You Can’t Build a Power Plant Overnight: The Infrastructure Reality Nobody’s Pricing In

Let me tell you about a conversation that’s been nagging at me since last night.

I was at an event, talking with a few folks from a few different companies. All smart people, all building genuinely interesting things. The conversation turned to their scaling plans, and I heard the same thing from multiple people: they’re planning to 10x their compute capacity in the next twelve to eighteen months to hit their roadmaps.

I asked the question that seemed obvious to me: “Where’s all that compute coming from?”

You should’ve seen the looks. Like I’d asked where the sun comes from.

“We’ll just buy more GPU time. Scale up our cloud spend.”

“And where do you think that GPU time comes from?” I pressed.

Long pause. Multiple long pauses, actually. The kind where you realize a whole room of people has never actually thought about the physical reality underneath the digital promise.

Here’s what I’ve been watching while everyone’s drunk on AI possibilities: we’re running headfirst into a brick wall made of concrete, copper, and the laws of physics. And that wall’s going to do more to pop this bubble than any market correction or regulatory framework.

Because you can’t run AI without massive computing power. You can’t generate that computing power without data centers. You can’t run data centers without electricity. And you can’t conjure electricity and the infrastructure to deliver it out of thin air, no matter how much venture capital you throw at the problem.

The Math Nobody Wants to Do

Let’s talk about what it actually takes to run this AI revolution everyone’s betting on.

Training a large language model takes enormous amounts of power. GPT-3 reportedly used about 1,287 megawatt-hours of electricity. That’s the equivalent of what 120 US homes use in an entire year. For one model. One time.

Now we’re training bigger models. Running inference at scale for millions of users. Every ChatGPT query, every AI-generated image, every “intelligent” feature being crammed into every product uses computational power. Real, physical, electricity-consuming power.

Goldman Sachs estimates that data center power demand could grow 160% by 2030. The International Energy Agency projects AI could require as much electricity as the entire country of Sweden by 2026.

That’s not speculation. That’s physics meeting reality.

The Constraint Nobody’s Pricing In

Here’s what keeps me up at night: everyone’s building business models assuming infinite scalability. Every startup pitch deck shows hockey-stick growth curves. Every AI company roadmap assumes they’ll have the compute they need when they need it.

But supply isn’t infinite. And more importantly, it can’t scale at the speed everyone’s assuming.

Building a data center takes 18-36 months minimum. That’s if you’ve got land, permits, and no significant community opposition. More complex projects take longer. Much longer.

Building the power infrastructure to support that data center? That’s measured in years, sometimes decades. You need power generation capacity. You need transmission lines. You need grid upgrades. You need regulatory approvals from multiple agencies at multiple levels of government.

I was talking to someone in the data center business last month. He told me they’re seeing demand that would require tripling US data center capacity in the next five years. When I asked if that’s possible, he just laughed. “We can’t get permits that fast. We can’t build transmission that fast. We can’t even get the transformers that fast because everyone else in the world is trying to build the same infrastructure.”

This isn’t a problem money solves. This is a problem time solves, and time doesn’t compress just because venture capitalists will it to.

The Power Problem

Let me paint you the picture everyone’s ignoring:

Data centers already consume about 1-2% of global electricity. That percentage is growing fast. In some regions, data centers are already consuming 10-15% of total power generation.

Every major tech company is announcing massive AI initiatives. Every startup’s building AI features. Every enterprise is deploying AI tools. They’re all competing for the same limited resource: electricity to power compute.

And here’s the thing about electricity infrastructure: it’s old, it’s stressed, and it can’t scale quickly.

The US power grid is aging. We haven’t built significant new generation capacity in years. We’re trying to transition to renewable energy while demand is spiking. We’ve got regions that already experience brownouts in summer because air conditioning maxes out capacity. Now we want to add data centers consuming as much power as small cities?

The math doesn’t work.

Nuclear power could theoretically help, but building nuclear plants takes a decade or more, costs billions, and faces massive regulatory and political hurdles. Solar and wind are great but intermittent, and battery storage at the scale needed doesn’t exist yet. Natural gas plants can be built faster, but that’s not going to fly in a world trying to decarbonize.

Every path forward takes time we don’t have if AI is supposed to scale the way everyone’s promising.

The Geography Problem

Here’s another constraint nobody talks about: you can’t just build data centers anywhere.

You need reliable power. You need network connectivity. You need cooling, which means you need water or climate that supports air cooling. You need to be somewhat near your users to minimize latency. You need friendly regulatory environments.

The places that check all those boxes? They’re already building data centers as fast as they can. And they’re running into limits.

Northern Virginia – the largest data center market in the world – is hitting power constraints. Dominion Energy told data center developers they can’t guarantee power for new facilities in certain areas. Ireland stopped approving new data centers in Dublin because the grid can’t handle more load.

These aren’t temporary problems. These are “we need to build multiple new power plants and upgrade the entire regional grid” problems. That takes years and billions in investment.

Companies are looking at more remote locations with cheaper power. But then you’re adding latency and connectivity challenges. And those remote locations still need massive infrastructure buildout.

The Supply Chain Reality

Even if you’ve got power and permits and land, you still need the physical equipment to build data centers.

Transformers. Chillers. Backup generators. Server racks. Network equipment. All of it’s seeing unprecedented demand. Lead times are stretching. Prices are climbing. And manufacturing capacity for specialized data center equipment doesn’t scale overnight either.

I’ve heard stories about companies paying premiums to jump the line on transformer deliveries. About data centers delaying openings by months because they can’t get the equipment they need. About projects that penciled out financially at one price point becoming marginal or unprofitable as costs climb.

This is the unsexy reality underneath the sexy AI promises: you need physical stuff, built by real people, using materials that have to be manufactured and transported. And all of that takes time and money in ways that don’t compress just because the market’s hot.

What This Means for the Bubble

Here’s why this matters for the AI bubble we’re living through:

Every company building on the assumption of infinite compute is building on sand. When they hit scaling limits, when they can’t get the GPU time they need at prices they planned for, their growth curves break. Their unit economics fall apart. Their promises to investors become impossible to keep.

Right now, cloud providers are the buffer. Companies buy compute from AWS or Azure or Google Cloud, and those providers worry about the underlying infrastructure. But the cloud providers are hitting constraints too. They’re competing for data center capacity, for power, for equipment. When they can’t scale fast enough to meet demand, prices go up or availability goes down.

Either way, all those AI startups with thin margins built on assumptions of cheap, infinite compute? They’re going to discover reality hurts.

I’m already seeing early signs. Projects getting delayed because GPU availability isn’t what was promised. Costs running higher than projected because compute pricing is climbing. Companies discovering that the economics that worked on paper don’t work when you’re competing with everyone else for limited resources.

The Timeline Problem

Here’s the thing that makes this different from other scaling challenges: the mismatch in timelines.

AI hype is moving at internet speed. Companies are promising revolutionary products in months. Investors are funding growth plans measured in quarters. Markets are expecting returns in years.

Infrastructure moves at construction speed. Projects are planned in years. Built over more years. Come online after even more years.

You can’t reconcile those timelines. You can’t build five years of power infrastructure in eighteen months. You can’t wish data centers into existence.

The companies that are being realistic about this? They’re planning slower growth, they’re being more selective about what they build, they’re managing expectations with investors about what’s actually achievable given infrastructure constraints.

The companies that aren’t being realistic? They’re going to hit walls. Hard.

The Winners and Losers

When this constraint becomes undeniable – and it will – here’s what I think happens:

The companies with long-term infrastructure deals, with their own data centers, with locked-in power agreements, they’ll have massive competitive advantages. The companies fighting for scraps of available compute will struggle.

The hyperscalers with deep pockets to build infrastructure will get even more dominant. The startups assuming they can just scale on cloud will discover moats they can’t cross.

The AI applications that are compute-efficient will survive. The ones burning compute recklessly because resources seemed infinite will die when resources get expensive or scarce.

The market will have to recalibrate around what’s actually possible given real-world constraints, not what pitch decks promised was possible in a fantasy world of unlimited resources.

Why Nobody Wants to Hear This

I know this isn’t the sexy take. It’s not the “AI is going to change everything” excitement that gets clicks and raises rounds and pumps stocks.

But here’s the cowboy code truth: reality always wins. Physics always wins. You can ignore constraints in your planning, but they don’t ignore you.

I’ve seen this pattern before. In the dot-com boom, everyone assumed bandwidth would scale infinitely to support streaming everything. It didn’t, not fast enough. In the crypto boom, everyone assumed energy costs didn’t matter. They did.

Now everyone’s assuming compute will scale infinitely to support every AI dream. It won’t.

Not because AI isn’t real or valuable. But because the physical infrastructure to support infinite AI growth doesn’t exist and can’t be built fast enough to support the growth curves everyone’s betting on.

What Smart People Are Doing

The folks I respect in this space? They’re planning for constraint, not abundance.

They’re building compute-efficient models instead of just throwing parameters at problems. They’re being selective about what actually needs AI versus what works fine with traditional approaches. They’re stress-testing their unit economics against higher compute costs and lower availability.

They’re talking to data center operators about real timelines. They’re understanding power markets. They’re building relationships with infrastructure providers instead of assuming resources will just be there when needed.

They’re not betting their entire strategy on being able to scale compute infinitely at dropping prices. They’re building businesses that work in a world where compute is the constraint it actually is.

The Honest Conversation

Look, I’m not saying AI isn’t transformative. I’m not saying it won’t change how we work and live. I’m saying the path from here to there runs through real-world constraints that don’t care about your pitch deck.

Data centers take years to build. Power infrastructure takes longer. Supply chains are stretched. Geography matters. Physics matters. And all the capital in the world doesn’t change those realities.

The AI bubble will pop partially because of market dynamics, partially because of overhype, but fundamentally because reality catches up with promises. And one of the biggest realities is this: you can’t scale AI without scaling physical infrastructure, and that infrastructure can’t scale at the speed the market’s pricing in.

Companies are going to hit these walls. Growth will slow. Costs will climb. Availability will constrain. And the market’s going to have to reprice everything based on what’s actually possible, not what seemed possible when resources felt infinite.

The Bottom Line

You can’t build a power plant overnight. You can’t will a data center into existence. You can’t ignore physics because you’ve got venture funding.

The infrastructure to support everyone’s AI dreams doesn’t exist yet and won’t exist soon enough to support the growth curves everyone’s betting on. That’s not pessimism. That’s just math and construction timelines.

The question isn’t whether AI is real or valuable. The question is whether the industry can scale at the rate everyone’s promising given real-world infrastructure constraints. And the honest answer, the one nobody wants to hear, is: probably not.

That gap between promise and reality? That’s one of the forces that’s going to pop this bubble. Not the only one, but a big one. Because eventually, you run out of compute capacity to buy. You run out of power to run it. You run out of ways to pretend the physical world doesn’t matter.

And when that day comes, all the pitch decks and market enthusiasm in the world won’t change the fact that you needed to start building infrastructure years ago to support what you’re promising today.

That’s the truth, straight and unvarnished. The kind of truth cowboys understand: you can’t build a ranch overnight, no matter how many cattle you plan to run. The land, the water, the buildings, the fences – they all take time. And time doesn’t compress just because you’re in a hurry.

The AI revolution will happen. Just slower and messier than everyone’s banking on. Because reality has constraints that don’t care about your timeline.

That’s what I’m seeing from where I sit. Make of it what you will.