The AI Data Center Boom Is Starting to Break
Big Tech is spending billions on AI infrastructure, but shaky returns are creating a reality check.
The AI boom was supposed to be a software revolution.
Chatbots, coding agents, image generators, enterprise copilots, automated search, synthetic video, AI assistants everywhere. The story sounded clean, weightless, and inevitable.
But behind every AI answer is not magic. It is a building. Usually a very large building. Inside it are GPUs, cooling systems, diesel backup generators, transformers, switchgear, water pipes, concrete, steel, and a terrifying amount of electricity demand.
The cloud was never really a cloud.
It was always concrete.
AI just made that impossible to ignore.
According to the International Energy Agency, global data center electricity use is projected to roughly double from about 485 TWh in 2025 to around 950 TWh by 2030. AI-focused data centers are expected to grow even faster than the sector overall.
That is the real story now.
AI is no longer only competing on models.
It is competing on power.
And power is where the fantasy starts hitting the wall.
Big Tech Is Spending Like This Is a New Industrial Revolution
Amazon, Microsoft, Meta, and Alphabet are committing hundreds of billions of dollars to AI infrastructure. The number often thrown around for 2026 alone is roughly $650 billion in combined capital spending.
That is not normal software money.
That is railway money. Space race money. Nation-building money.
To understand how strange the scale has become, compare it with older mega projects. The Marshall Plan rebuilt Europe after World War II. The Manhattan Project built the atomic bomb. Apollo put humans on the Moon. The International Space Station became one of the most expensive science projects ever built.
The AI data center boom is now being discussed at a scale larger than all of those combined.
That is what makes this moment so surreal. We are not just building server rooms anymore. We are funding a new industrial layer of the internet before anyone has fully proved the business model can pay it back.
The logic is simple. If AI becomes the new operating layer of the internet, whoever controls the compute controls the future. More compute means better models, faster products, more enterprise deals, and a bigger moat.
So everyone is rushing to build.
Private equity wants in. Utilities want in. Chipmakers want in. Real estate developers want in. Governments want the investment. Wall Street wants the growth story.
The pitch sounds clean:
Promise | Reality |
|---|---|
AI demand will keep exploding | Demand is real, but the returns are still uncertain |
Data centers will bring jobs | Construction jobs are real, but permanent jobs are often limited |
Big Tech will pay for the buildout | Communities may still face higher bills, grid stress, and local disruption |
Clean energy will solve the problem | New demand is arriving faster than clean supply and transmission |
The cloud can scale instantly | Land, power, permits, transformers, and substations cannot |
This is why the current data center boom feels so strange.
The money is real. The demand is real. The projects are real.
But so are the delays.
A Lot of the Boom Still Exists on Paper
Announcing a data center is easy.
Building one is not.
A company can announce a multi-billion-dollar AI campus, show a render, talk about gigawatts, and dominate headlines. But before that campus can actually run, it needs grid connection, power agreements, water planning, backup generation, local approval, equipment, labor, and years of construction.
That is where the slowdown begins.
A Bloomberg-linked report claimed roughly half of planned U.S. data center builds for 2026 had been delayed or cancelled because of power infrastructure shortages and equipment constraints. The exact numbers vary by report, but the direction is clear: the AI buildout is already running into the physical limits of the grid.
The AI industry is discovering something old industries already knew:
Infrastructure does not move at software speed.
A model can be updated overnight. A data center campus cannot.
A product can go viral in a week. A transmission line can take years.
That mismatch is now one of the biggest risks in AI.
The Real Bottleneck Is Electricity
For most of the AI boom, people talked about GPUs as the shortage. Nvidia chips became the symbol of the entire race.
But the bigger shortage is now electricity.
The U.S. Department of Energy says data centers consumed about 4.4% of total U.S. electricity in 2023. By 2028, that could rise to between 6.7% and 12%.
That is a massive jump for one sector.
Modern AI servers are far more power-dense than older cloud workloads. Training large models requires huge clusters of accelerators. Running those models for millions of users creates constant inference demand. Video generation, agents, search answers, coding tools, enterprise automation, and AI assistants all add more load.
This is why data center projects are now measured in megawatts and gigawatts.
Some proposed campuses sound less like tech buildings and more like industrial energy projects.
The problem is that the grid was not built for this kind of sudden concentrated demand. Power plants take time. Transmission takes time. Substations take time. Permits take time. Transformers are already in short supply. Skilled electrical labor is limited.
The AI boom is asking the power grid to behave like the internet.
It cannot.
The Supply Chain Is Also Cracking
The electricity problem is not just about generating power.
Data centers need transformers, switchgear, backup batteries, generators, cooling systems, fiber, electrical equipment, construction crews, engineers, and technicians.
If one critical piece is delayed, the whole project can slip.
That creates a weird situation where the most futuristic technology in the world can be blocked by old-school industrial shortages.
You can have the GPUs.
You can have the land.
You can have the customer.
You can have the financing.
But if the transformer is not available, nothing turns on.
The AI economy is not only built on chips. It is built on everything required to power and cool those chips.
And that supply chain is now under pressure.
Even Microsoft and Stargate Are Hitting Reality
The delays are not only happening to random developers.
Even the biggest names are running into limits.
In 2025, TD Cowen analysts said Microsoft had walked away from data center projects in the U.S. and Europe totaling up to 2 GW of capacity, with reports framing the pullback as a sign of oversupply relative to near-term demand forecasts. Microsoft disputed the broader panic around its AI plans, but the report still mattered because it showed the market was starting to question whether every planned AI data center would be needed immediately.
The same pressure has touched Stargate, the giant OpenAI, Oracle, and SoftBank AI infrastructure project. In March 2026, Reuters reported that Oracle and OpenAI had shelved plans to expand a major AI data center in Abilene, Texas, because of financing talks and changes to OpenAI’s requirements.
The companies were still moving ahead with capacity elsewhere.
But that is exactly the point.
Even the biggest AI infrastructure project in America is not immune to financing, capacity, and execution problems.
Everyone says AI demand is infinite.
But the builders are already making adjustments.
The Kevin O’Leary Data Center Became a Warning Sign
One of the clearest examples of the backlash came in Utah.
Kevin O’Leary’s proposed Stratos Project in Box Elder County was originally planned as a massive 40,000-acre AI and defense data center campus. The scale triggered immediate alarm over water use, energy demand, environmental damage, heat, and the impact on the nearby Locomotive Springs Waterfowl Management Area.
After political and public pressure, O’Leary agreed to shrink the project. The Verge reported that nearly 19,430 acres would be removed from the proposal, especially around sensitive environmental areas. Business Insider reported the project could still require 7.5 to 9 GW of power, even after the changes.
That number is insane.
This is not a small server farm hidden behind an office park. This is industrial infrastructure at the scale of a regional power system.
The Kevin O’Leary story matters because it shows the new politics of AI infrastructure.
A billionaire can announce a giant data center.
But the town, the state, the water, the birds, the grid, and the people still get a vote.
Communities Are Starting to Hate These Projects
For years, most people did not think much about data centers. They sat in industrial zones and quietly ran the internet.
AI changed that because the scale changed.
The new facilities are bigger, louder, more power-hungry, and harder to ignore. Residents are pushing back over noise, water use, diesel generators, land use, tax breaks, and rising electricity bills.
That backlash is no longer fringe. Heatmap reported that 25 U.S. data center projects were cancelled in 2025 after local pushback, four times the number in 2024.
Polling also shows the mood is getting worse. A Quinnipiac poll found 65% of Americans opposed an AI data center being built in their community. A later Gallup poll found seven in ten Americans opposed local AI data center construction.
This is becoming a kitchen-table issue.
Not because people suddenly care about server architecture.
Because they care about their bills, their water, their sleep, their town, and whether they were ever asked.
A trillion-dollar company gets compute capacity.
A town gets the hum.
The Tax Breaks Make People Even Angrier
The anger gets worse when tax incentives enter the story.
Many data center projects are sold with promises of economic development. But some deals also come with property tax abatements, sales tax exemptions, infrastructure support, or other incentives.
That creates a political problem.
If Big Tech receives tax breaks while ordinary residents face rising utility bills, the project starts to look less like development and more like a subsidy.
Virginia’s “data center alley” in Loudoun County became the symbol of this bargain. The region attracted massive data center investment, but also growing complaints about land use, power demand, transmission lines, diesel backup generators, and the long-term shape of the local economy.
Oregon has had a similar debate. Lawmakers pushed the POWER Act to make sure residential customers are not forced to subsidize grid upgrades needed by large data centers.
That is the political shift.
The question is no longer just “will this data center bring investment?”
The question is “who pays for the infrastructure it needs?”
Water Is Becoming the Emotional Flashpoint
Power is the main bottleneck.
Water is the emotional one.
Data centers generate heat, and that heat has to go somewhere. Some facilities use air cooling. Some use liquid cooling. Some use evaporative cooling, which can consume large amounts of water depending on the location and design.
AI makes this harder because GPU racks are denser and hotter than older server setups.
A Reuters report on UN research warned that data centers could double both power and water consumption by 2030 as AI demand grows. A separate Guardian analysis found that many planned U.S. AI data centers are being built in drought-hit areas.
Australia shows how global this issue is becoming. ABC Australia reported that data centers could demand more new electricity over the next 15 years than Australia’s cars or homes, with demand potentially rising from 2% to 13% of the country’s total energy use by 2040. AAP reporting also cited Sydney Water warnings that data centers could use up to 25% of the city’s drinkable water by 2035.
That is where the politics becomes explosive.
People may accept water and electricity use for hospitals, banking, logistics, emergency systems, and essential cloud services. They are less patient when the same resources are used to generate low-value content, spam, ads, memes, or AI slop.
That is the moral problem of AI infrastructure.
Not all compute feels equally worth it.
The Jobs Argument Is Getting Weaker
Data centers are often sold to communities as job creators.
That is partly true. Construction jobs can be significant, and large projects can bring short-term economic activity. But once the facility is complete, the permanent workforce is usually much smaller than the size of the building suggests.
If a facility consumes enough electricity for hundreds of thousands of homes but only creates a few hundred permanent jobs, residents will question the trade.
Especially if the company also receives tax breaks.
The Memory Shortage Shows the Boom Hits Consumers Too
The data center boom is not only affecting people who live near data centers.
It is also showing up in consumer hardware.
AI servers consume huge amounts of memory, especially high-bandwidth memory for accelerators and enterprise DRAM for servers. As manufacturers shift capacity toward AI hardware, normal consumer markets can feel the squeeze.
IDC warned that the global memory shortage is being shaped by a strategic reallocation of wafer capacity toward AI and data center demand. Counterpoint Research reported a sharp rise in memory pricing, including 64GB RDIMM prices jumping from $255 in Q3 2025 to $450 in Q4 2025, with targets reaching $700 by March 2026.
That is the consumer version of the AI infrastructure boom.
You may not live near a data center.
You may not buy cloud compute.
But if AI demand absorbs memory supply, your next PC, laptop, server, or upgrade can still get more expensive.
The boom leaks into electricity bills, water politics, construction markets, and even RAM prices.
The Allbirds Pivot Was Peak AI Mania
Nothing captures the mood better than Allbirds.
Allbirds was known as a struggling sustainable shoe company. Then in 2026, it announced a dramatic pivot away from consumer footwear and toward AI infrastructure, rebranding around GPU-as-a-Service and AI cloud solutions.
The market loved it. Tom’s Hardware reported that the stock surged by around 580% in a single day after the pivot.
That is funny.
It is also a warning sign.
When a shoe company can say “AI data centers” and suddenly become a market darling, you are no longer just looking at infrastructure demand. You are looking at narrative heat.
The dot-com bubble had companies adding “.com” to their names.
The AI bubble has companies adding “GPU cloud.”
The Bubble Question Is Getting Harder to Ignore
The most uncomfortable part of the data center boom is not whether AI is useful.
AI is useful.
The real question is whether the current level of infrastructure spending can generate enough return fast enough.
Hyperscalers are spending enormous amounts of money. AI labs are raising huge sums. Infrastructure firms are building for future demand. Debt markets are financing new capacity. Everyone is betting that AI usage will keep rising and that customers will pay enough to justify the buildout.
But the revenue side is still uncertain.
Consumers like AI tools, but many do not want to pay high monthly prices. Enterprises are experimenting heavily, but not every pilot becomes a profitable deployment. Open-source models are improving. Smaller models are becoming more capable. Local AI is getting better. Inference costs may fall. Some tasks may move from giant cloud systems to laptops, phones, and edge devices.
That does not kill the data center business.
But it does challenge the assumption that every planned gigawatt of AI compute will be needed at profitable prices.
This is not a simple dot-com repeat. Data centers are real assets. Cloud demand is real. AI demand is real. Power constraints are real because usage is real.
But parts of the market now look overheated.
Some companies are spending ahead of proven returns. Some projects are being announced before power plans are clear. Some investors are treating anything attached to AI infrastructure as if it is guaranteed to win.
That is not a stable situation.
Fermi America Is the Perfect Reality-Check Story
Then there is Fermi America.
Fermi, co-founded by former U.S. Energy Secretary Rick Perry, pitched one of the most ambitious AI power and data center projects in the world: Project Matador, a massive Texas Panhandle energy and intelligence campus intended to deliver gigawatt-scale power for AI.
On paper, it sounded like exactly what the AI boom needed: huge land, huge power, huge ambition, and a story built for investors.
Then reality hit.
Fermi’s own site now pitches an 11 GW private HyperGrid campus on 7,570 acres near the Pantex DOE complex. But the company has faced major questions around financing, tenants, leadership, and execution. Barron’s reported that CEO Toby Neugebauer and CFO Miles Everson stepped down as Fermi struggled to secure a cornerstone tenant for Project Matador. The Houston Chronicle reported that the company had not begun construction, had suffered a major stock drop, and was facing legal and financial pressure.
That is why Fermi matters.
It is not just one troubled company.
It is a symbol of the entire AI infrastructure moment.
The story sounds enormous. The presentation looks futuristic. The numbers are huge. But at the end of the day, someone still needs to sign a lease, finance the project, build the power, secure the equipment, and turn the thing on.
AI hype cannot replace an anchor tenant.
The Political Fight Is Just Beginning
The backlash is now turning into policy.
New York lawmakers have moved toward a one-year pause on large AI data centers while the state studies energy, water, labor, and ratepayer impacts. The Guardian reported that New York could become the first U.S. state to temporarily ban large data center construction if the bill is signed.
Cities are moving too. Monterey Park, California, became a major example after voters overwhelmingly backed a permanent data center ban following opposition to a proposed project near residential areas. The Guardian reported that early results showed more than 86% support for the ban.
That is not a small zoning dispute.
That is a revolt.
The data center industry can still win communities over, but the old playbook is weakening. Tax breaks, job promises, and “innovation” language are not enough anymore.
Residents want details.
How much power?
How much water?
How much noise?
Who pays for the grid?
Who gets the tax break?
Who is responsible if the project fails?
The AI Boom Has Entered Its Reality-Check Phase
The first phase of AI was hype.
The second phase was adoption.
The third phase is infrastructure.
And infrastructure is where the clean story gets messy.
AI companies cannot assume infinite power. They cannot assume infinite water. They cannot assume infinite grid capacity. They cannot assume infinite local patience. They cannot assume every community will accept a massive industrial facility because someone says it is important for the future.
The real world now gets a vote.
That does not mean the boom is over.
It means the easy part is over.
The companies that win the next phase of AI may not just be the ones with the best models. They may be the ones with the best power strategy, the best cooling, the best community agreements, the best supply chain, and the most realistic economics.
In other words, the next AI moat may be a grid connection.
The Future May Be Smaller, Cleaner, and More Distributed
There are better ways to build this.
Some data centers are moving toward closed-loop liquid cooling, waste-heat reuse, geothermal energy, nuclear power, renewables, batteries, and smarter grid integration. China has also tested underwater data centers that use naturally cold seawater for cooling.
Local AI may also reduce pressure on centralized infrastructure. If phones, laptops, and edge devices can run smaller models that are good enough for everyday tasks, not every request needs to travel to a massive AI facility.
That does not mean giant data centers disappear.
Frontier models, enterprise AI, scientific computing, video generation, and global cloud services will still need enormous centralized compute.
But the future may not be one giant cloud doing everything.
It may be split between hyperscale data centers for the hardest workloads, regional facilities for business systems, and local devices for simple AI tasks.
The Cloud Has Become Concrete
The biggest lesson from the current data center situation is simple:
AI is not weightless.
Every prompt has a physical shadow. Every model has an energy footprint. Every chatbot depends on land, water, chips, cooling, and electricity.
The digital future is being built out of very physical materials.
That does not mean AI is doomed. It means the fantasy version of AI is over.
The question is no longer just whether AI can get smarter.
It is whether the world can afford to build the infrastructure AI companies are promising.
Because right now, the AI industry is trying to scale intelligence faster than the grid, faster than water systems, faster than supply chains, faster than local politics, and maybe faster than the business model itself.
The future may still be powered by AI.
But first, someone has to explain who pays for the power.
