challenges that indicate and affect where they can be built. Lets go over that in no particular order.

▶︎ Power availability: Is there excess grid capacity in the region? And if not, is the region amenable to expansion? Or will the datacenter owner build a power plant?

▶︎ Cooling feasibility: What's the average year-round climate in terms of heat, humidity, precipitation, etc., and water availability?

▶︎ Network connectivity: Relative proximity to fiber infrastructure and IXPs (Internet Exchanges)

▶︎ Land: Depending on the planned size, anywhere from 50 to 200+ acres could be needed for a datacenter campus.

▶︎ Local Acceptance: Is the locality amenable to having a datacenter nearby? Or will permitting and community buy-in be a long, difficult, or impossible process?

▶︎ Incentives: Will the region offer sales and property tax abatements and other incentives?

Surrounding semi-rural exurban counties tend to work best. Generally not close enough to a large city to be heavily negatively affected but close enough to benefit from its proximity.

e.g. To the extent that sufficient power, feasible cooling, and networking is available (or can be) then remaining items tend to be easier. Land is much cheaper, taxes likely lower, community acceptance more likely, or at least less pushback, etc.

Jobs Jobs Jobs

Datacenter operators will tout the numerous jobs they'll create. And what small town council or county commissioner doesn't like that?

Some facts about those jobs.

Construction

Construction jobs are very real and pretty well paying, however... A typical mid-sized datacenter takes about a year or so to build from start to end. After that, it's done, and the construction jobs go away. But even while its under construction, there's a good chance that some/most(?) of that labor will be "imported" -- that is, brought in from outside the county or even the state.

Why is that?

Unless the datacenter is going into or near a fairly major population center, which is often not the case, then your rural(ish) community or county may simply lack the labor pool to turn out in the numbers the project will require. Even a mid-sized datacenter requires potentially several hundred skilled construction workers in order to meet scheduling milestones.

Then there's the issue of expertise. Most rural local construction labor has probably not worked on a datacenter before. While the exterior prep work (grading, concrete, etc.) and the building shell itself is pretty standard, much of the interior fit-out work for a datacenter requires specialized knowledge that local workers may lack. Consequently, many datacenter interiors, especially those in more rural areas, are built by traveling crews of workers.

Those crews roll into town, staggered by construction phase expertise, shack up in the local hotels, motels, extended-stays, etc. until their part is done. Then they pack up and leave.

It's good but very temporary business for local accommodations, restaurants, that sort of thing.

Ongoing Staff

When the datacenter is finished and opens for business, the amount of ongoing onsite-labor required is surprisingly minimal for something that big. Datacenters are highly automated and remotely administered installations requiring little hands-on staffing.

Here's some of what the permanent on-site staff does.

• Facilities and infrastructure, electrical and mechanical
• I.T. operations (install and replace servers and networking gear)
• Shipping and receiving (compute equipment and other stuff)
• Site security
• Interface with visiting vendors and clients
• Janitorial

A medium sized datacenter may only need 25 to 50 people to cover everything 24/7/365. Some of those staffers may be hired from the local labor force, but others may be transferred-in especially for the more technical roles.

Contrast that with, say, an Amazon fulfillment center of similar size, that would employ well over a thousand local workers. A Walmart store, also similar in size, typically needs 200 to 400 workers. And those jobs don't require the level of technical knowledge that datacenters require, making them more compatible for entry level workers.

Local Politics

Datacenter companies know they aren't particularly welcome in some communities. So they have a playbook of strategies for helping to win over local opposition. I've already touched on some of this above, but here's a more complete and tidy list of strategies designed to overcome objections.

• Construction jobs (These are temporary jobs and often include imported workers)
• Ongoing permanent jobs (These jobs are far fewer in number and may also include imported workers)
• Various tax windfalls (But may be reduced through negotiation or by other abatements)
• Upgrades to local infrastructure: roads, power, fiber internet (this can actually happen, not just puffery)
• Direct investment/donations/sponsorships to schools, workforce training, non-profits, civic affairs like police, fire, EMS (often a one-shot injection)
• Claims of environmental stewardship
• Local bump to economy (real, but limited to certain businesses and mostly during construction, nothing much enduring)

Some of these incentives are real and would be welcomed by most communities. But, like any advertising, you're only getting one side of the narrative. Best-case outcomes are presented as typical and routine. It's important to maintain incredulity and not simply believe what is promised without iron clad performance guarantees.

The fuller truth is many of these incentives are shiny superficial objects, designed to inexpensively flatter and appease voters and local officials. They're often door prizes in the grand financial scheme of datacenter operations which can easily reach into the billions of dollars.

♫ NDA, Well, You Don't Say ♫ *

When a Big Tech company approaches a city or county expressing interest in building a local datacenter, they will require local officials and other key players to sign an NDA (Non Disclosure Agreement) that gags those officials and other signers from discussing the project with anyone else. You can bet your bottom dollar that NDA is designed to protect the company, not the locality.

Be warned: Transparency is a must if you expect to win community buy-in. Without it, expect howling opposition and lawsuits. That may happen anyway if the project's metrics don't pencil out. But without transparency, pushback is virtually guaranteed. Few things rile up people more than government being tight-lipped or seen to be hiding something.

This isn’t Costco, after all, guarding a negotiating edge with a property owner. Datacenters bring significant externalities on power, water, land use, and infrastructure that can reshape a region in ways that a typical commercial development never could.

* Here's a fun animated musical video short about NDAs (72 seconds) in the style of Schoolhouse Rock! (remember that from your childhood?) from the satirical TV law show The Good Fight. I'm sure you'll enjoy it. Come back here when it's done.

A Cautionary Tale

Local opposition to new datacenter construction projects are growing all over the US as the curtains are being pulled back on the often lopsided nature of datacenter contracts in terms of power use, water use, land use, and other undesirable externalities.

Local politicians and public officials would do well to learn from the story in unfolding in Festus, Missouri (a town just south of St. Louis), where half the city council was ousted in the recent elections and the other half along with the mayor are likely facing a recall vote for their role in approving a new datacenter project.

Festus city officials were accused of secretive negotiations, not listening to residents, and generally trying to quell opposition. Turns out the people didn't like that and boy did they take action. Who'da thought?

Read about that here: Missouri town fires half its city council over datacenter deal (Politico)

Another town, Port Washington, Wisconsin, passed a first-of-its-kind referendum to kneecap new datacenter projects (Politico)

More on datacenter construction pushback: Big delays in US datacenter construction (Ars Technica)

AI Fatigue

Even if -- and boy howdy that's a whopping big "if" -- Big Tech were completely honest, ethical, and forthright in their datacenter proposals, there would still be pushback. The mood toward AI is shifting negatively which is one (of the many) reasons why datacenters projects are having facing such opposition.

We are absolutely inundated with AI hype, advertising, and slop* of every conceivable kind. People are getting sick of it.

* Slop: Bulk, low effort, low quality AI produced content. Not necessarily hallucinations -- those are two different things.

But the attack on jobs is probably the biggest concern. Some 50,000 jobs have been lost so far just this year specifically due to AI according to outplacement firm Challenger, Gray, and Christmas. Softer hiring -- jobs not filled in the first place -- is seeing even larger numbers.

This is due to 1) companies automating certain jobs thus reducing the need for personnel and 2) significantly more spending on AI initiatives that reduce funding available for increased hiring.

As companies charge ahead with AI initiatives, an increasing variety of jobs are being replaced with AI, not just more of the same jobs. And a fair chunk of those employees that remain are spending more time using and training AI on their systems and processes which just makes the outlook for workers that much worse going forward.

Even if your job is safe from AI (maybe you're a plumber?), there's a good chance you know people whose jobs may not be so safe. With such slow hiring today, college grads and others entering the workforce are finding fewer opportunities. And that's amid record gains in the financial markets.

Looks like the skilled trades are gonna get more recognition! I have mad respect for people that do tough jobs with their hands for a living. I'm a pretty handy DIYer around the house, fixing most things that need it -- plumbing, electrical, construction. So I've got a small taste of what that's like.

So it's my opinion that at least some of the pushback against datacenters is borne from deeper societal concerns -- not just electricity, water, and pollution discussed above.

Final Comments

In all this, it may appear that I'm against datacenters, but that is not necessarily true! No part of today's modern world would be possible without them.

So why then do I come across so negative even while acknowledging their necessity?

Because Big Tech doesn't need my skinny ass helping to tout the benefits of having a datacenter move in. With a market cap in the 14 figures, Big Tech is perfectly capable of expressing themselves.

But what is needed is a detailed counternarrative, discussing the other side of the datacenter coin. Because you aren't going to hear any of that from Big Tech. They don't really care about you, your town, or your county. They care about making money, full stop.

That's why I sound so negative.

And that's why a competent negotiating strategy is important.

Towns and counties that are considering welcoming a datacenter would do well to: Exercise all due caution • Stress transparency. Resist signing broad NDAs; this really nukes community trust! • Negotiate aggressively and wisely • Consider all the impacts • Contact other towns and counties that have been down this road already • Hire outside legal council • Don't get distracted by shiny things • And make damn sure the good of the community is always top of mind.

If you made it to here, then thank you for your time. This was a long one. But you should have a pretty good idea of what goes on inside a datacenter, how a datacenter affects a locality, and how Big Tech tries to sell datacenters. Use this as a step toward better educating yourself on the pros/cons of having a datacenter and perhaps even becoming a community voice.