Yves here. This post does a good, layperson-friendly job of describing how the tech-overlord-envisaged explosion in data centers is even more problematic than you might have realized. It adds in a new impediment to the buildout, which is transformer scarcity.
By Michael Kern, newswriter and editor at Safehaven.com and Oilprice.com. Originally published at OilPrice
AI’s shift from CPUs to ultra–power-dense GPUs has created a structural surge in electricity demand that is outpacing efficiency gains and overwhelming the grid.
The need for 24/7 baseload power is delaying coal retirements, boosting natural gas, and shifting data center growth to regions with cheap land, weak regulation, and vulnerable communities.
Global transformer shortages, mineral constraints, and interconnection bottlenecks mean that compute—not oil—may become the next strategic resource nations hoard.
The Cloud” might be the greatest branding trick in history. It sounds fluffy, ethereal, and notably light.
It implies that our digital lives…our emails, our crypto wallets, our endless scrolling…exist in some vaporous layer of the atmosphere, detached from earthly constraints.
But if you actually drive out to Loudoun County, Virginia, or stare at the arid plains of Altoona, Iowa, you realize the Cloud is actually just a very big, very loud, and very hot factory.
We’ve been telling ourselves a lovely story about the energy transition. We were retiring coal plants, building wind farms, and decoupling economic growth from carbon emissions. It was all going according to plan.
For years, the tech sector achieved relative decoupling…
Moore’s Law kept server efficiency gains ahead of the curve, allowing internet traffic to surge while power demand grew slowly.
The exponential curve of AI, however, has shattered this delicate balance. AI workloads are so compute-intensive that demand is now skyrocketing faster than efficiency gains can compensate.
This is a re-coupling with physics..and the defining narrative of the next decade isn’t about supply anymore.
Now it’s about a structural shift in demand that almost nobody priced in: The thermodynamics of Artificial Intelligence.
According to the International Energy Agency (IEA), global electricity demand from data centers is projected to more than double by 2030. This is the same as the entire annual electricity use of a country like Japan.
The invisible hand is hitting a concrete wall.
The question is no longer if the grid can handle it, but what is making the demand curve look like a rocket launch. The answer isn’t better software or smarter algorithms; it’s the raw physics happening inside a rack that now demands the power of a city block.
The Thermodynamics of “Thinking”
To understand why the grid is struggling right now, you have to look at the silicon.
For a long time, we ran the internet on CPUs (Central Processing Units). These are the general managers of the chip world. Efficient, predictable.
But Generative AI doesn’t want a manager. It wants a battalion of mathematicians. It runs on GPUs (Graphics Processing Units), specifically monsters like Nvidia’s H100.
Here’s what that actually means for power draw:
Traditional server rack: Draws about 5 to 10 kilowatts (kW).
Modern AI rack (H100s/Blackwell): Draws 50 to 100 kW.
We have effectively moved from powering a toaster to powering a neighborhood, all inside the same metal box. Air cooling…fans blowing over hot metal…doesn’t work anymore. Air just isn’t physically dense enough to move that much heat away.
We are now plumbing data centers like chemical refineries, running liquid coolant loops directly to the silicon die.
This is the new reality of Direct-to-Chip (DTC) cooling. It is already happening in cutting-edge AI centers because it is the only way to manage the extreme heat density of chips like the H100.
Liquid cooling saves energy compared to air conditioning. While the chip itself still draws 100 kW, the overall cooling system…the pumps and chillers…consumes far less power than running massive air handlers for the whole room. This makes it an efficiency measure born of necessity.
The next step is Immersion Cooling, where entire server racks are submerged in a non-conductive fluid. This is also being deployed now, often in pilot programs and specialized facilities.
This shift from fans to specialized plumbing and chemically inert fluids is the physical realization of the industrialization of thought.
Just like the industrialization of textiles or steel, it requires massive inputs of raw power and exotic, specialty materials. This industrial intensity demands something traditional renewable sources…intermittent solar and wind…struggle to provide: reliability.
When an AI training run costs tens of millions of dollars, a 1% flicker is an existential threat.
The Dirty Secret of the “Green” AI Boom
Every major tech CEO is currently on a podcast tour talking about their “Net Zero” 2030 goals. And sure, they are buying a lot of paper credits.
But physics doesn’t care about carbon offsets. The reality is that AI needs baseload power. It needs to run 24/7/365 with “five nines” (99.999%) of reliability.
You know what provides that?
According to IEA data, coal still accounts for about 30% of global data center power. And in the U.S., natural gas is doing the heavy lifting, covering over 40% of demand.
The irony is palpable. We spent billions trying to kill coal, only to have the most futuristic technology on earth, AI, throw it a lifeline.
In places like Virginia or Kansas, utilities are delaying the retirement of coal plants. They simply cannot risk the grid instability when a gigawatt-scale data center comes online.
The “future” is being powered by the “past.”
The need for this reliable baseload power, combined with the sheer gigawatt-scale hunger of these new facilities, is now fundamentally reshaping the American power landscape. Capital always flows to the path of least resistance—and right now, that path runs right through communities that have never seen a single dollar of tech prosperity.
The New Geography of Power (and Inequality)
This energy hunger is redrawing the map. We are seeing a “K-shaped” geography of infrastructure.
In the U.S., “Data Center Alley” in Northern Virginia supposedly handles 70% of the world’s internet traffic. But the grid there is tapped out. You can’t get a new hookup for years.
So, the capital is fleeing to places with looser regulations and cheaper land: Texas, Ohio, Arizona.
But this brings us to the friction point. These facilities are neighbors. And they are often bad neighbors. They are loud, they consume massive amounts of water for cooling, and they raise local utility rates.
There is also a significant Environmental Justice component here. Industrial infrastructure is rarely sited in wealthy neighborhoods.
According to the NAACP’s “Fumes Across the Fence-Line” report:
African Americans are 75% more likely than white Americans to live in “fence-line” communities (areas adjacent to industrial facilities).
A disproportionate number of fossil-fuel peaker plants, which fire up when data centers max out the grid, are located in low-income areas and communities of color.
This directly contributes to higher rates of asthma and respiratory issues.
While the “invisible prosperity” of AI stock gains flows to portfolios in San Francisco and New York, the “visible decay”…the pollution, the water usage, the hum of the cooling fans…is localized in communities that often see none of the upside.
Even if a community were willing to bear the cost, the industrial machine that once smoothly supplied the electrical grid is choked.
The problem is no longer just where to put the data center, but how to physically connect the massive, power-hungry factory to the existing grid infrastructure. This process is crippled by a global bottleneck of essential, non-digital hardware.
The Great Transformer Shortage
Let’s say you have the money, the land, and the permits. You still have a problem. You can’t get the gear.
The lead time for a high-voltage power transformer used to be 12 months. Today? It’s 3 to 5 years.
We are trying to rebuild the electrical grid at the exact moment everyone else is trying to electrify cars and heat pumps. The supply chain is fractured.
We are also running out of the raw stuff: Copper. Lithium. Neodymium for the magnets in the cooling fans.
We are dependent on China for the processing of nearly all these critical minerals. As I explained in this “Data Center Guide,” we are realizing that the digital economy is actually a material economy.
If China restricts graphite or gallium exports (which they have started doing), the Cloud stops growing.
The “Trust Me, Bro” Efficiency Pitch
The counter-argument from Silicon Valley is the “Handprint” theory. The pitch goes like this: Yes, training the AI uses a lot of energy, but the AI will make the rest of the world so efficient that it pays for itself.
The IEA models suggest that AI could optimize logistics, manage smart grids, and reduce building energy usage by 10-20%.
And honestly? It’s a compelling argument. If AI can figure out how to drive a truck platoon 5% more efficiently, that saves more carbon than the data center emits.
But this is a long-term bet against a short-term, guaranteed withdrawal of power.
The core efficiency problem is two-fold:
Training vs. Inference: Training a colossal model takes a massive, months-long burst of power. The resulting AI is then put to work performing inference…answering questions. While inference is far cheaper per interaction than training, its global volume is exponentially growing, turning tiny energy costs into a massive, persistent drain.
The Hardware Treadmill: A high-end CPU might last 5-7 years in a data center. The new AI GPUs are considered obsolete in as little as two years. This brutal, accelerated hardware cycle…the constant replacement of power-hungry H100s with even more power-hungry Blackwells…means that the embodied carbon and raw materials tied up in the silicon are never given a chance to pay back their energy debt over a reasonable lifespan.
We are spending the carbon now in hopes of efficiency later. While the industry is working on “smarter” silicon, efficient ASICs for inference, that transition won’t arrive fast enough to save the grid from the current exponential surge.
What Comes Next?
We are moving from an era of Generation Constraints to Connection Constraints.
The most valuable asset in the world right now isn’t the H100 chip; it’s a signed interconnection agreement with a utility company. The “queue” to get on the grid is the new velvet rope.
This is going to force a few things:
Off-Grid AI: Tech giants will stop waiting for the utility. They will build their own SMRs (Small Modular Nuclear Reactors) or massive solar farms with battery storage, effectively taking their ball and going home.
Sovereign Compute: Nations will realize that “compute” is a strategic resource like oil. You will see countries hoarding power to feed their own AI models rather than exporting it.
The Efficiency Wall: We will hit a point where the cost of power makes brute-force AI training uneconomical, forcing a shift to “smarter” chips (ASICs) and maybe, eventually, neuromorphic or photonic computing.
The invisible hand is dealing cards, but the laws of thermodynamics are calling the bluff. The virtual world requires real power, and for the first time in a long time, we are realizing that “unlimited data” was a temporary illusion.
