By Oluwaseyi Ayodeji, Published on oluwaseyiayodeji.com | Regal Stack Newsletter
The question that found me on the school run
Earlier this year, I started a journey I wasn’t sure I had any business starting. I wanted to understand AI infrastructure: the chips, the power, the money, the buildings, and how it all fits together. My biggest fear was that I didn’t know enough to say anything useful, and that I’d run out of things to say within weeks.
The opposite happened. The deeper I read, the more I noticed things that had been there all along. Ideas started showing up uninvited. In the shower. On the highway.
This one arrived on the drive home after picking up my daughter from preschool. I’ve spent years in cloud infrastructure program management, and I know one pattern well. When the people above you need something delivered yesterday, you execute. Fast. Sometimes frantically. And when everyone is heads-down, very few people look up at the bigger picture creating all that pressure.
So I looked up. With everyone racing to build data centers, the giant cathedrals of compute that power AI, are we building enough? And if we’re not, is building more the only answer?
Then I went looking for the questions other people are asking. Residents want to know what a data center will do to their power bill. Investors want to know if this is a bubble. Business leaders want to know why AI capacity is so hard to get. This piece tries to answer them honestly.
First, let me read the room
The data points to a shortfall, and I’ll show you why. But I know how a lot of people feel about these buildings. In an April 2026 Yale and George Mason University survey, 56% of Americans said they’d oppose a data center near them, and 59% expected one to push up their electricity bill [1].
Here’s what struck me. When another national poll asked how America should meet AI’s power needs, the most popular first step wasn’t building more power plants. It was improving energy efficiency [2]. I think the public is onto something. Building more will be part of the answer, but the fastest and least painful wins are inside the buildings we already have.
Is AI demand real, or is this a bubble?
The demand is real today, and tokens show it. A token is a small chunk of a word, and more tokens means more AI work. In its October earnings report, Google said its AI tools for developers now process nearly 32 billion tokens every minute, more than four times the pace of a year ago [3]. Customers have also signed contracts they’re still waiting on: Google’s cloud backlog stands at $561 billion [3], and Microsoft’s at $677 billion [4]. And the clearest sign of all: renting a three-year-old Nvidia H100 chip got about 40% more expensive between October 2025 and March 2026 [5]. Old tech normally gets cheaper. Remember used car prices in 2021? When old things get pricier, there aren’t enough new ones to go around.
“The demand is real today, and tokens show it.”
The money is real too. Alphabet, Amazon, Meta and Microsoft now expect to spend between about $725 billion and $745 billion this year, mostly on AI infrastructure [3][4][6][7]. Alphabet says its spending will rise significantly again in 2027 [3]. Meanwhile, Anthropic and OpenAI alone bring in revenue at a rate of roughly $100 billion a year [8].
The bubble risk sits in the financing. In June, the Bank for International Settlements, often called the central bank for central banks, warned that weak returns on AI spending could trigger an “investment bust” [9]. That’s also an argument for squeezing more out of existing buildings before borrowing billions for new ones.
Do we actually have enough data centers?
Not right now, and the gap isn’t only America’s problem. America hosts most of the world’s AI computing, and customers from every continent rent it [10]. So demand on US data centers is the world’s demand, landing on American soil.
Data centers are measured in gigawatts: how much electricity they can draw at once. One gigawatt is roughly the output of a large nuclear plant. Goldman Sachs expects demand on US data centers to jump from about 41 gigawatts this year to 66 next year [11]. The investment bank Jefferies thinks the country can realistically deliver 15 to 20 gigawatts a year, and found only 12 of the 24 gigawatts scheduled for 2026 actually under construction [12].
That leaves a rough gap of 5 to 10 gigawatts a year, by my back-of-the-envelope math. Nvidia, the biggest AI chipmaker, called its own growth forecast limited by supply [13]. Google says it’s working in a “supply constrained environment” and started renting other companies’ computing capacity at the end of September. Its cloud chief said Google “could obviously serve a lot more demand” with more capacity [3].
Why can’t we just build faster?
Money isn’t the bottleneck. Three other things are.
Permission. In the first three months of 2026, local opposition blocked or delayed at least 75 projects worth about $130 billion, roughly all of 2025’s total in a single quarter [14].
Parts. The biggest grid transformers now take three to five years to arrive [15]. You can’t plug in a building whose plug is still at the factory.
A place in line. Texas alone is tracking about 438 gigawatts of requests to connect to its grid, though officials there say much of it reflects interest rather than firm plans [16].
So will all this cause blackouts? Not under normal conditions, according to NERC, the grid’s reliability watchdog. Its 2026 summer assessment found enough reserves across the regions it studies, with risk mainly during extreme heat [17].
Will a data center raise my electricity bill?
It can, and in some places it already has. In the PJM grid region, which serves 13 states and Washington, D.C., the price utilities pay to reserve power rose roughly elevenfold in a few years. The market’s independent monitor blamed data centers for 63% of one year’s increase, about $9.3 billion that ordinary customers will pay. In Ohio, higher capacity prices add about $16 a month to the average home bill [18].
The fix is rules that make data centers pay their own way. Ohio now requires large new data centers to pay for at least 85% of the power capacity they reserve, for up to 12 years, whether they use it or not [19]. If a project fizzles, households aren’t left holding the bill.
What about water and jobs?
Water is a local question more than a national one. Amazon’s data centers withdrew about 2.5 billion gallons worldwide in 2025, a tiny fraction of US water use. Yet a single Meta data center in Newton County, Georgia, uses about 10% of the county’s supply [20]. Where a data center sits matters more than how big it is.
Jobs are modest. Even the largest US data centers employ only up to about 150 people once running. The bigger local prize is tax revenue. One Kansas city expects about $10 million a year from a single project, more than its entire annual budget [21].
How do we close the gap without paving over everything?
Nearly everything that can come online in the next two years is already being built. So the fastest fix is wringing more out of what we have:
Make every answer cheaper. Google says the energy behind a typical Gemini answer fell 33-fold in one year [22]. Microsoft says software alone got it 90% more output from the same chip [23].
Use the right-sized model. Not every question needs the biggest, most power-hungry AI. You don’t need an 18-wheeler to pick up groceries.
Fill the empty seats. Goldman’s own forecast assumes data centers run at about 70% of capacity on average [11]. Moving work that can wait into quiet hours recovers some of that slack.
Swap the engine, keep the car. Google says its newest chip for running AI delivers 80% better price performance than the last one [3]. Upgrading the hardware adds capacity without a new permit or a new transformer.
Let data centers flex. A Duke University study found the grid could absorb about 76 gigawatts of new demand if big users dialed down for roughly 0.25% of the year, during peak stress [24]. That’s about three years of new demand, served by power plants that already exist, though the study didn’t model transmission limits.
Won’t efficiency just make people use more AI?
Probably, yes. Economists call it the Jevons paradox: make something cheaper and people use more of it.
I see it the other way. Efficiency lets the same building, on the same power line, serve far more people. That buys time to build the next data center where it’s welcome, on terms that protect its neighbors’ bills.
What does this mean for emerging economies?
When the world’s biggest AI landlord runs short, the smallest customers wait longest. That often means startups and governments from Lagos to Nairobi to Jakarta. Their lesson: secure reliable power and community buy-in before building, start with smaller and cheaper-to-run models, and treat firm electricity as a strategic asset that attracts investment money alone can’t buy.
What this can’t tell you
Google is one of the few companies that publishes token counts, so the picture is partial. Estimates of America’s share of global AI computing swing widely: one says 93% [10], while the International Energy Agency put America’s share of global data center electricity at 45% in 2024 [25]. And if chips are sold out, the usage we can measure understates what people actually want. The real gap may be bigger.
Zoom out, then act
That drive home reminded me that the most useful question is often the one nobody has time to ask. Over the next two years, the winners will be whoever gets the most out of every watt they already have, and earns the trust to build the next one.
If a data center is proposed near you, ask whether it will pay its own grid costs, how much water it will use and what your town gets back.
If you buy AI for your business, ask what the same work would cost on a smaller model before asking for more capacity.
If you build or run AI infrastructure, treat efficiency and flexibility as capacity you can sell, and as your best argument at the town hall.
If you’re in an emerging market, start the conversation about power and community buy-in now, before investors arrive asking for both.
Would you welcome a data center in your town? What would it take to earn your yes? Leave a comment, and share this with someone who’s been too busy executing to zoom out.
References
Fortune (republishing The Conversation, on the Yale and George Mason University Climate Change in the American Mind survey, April 2026), Data center opposition poll, October 2026.
Issues & Insights, Americans fear higher costs, strains on electricity grid from data centers: I&I/TIPP poll, August 2026.
The Motley Fool, Alphabet (GOOGL) Q3 2026 earnings call transcript, October 2026.
Stock Titan, Microsoft earnings: Azure grew 48%, beating Microsoft’s own forecast again, October 2026.
Crypto Briefing (summarizing SemiAnalysis), Nvidia H100 GPU rental costs surge, 2026.
TradingKey, Meta Q3 net profit surges 264% YoY, raises full-year capex slightly, October 2026.
The Motley Fool, Stock market today, Oct. 29: Amazon falls on Q3 earnings report and capex concerns, October 2026.
Epoch AI, An update on AI’s most important number, 2026.
Semafor, AI spending boom could end in global bust, June 2026.
Financial Investigator (Payden & Rygel), The US leads the way in AI data centre capacity, July 2026.
Goldman Sachs Research, US data center power demand projected to double by 2027, May 2026.
The Register (on a Jefferies research note), Only half of US datacenter capacity planned for 2026 is actually under construction, 2026.
Rivkin, Nvidia Q2 FY2027 results, August 2026.
Data Center Watch, Q1 2026 report, 2026.
The Next Web, US power firms scramble for transformers and turbines as data centres strain supply, 2026.
New Project Media, ERCOT large-load requests climb to 438 GW, 2026.
America’s Electric Cooperatives, NERC: Generation growth will ease summer strain; blackout risks remain in some areas, May 2026.
IEEFA, Projected data center growth spurs PJM capacity prices by a factor of 10, July 2025.
POWER Magazine, Regulator approves AEP Ohio’s landmark data center tariff, July 2025.
Benton Institute (republishing Ars Technica), When it comes to total water use, AI data centers are a drop in the bucket, June 2026.
Straight Arrow News, The big reason small cities want data centers. Hint: it’s not permanent jobs, July 2026.
Google Cloud, Measuring the environmental impact of AI inference, August 2025.
Tom Tunguz, Beyond a trillion: the token race, September 2025.
Utility Dive (on Duke University’s Nicholas Institute), US grid headroom for flexible load, February 2025.
International Energy Agency, Energy and AI: executive summary, April 2025.
