We’ve all heard the pitch. AI is going to solve cancer, write your emails, replace interns, and maybe even find your missing socks. It feels like magic. You type a prompt into a box, and a few seconds later, a computer spits out something that looks smart enough to charge consulting fees. But here’s the thing: magic isn’t free. Behind every "magic" response is a massive, humming building full of chips, cooling gear, backup systems, and power equipment that eats electricity like a hungry teenager at a buffet.
At Regular Guy Economics, we like to pull back the curtain on the numbers that actually matter to your wallet. Right now, the tech giants are engaged in one of the biggest spending frenzies we’ve ever seen. They are building data centers: giant warehouses full of super-hot, power-hungry computers designed to train and run AI models.
And this is not some abstract Silicon Valley hobby. This buildout touches the electric grid, local water systems, land use, mining, and utility bills. The bill is coming. Not just for Microsoft or Google, but for you.
The funny part is that AI is sold as weightless. "The cloud" sounds soft and fluffy, like your photos are floating around on a happy little puff of vapor. In reality, the cloud is concrete, steel, copper, diesel backup, cooling pipes, substations, and transmission lines. It is a giant industrial machine wearing a clean white tech hoodie.
That matters because when Wall Street hears "AI," it hears growth. When your local utility hears "AI," it hears new generation, new wires, new transformers, and new rate cases. When your town hears "AI," it should probably hear one simple question: who pays for all this?
The Sheer Scale: Hundreds of Billions, Maybe Trillions
The money being thrown at AI right now is hard to wrap your head around. Hyperscalers like Microsoft, Google, Amazon, and Meta are leading the charge. In 2025 alone, global spending on data center infrastructure is expected to hit $580 billion.
To put that in "Regular Guy" terms: that’s more than the entire GDP of countries like Norway or Thailand spent on building server rooms, chip clusters, substations, cooling systems, and transmission support in one year. By 2030, analysts expect this total to balloon toward $3 trillion.
Companies like CoreWeave are popping up to meet the demand, reporting revenue backlogs of $99 billion. That sounds amazing until you remember backlog is not the same thing as cash in the bank. It means customers are lined up. It does not mean the economics are proven, the margins are safe, or the hardware will age gracefully.
And the upfront price tag is nasty. A lot of these AI-heavy projects now pencil out at roughly $11 million to $13 million per megawatt of capacity. That is eye-watering money. If you are trying to stand up a 100-megawatt campus, you are not talking about a nice office expansion. You are talking about a multibillion-dollar industrial project with real construction risk, power risk, financing risk, and demand risk.
So yes, the market is excited. But let’s be honest about what is happening. We are betting the farm on future AI revenue streams that have not fully shown up yet. We have excitement, projections, and PowerPoint decks. What we do not have, at least not yet at the scale needed, is proof that all this spending will create enough durable profit to justify the tab.
The Grid Squeeze: From Background Load to Main Event
Here is where the math gets messy for the rest of us. For decades, data centers were a smaller blip on the national power grid, using maybe 2% to 4% of all U.S. electricity. But by 2028, that number is projected to leap to 12%.
Think about that. One out of every eight lightbulbs in America, in effect, could be powering a GPU somewhere.
But the real problem is not just how much electricity AI uses. It is how it uses it. AI workloads can create giant power ramps, especially during intense training runs. Demand can jump by hundreds of megawatts in seconds. That is not a normal office park. That is "steel mill" behavior. Our grid was not built for a wave of facilities that can hit it with industrial-style jolts on command.
Utilities like steady, predictable demand. They can plan around that. AI clusters are a different animal. They can go from cruising along to hammering the system in a blink. That forces utilities and grid operators to think about more generation, more reserve margins, more transmission, more substation upgrades, and more expensive equipment that can handle the swings.
It’s like plugging a casino, a steel mill, and a space heater the size of Ohio into the same extension cord and then acting surprised when somebody needs a new transformer.
And remember: America’s grid is not exactly fresh off the showroom floor. A lot of the transmission system is old. Interconnection queues are backed up. New power plants take time. New high-voltage lines take even longer. So when AI demand shows up fast, the physical grid cannot just stretch like sweatpants. Somebody has to build the hardware.
The Financials: A Long-Term Gamble With a Very Short Fuse
If you look at the books, the CAPEX is brutal. Building these centers costs billions upfront. Land, power equipment, switchgear, generators, batteries, chillers, cooling towers, networking, and the chips themselves all cost real money right now, not in some happy future quarter. Then you have depreciation. In the tech world, a chip is "old" fast. That means these companies need to earn back the investment before the hardware turns from elite machine into yesterday’s expensive toaster.
This is why the $99 billion backlog at CoreWeave needs to be read with both eyes open. Yes, it shows demand. Yes, it shows customers want capacity. But the spending required to serve that demand is staggering, especially when projects can run $11 million to $13 million per megawatt. That is a gigantic upfront wager on future AI demand, future pricing power, and future utilization staying high.
That is the key point: we are still in the "build it and they will come" phase. The industry is spending like the revenue side is already mature. It isn’t. Some AI tools make money. Some save labor. Some are impressive demos with unclear business models. There is still a huge gap between "this is cool" and "this throws off enough cash to cover the buildout."
While the backlog numbers look great on paper, the Return on Invested Capital is a long game. If AI software turns into the backbone of the global economy, maybe these investments age like fine wine. If not, we may discover that we built a mountain of expensive infrastructure for lower-margin products than investors were promised.
That is the uncomfortable part of the story. Everyone talks about the upside. Not enough people talk about the possibility that the infrastructure bill is real, immediate, and physical, while the payoff is still partly theoretical.
The Hidden Tab: Water, Mining, and Concrete
Most people think of "The Cloud" as something ethereal and weightless. It’s not. It’s heavy, dirty, and very thirsty.
-
The Water War: A single large data center can use around 1 million gallons of water a day for cooling, depending on design and climate. But the sneaky part is the hidden water use. The power plants feeding those facilities can use 10 to 18 times more water than the data center itself, especially if the power comes from thermal generation. So when a tech company says, "Don’t worry, our direct water use is manageable," that is a little like telling your spouse the boat was cheap while leaving out the marina bill, the gas bill, and the repair bill.
And this is not happening in water-rich wonderlands. In the Texas Panhandle, projects like "Matador" are looking at tapping the Ogallala Aquifer, which is not some random puddle. It is the lifeblood for local agriculture, especially cotton farming. In Arizona, data centers are showing up while residents and farmers deal with Colorado River cuts and tighter water realities. Strip away the buzzwords and it becomes very simple: farms vs. servers.
-
The Mining Tax: AI hardware needs more than electricity. It needs specialty metals, magnets, and components built from materials like Neodymium and Dysprosium, which are key rare earth elements used in high-performance systems. And getting those materials is ugly. For every ton of rare earth minerals mined, the process can generate roughly 1.4 tons of radioactive waste and about 60,000 cubic meters of toxic gas, including compounds tied to hydrofluoric and sulfuric acid processing. This is not just "mining." It is a chemistry set gone wrong. It can contaminate groundwater, damage soil, and leave communities with the cleanup tab while someone else enjoys the stock upside.
Add copper to the list too. AI needs mountains of copper for cables, transformers, substations, and grid expansion. Analysts are already warning about a potential major copper shortfall by 2035. So the AI buildout is not just a software story. It is a giant raw-material story.
-
The Carbon and Concrete Tab: We often talk about emissions from the power plants, but don't forget the concrete and steel. Building these giant campuses requires huge volumes of cement, one of the most carbon-intensive materials on the planet. So even before a chatbot answers its first dumb question, there is already a sizable environmental tab baked into the slab, walls, and utility infrastructure.
This is the part nobody puts in the commercial. AI looks clean on your screen. Off-screen, it is pipes, mines, chemicals, cooling towers, diesel backup, and a whole lot of water moving around dry places.
The "Regular Guy" Perspective: Who Pays the Bill?
Now, let’s get to the part that affects your dinner table. Who pays for the new power plants, substations, transmission lines, and grid upgrades needed to support these AI giants?
In many cases, the utility companies pass a meaningful chunk of those costs down to consumers over time. If a data center moves into your county and the utility has to build a new $500 million substation, upgrade local lines, and chase new generation just to keep the servers happy, do not be shocked when your monthly bill starts looking puffier.
This is the part that should make the average person sit up. If tech giants are willing to pay almost any price for power, local utilities have every incentive to build more infrastructure. And utilities do not build things out of kindness. They build them and then go looking for a way to recover the money. Guess where that recovery usually lands? On ratepayers.
And it is not just the grid. It is the water systems too. Bigger pipes, more treatment, more cooling support, more pressure on local supply. If a data center can use town-sized water volumes while the grid absorbs steel-mill-style power spikes, the average resident is basically competing with a robot warehouse for the basics.
So here is the regular guy question: Is your $200 power bill helping subsidize a billionaire’s AI chatbot?
That is not a crazy question. It is the obvious question.
AI data centers can justify paying a premium because the upside for them is massive, at least in theory. But for the small business owner, the retiree on a fixed income, or the family of four already juggling groceries, insurance, and housing, electricity is just another bill that keeps getting meaner. They do not get equity in the AI boom. They just get the rate increase.
If this buildout keeps accelerating, we need honest answers on cost-sharing, water access, land use, and grid priority. Because if the public gets the higher bills, the water stress, and the infrastructure burden while private firms get the upside, that is not innovation. That is a transfer.
A Digital House of Cards?
So, does the math work?
Maybe. But "maybe" is doing a lot of heavy lifting here.
If AI becomes the brain of the global economy, maybe the trillion-dollar buildout ends up looking smart. Maybe the chips, the campuses, the water demand, and the grid expansion all pay off because AI becomes as essential as electricity itself. That is the bull case, and it is not impossible.
But there is a much less glamorous possibility. We may be spending extraordinary amounts of real-world capital to support business models that are still being figured out in public. If it turns out we built a giant industrial base mostly to generate deepfake images, summarize meetings nobody wanted, and write marketing copy that sounds like it went to business school for six weeks, then we have a problem.
We are trading real-world resources, water, copper, rare earths, concrete, grid capacity, and your hard-earned money, for a digital bet. And unlike a bad app launch, these mistakes do not disappear with a software patch. Mines stay mined. Aquifers stay depleted. Rate hikes have a funny way of hanging around.
That is the bigger point from all this new water math. The AI story is not just about clever software. It is about a prompt that can carry a water cost, a chip that can arrive pre-soaked from manufacturing, a server farm that drinks like a town, and a power system that often burns even more water behind the curtain. That is a lot of hidden gulp for a product marketed as frictionless.
That is why the AI story should not just be told as a tech story. It is an infrastructure story, a utility story, a water story, and a public-cost story. The economics of "more" can only go so far before the grid, the environment, or our bank accounts start waving the white flag.
So before we all clap for the next chatbot demo, it might be worth asking one plain English question: is this revolution making life cheaper, easier, and better for regular people, or are regular people just being asked to finance the extension cord?
A big thank you to my good friend Devon for asking me to dig into the real costs of these data centers. Sometimes you have to look under the hood to see how much gas we're actually burning—or in this case, how much water we're gulping down.
Be mindful, be watchful and good luck.