The AI Hype Cycle Faces Its Biggest Test: A Shift You Can’t Ignore

Investors and builders alike are starting to question the (AI Hype Cycle) story.

AI is currently facing an inescapable problem. Over the past few years, we’ve been living through one of the biggest technological shifts humanity has ever experienced.

You know, AI has taken over the world, and companies like Nvidia, Oracle, OpenAI, Microsoft, Google, Meta, Amazon, Apple, and Tesla are investing 100 of billions of dollars each to build out data centers that will power this new technological age. And investors have been loving it, too.

Just look at the S&P 500 or the Nasdaq. Very frothy. But despite the technology being revolutionary, the growth of AI is facing a major choke point.

It’s a problem that economists have been warning about for years. In fact, this problem has quietly grown into such a size that now even the pro-business administration is stepping in to get things under control.

The problem can be seen here. This is one of the world’s largest AI training clusters, dubbed Colossus, built by Elon Musk’s xAI.

Now, this is an amazing feat of engineering, and the world has not seen such a powerful data center built in such a short a time frame, just 122 days.

But interestingly, when this data center came online, residents in the area noticed some strange changes to their surroundings.

All of a sudden, their neighborhood featured a constant hum of distant jet engines, and the area started to smell like gas.

The culprit? These mobile gas turbines, portable short-term power generators, are typically designed as backup power for hospitals or military bases or deployed for disaster response.

The interesting thing about this facility is that they didn’t just have one or two backup generators.

They had 35 turbines on site just to generate enough power to keep this facility running, chewing through gas just to keep this giant computer on.

This is the problem that the AI industry faces. It has a massive power problem.

According to Business Insider, companies in the US had filed permits to build 311 data centers nationwide as of 2010. By the end of 2024, that number had nearly quadrupled to 1,240.

Right now, warehouses like this are popping up across the United States as these tech giants invest heavily in their AI futures.

But the thing about these warehouses is that instead of holding furniture or parcels, these warehouses are housing thousands and thousands of power-hungry GPUs.

When I say power-hungry, I mean it. Each one of these data centers can consume as much power as a small city.

We’re talking power consumption of around 100,000 homes, the lights, the HVAC systems, the EV charging in the garage, everything.

And in fact, the biggest data center under development right now will eat up to 20 times that amount.

Now, imagine we’re popping one of these into existence every single week. It is the largest industrial buildout of the 21st century without question.

What’s really fascinating to me is how much it clashes with the mental image we all have of big tech.

You know, the marketing is always so sleek. It’s the cloud. It’s clean. It’s white. It’s silent. It’s weightless.

But the reality is the cloud is very heavy. It’s extremely hot, loud, and power-hungry.

And it’s becoming such a problem that just the other week, the extremely pro-business president announced he would be meeting with these tech giants to put the brakes on.

It really feels like a watershed moment where the government finally just admitted out loud, “Okay, the grid is actually at risk here.”

But before we get to politics, I want to start at the bottom of the stack. I want to talk about the hardware itself, the silicon, because none of this, the blackouts, the noise complaints, the political maneuvering, none of it makes sense unless you understand the physical object that’s driving it, the chip.

Now, computers have been around for 50-plus years. The internet has been massive for 30 years. So, why is 2026 different?

Why are we seeing an energy crisis right now that we didn’t see in, say, 2020?

Well, it comes down to a fundamental architectural shift in how we compute.

The internet basically ran on CPUs, central processing units, you know, the chips in our laptops.

They’re generalists. They’re very efficient. They can run a spreadsheet, play a movie, manage a database, and do a whole bunch of things.

But the AI boom, specifically the generative AI we’ve been living with for the last few years, no longer prefers CPUs.

It runs predominantly on GPUs, graphics processing units, originally designed for video games, designed to render millions of pixels simultaneously on a screen.

That specific ability, parallel processing, turns out to be exactly what you need to train a massive neural network. But the tradeoff is energy.

GPUs are incredibly power-hungry. How hungry?

A standard server rack, the kind that ran Google or Facebook back in 2015, pulls maybe 7 kW, right?

Think of that as running a couple of household ovens. It’s totally manageable.

You can cool it with basic air fans. But for an AI server rack, the specs we’re seeing now are 30, 50, even over 100 kW per rack.

It’s a 10-to-15-fold increase in power density. You are effectively packing the energy demand of a whole neighborhood into a cabinet the size of a vending machine.

Now, that is insane. And if you track the chip specifically, the line on the chart is only getting steeper.

Just a few years ago, the A100 chip was the industry standard. It consumed about 400 W. Then the H100 jumped to 700 W.

And now the Blackwell B200, which is filling up these new data centers right now, is pushing 1,000 W per chip. And a single server might have eight or 16 of them bolted together.

This is the tricky part to wrap your head around, because with technology, we’re so used to it getting more efficient.

You know, my fridge uses way less power than my grandmother’s fridge did.

My LED lights use way less power than the incandescent bulbs used to be. So, why are these AI chips getting thirstier?

Well, the interesting thing is that technically, they are more efficient. If you look at performance per watt, they are absolutely miracles of engineering these days.

But unfortunately, now we’ve run headfirst into Jevons paradox, which is the idea that making a resource more efficient actually increases the total consumption of it, precisely because we made the thing cheaper and more efficient.

We didn’t just say, “Great, let’s use less power to do the same stuff.” We said, “Well, now let’s build models that are a thousand times bigger.”

And now we’re integrating AI into everything from word processors to cars to medical imaging.

So, even though the car’s engine is technically more efficient, we’ve just decided to drive it at a billion miles per hour instead of 10.

And driving it that hard takes a heck of a lot of power, and it creates an unbelievable amount of heat.

This is the other part of the hardware hunger that we need to touch on. It’s not just electricity. It’s water, too.

Once you move towards 50, 80, or even 100 kW per rack, traditional air cooling becomes extremely inefficient.

That’s why hyperscalers are rapidly shifting towards liquid cooling and large-scale evaporative systems.

Even in 2022, which is basically ancient history in AI terms, Google and Microsoft together reported using roughly 25 to 30 billion liters of water primarily for cooling their data centers.

But with AI rack densities climbing and new gigawatt-scale campuses coming online, that demand is skyrocketing.

And a lot of these data centers today are being built in the desert. Take Arizona, for example.

Here, the data centers are now competing with local farmers and residential subdivisions for the water table, and it’s putting stress on the water supply of these local communities.

It’s not just power, but water, as well. And it doesn’t stop there, because these data centers are also indirectly causing sound and air pollution amongst residential communities across the US.

That was the main problem with Colossus, the xAI facility in Memphis. The investigative reporting coming out of there paints a very different picture than the sleek AI commercials we’re seeing.

But why? Well, fundamentally, it’s because usually a data center of that scale takes three to four years to plan, permit, and build.

But to move as fast as Elon did, and a lot of these hyperscalers want to, you have to bypass the normal infrastructure, specifically the power lines.

The local utility in Memphis couldn’t just magically conjure 150 MW of electric power in three months. That is just not how grids work.

The result is you get these 35 gas turbines, and along with it, neighbors complaining of nausea, worsening asthma, a smell of gas, and a constant hum.

Now, to be fair to the timeline, xAI has since hooked up to the main grid and removed some of those turbines.

But the story doesn’t end there. They are expanding just across the border to South Haven, Mississippi, and it’s a similar story.

They’ve installed 27 temporary turbines there, which are vibrating residents’ walls and stopping people from getting to sleep.

This really highlights the disconnect. You have the mayor of South Haven calling this an amazing investment in the area because he sees tax revenue and he sees jobs.

Then the residents just see, or rather hear, a nightmare. It’s the classic battle between economic development and quality of life, but supercharged by the sheer scale of energy that these things need.

That is exactly why this issue has now landed in the Oval Office. February 25th, 2026, President Trump summons the heads of the tech giants, Amazon, Google, Meta, Microsoft, Oracle, and xAI.

It’s really a who’s who of power consumption. And he announces the ratepayer protection pledge.

What is this? In essence, it’s the White House telling Big Tech, “Look, if you want to build these massive AI factories, you can’t just plug into the grid and drive up the price of electricity for everybody else. You need to build, bring, or buy your own power generation.”

So, if Amazon wants to build a gigawatt data center, it needs to build a gigawatt power plant.

Americans are also concerned that energy demand from AI data centers could unfairly drive up their electric utility bills.

Tonight, I’m pleased to announce that I have negotiated the new ratepayer protection pledge. You know what that is?

We’re telling the major tech companies that they have the obligation to provide for their own power needs.

They can build their own power plants as part of their factory so that no one’s prices will go up, and in many cases, prices of electricity will go down for the community, and very substantially down.

Now, that certainly sounds like a step in the right direction, and hopefully it does help boost power supply, but the plan has already drawn fevered criticism.

The main point is that building a power plant does not necessarily solve the problem of moving that power to the data center.

You know, the electrical grid is an old and complex network of wires and transformers, and it’s already extremely congested.

And even if a company builds its own gas plant nearby, they still have to use public transmission lines to deliver that electricity.

Now, the problem with that is that it will not only take significant time and significant money to upgrade that grid infrastructure, but the cost of maintaining those poles and wires is traditionally shared, and these expenses will just end up on the bills of the residential ratepayer anyway. So, are residents really escaping increased costs?

That’s really the first thing. Then the other criticism is the environmental cost, because the administration isn’t just asking the tech companies to build the power generation; they’re also promising that the rules are not going to get in the way.

This is what they’re calling the energy dominance strategy, and as a part of that, we are seeing a push from the administration and EPA head Lee Zeldin to slash EPA budgets and roll back standards on particulate matter.

Trump’s EPA has begun what it calls the biggest deregulatory action in US history.

To fulfill President Trump’s promise to unleash American energy and give power back to the states.

It eliminated the EPA’s office that assists heavily polluted communities like South Memphis and canceled billions of dollars of funding for addressing pollution.

The argument the administration is making is purely geopolitical: to beat China.

They argue we are in an AI arms race with China, and we simply cannot afford to let environmental reviews slow us down.

And I’m not taking a stance on the politics of deregulation, but we have to look at the practical effect.

The strategy is essentially to build whatever you need, whether it’s coal or gas or nuclear, and the government will make sure the red tape doesn’t stop you. That’s the explicit goal. It’s a move-fast mentality applied to national energy policy.

That begs the question, what are they looking to build?

Well, it really comes down to timelines. The faster you need your data center online, the more likely you are to use fossil fuels as the energy source.

That’s what we saw in the case of xAI, but Silicon Valley is also obsessed with nuclear.

Microsoft recently penned a deal to restart Three Mile Island. Yes, we are talking about that Three Mile Island, but don’t worry, the reactor they’re restarting is the one that didn’t melt down.

But the symbolism is very heavy, and billions are pouring into small modular reactors, too.

I mean, they’re carbon-free, they run 24/7, but there’s a catch. You can’t just order a small modular reactor on Prime, and you can’t even build one in 5 years, usually.

So, to meet the demand today, right now, we are seeing a massive resurgence of fossil fuels.

Utilities in places like Virginia, Michigan, and Pennsylvania are literally canceling their plans to retire old coal, oil, and gas plants. In some cases, they are building brand new natural gas peaker plants because the AI data centers will never sleep.

When it comes to these tech companies taking matters into their own hands and building their own power generation solutions, gas is proving to be the way to go.

In fact, CleanView found 75% of the new power generation equipment being ordered and installed by these companies is being powered by gas.

And with the agentic age of AI only just starting, that power demand is only going to rise. Here are some stats for you.

The IEA projects global electricity demand for data centers will more than double by 2030, from 415 terawatt hours in 2024 to around 945 terawatt hours in 2030.

And when you look at just America, the numbers get even more terrifying. Reports model that data centers could chew through up to 9.1% of America’s total power consumption by 2030.

Think about that. Of all the things we use power for, heating, cooling, lights, appliances, charging cars, running businesses, powering factories, 10% of all power will be going straight to data centers.

That’s tens of millions of homes, or think about it like this, that is three New Yorks worth of power just running data centers. But power just can’t come online fast enough.

Because of this insatiable demand for power, now even those gas turbine manufacturers have said lead times for their equipment stretch as long as 5 to 7 years.

While all the news and hype and stock market rallies center around AI’s explosive growth, I hope this story has shown you that, actually, there’s a much tougher problem to solve: energy, power, grid upgrades, and how you do it in a way that doesn’t actually end up hurting the average US citizen.