Electricity Isn’t Boring Anymore
AI Turned the Grid Into the Real Battleground—and Why Geothermal Is the Most Underpriced Constraint‑Release Valve
I used to think electricity was the wallpaper of modern life.
You don’t discuss wallpaper. You don’t build a worldview around it. You notice it only when it peels.
Then AI showed up and started peeling the wall.
In 2024 the United States generated about 4,391 TWh of electricity. China generated about 10,087 TWh—more than double, with a completely different posture: build everything, everywhere, fast.
The U.S. spent years living inside a stable electricity budget. That era is ending.
AI doesn’t politely sip power. It gulps it. Large, fast, near-constant power—in specific places—on timelines utilities were never built for. And it wants that electricity to come with a clean label.
Once you see the situation that way, the argument shifts. You stop debating whether a model is “the next iPhone,” and you start asking the question that feels uncomfortably geopolitical:
Who gets the electrons?
That’s my thesis: AI turns electricity from background utility into a strategic bottleneck. If you want to understand the next decade of tech, you have to understand the grid the way you’d understand a wartime supply chain: where it pinches, who controls it, how it breaks.
And once you look for pinch points, you notice the split between what sounds good and what ships.
Space data centers make a great slide. As a 5–10 year fix for terrestrial AI power demand, they look like science projects.
The near-term winners will be the companies that unclog the boring stack: interconnection, substations, transformers, turbines, transmission, plus credible firm power.
The most interesting firm-power candidate that’s clean-ish and plausibly buildable before the mid‑2030s isn’t exotic.
It’s geothermal.
Geothermal matters for a harsh reason: it maps to capabilities we already have—drilling, reservoir engineering, project finance, and plants that run when the sun doesn’t.
The U.S. grid story we misremember
The last two decades of U.S. electricity didn’t run on a clean-energy fairy tale. It ran on fuel switching.
Coal fell hard.
Natural gas rose hard.
Wind and solar went from rounding error to material.
Nuclear mostly stayed where it was.
I’m not trying to score political points here. I’m trying to explain why the U.S. got away with change: load growth was quiet.
For a long stretch, U.S. electricity demand was close to flat. That created a kind of national illusion: we could clean up the grid without expanding the grid.
AI breaks that illusion.
EIA’s AEO 2026 framing is blunt: after roughly 15 years of near-flat consumption, electricity demand is rising again, with data centers as a major driver.
DOE’s public summary of the LBNL work is even more direct: domestic data-center energy use is expected to double or triple by 2028.
If you’ve ever waited on a transformer, you can feel what “double or triple” means.
The change goes beyond volume. AI load has a different shape: steady, relentless, hard to shed.
The numbers that matter (because units are the trap)
Most debates about AI and energy turn into a mess because people mix capacity, generation, load, and primary energy like they’re interchangeable.
They’re not.
Here are the anchors I keep on a sticky note:
1 GW of 24/7 average load = 8.76 TWh/year.
If U.S. data centers in 2030 consume 300–600 TWh/year, that’s roughly 34–69 GW of average load.
Average load is the brutal part. A data center behaves less like a stadium and more like a factory that refuses to sleep.
Now zoom out.
The U.S. generating 4,391 TWh in 2024 sounds huge until you imagine data centers alone taking 300–600 TWh. At the high end, that’s a big slice of the pie.
Then you look at China: 10,087 TWh in 2024.
Different order of magnitude.
Mental model: the U.S. has been managing a mature system; China has been building a new one. AI lands like a load-growth shock on the first kind of system.
The bottleneck stack (why no single tech bails you out)
When people say “just build more renewables” or “just build more nuclear” or “just use gas,” they’re taking a multi-constraint problem and turning it into a slogan.
In reality, AI power is a stack of constraints:
Interconnection queues (the place projects go to die quietly).
Transmission (the thing everyone agrees we need and almost nobody can permit quickly).
Substations and distribution capacity (you can be near transmission and still be power-poor locally).
Transformers (lead times that feel like a joke until you need one).
Gas turbines (fastest dispatchable solution, but also a supply chain with its own backlog).
Cooling water and heat rejection (data centers are thermal machines).
Permitting (the invisible hand that often slaps).
This is why the “AI power boom” doesn’t translate neatly into “build X.”
It’s also why utilities in data-center-heavy regions suddenly talk like growth companies. Why hyperscalers chase 24/7 power contracts instead of buying annual credits. Why nuclear restarts show up in conversations again.
And why geothermal—tiny on paper—starts looking strategic.
Geothermal: real, difficult, and worth taking seriously
Geothermal has a reputation problem.
It gets filed mentally next to “small hydro,” “wave power,” and “cool but irrelevant.”
Conventional geothermal is absolutely real. It has run for decades. It produces high-capacity-factor electricity.
The problem is geography.
Classic hydrothermal resources are not evenly distributed. If geothermal were as easy as “drill anywhere,” we’d already be doing it.
So the story shifts to Enhanced Geothermal Systems (EGS)—creating permeability and circulation in hot rock where nature didn’t hand you an easy reservoir.
DOE’s definition is straightforward: inject fluid under controlled conditions to create or re-open fractures, circulate water through hot rock, bring it back up, generate electricity.
The implication is the point:
EGS tries to turn geothermal into an industrial drilling business.
That’s why it’s plausible. The oilfield supply chain already knows horizontal drilling, stimulation, sensing, and iterative learning.
Plausible doesn’t mean guaranteed.
NREL’s 2024 Annual Technology Baseline is careful here: near-term EGS costs are modeled predictions and do not have calibration to commercial-scale dedicated EGS plants operating in the U.S.
That’s a polite way of saying: don’t assume the cost curve.
So what’s the right stance?
Treat geothermal like a manufacturing process you have to prove in the field—milestone by milestone.
Which brings us to the flagship case.
Fervo: the flagship case for “geothermal that behaves like a project”
If you want one company that best represents the current attempt to industrialize EGS, it’s Fervo Energy.
The company’s story matters less than its milestones.
Project Red: a 3 MW commercial pilot in Nevada, supplying Google.
Cape Station: planned at 500 MW in Utah, with first power targeted late 2026 / early 2027, and further ramp targets discussed for 2028.
The numbers are small and large at the same time.
3 MW barely moves a national grid. As proof of approach, it matters.
500 MW is big enough that the grown-up problems show up: financing, permitting, reservoir performance, construction schedules.
This is the point: geothermal is the opposite of software. You don’t “ship” it. You build it. And the only way it becomes bankable is by surviving reality.
There’s a quick mental math that keeps you honest:
A hypothetical 10 GW geothermal fleet running at 90% capacity factor produces about 79 TWh/year.
That’s meaningful.
It’s also nowhere near enough to solve a world where U.S. data centers might be consuming 500–900 TWh/year in the 2030s.
So why care?
That hardest slice of the stack has a name: clean, 24/7, high-capacity-factor power.
That’s what hyperscalers actually need if they mean their carbon commitments literally.
“Just put the data centers in space” (and other elegant ideas)
The space data center idea is catnip.
Unlimited sunlight. Cold background. No land disputes. A science-fiction feel.
As a concept, it sounds like a cheat code.
Then you remember data centers aren’t just chips.
They’re industrial facilities designed around:
power conversion
reliability and redundancy
maintenance and replacement cycles
networking and bandwidth
cooling and heat rejection
physical security
Space makes many of those harder.
The constraint people underestimate most is thermal.
On Earth, you move heat with air and water and engineered systems you can service. In orbit, there’s no convection. You radiate heat. Radiators cost mass and area. AI compute produces brutal heat flux. The cooling problem doesn’t disappear in space; it mutates into a mass problem.
Then there’s radiation hardening, orbital debris, replacement logistics, and the simple fact that many AI workloads are bandwidth-hungry and latency-sensitive.
Could space compute exist as a niche? Sure.
Could it “solve” terrestrial AI power demand before 2035? That’s the leap I don’t buy.
Geothermal has a path to relevance in 5–10 years. Space data centers have a path to fascinating demos.
Those are different timelines.
The contrarian frame: the real winners are boring
When an economy hits a bottleneck, the market usually overpays for the sexy narrative and underpays for the unglamorous constraint-relief.
AI is the sexy narrative.
Constraint-relief is:
transformer manufacturers
grid equipment
engineering and construction
turbine makers
utilities that can rate-base the build
firms that can deliver firm power fast
This doesn’t mean those stocks go up. It means those categories will matter—and they’ll be argued about, regulated, subsidized, and fought over.
And it means the “AI trade” is quietly becoming an infrastructure trade.
Market watchlist (indirect exposure / monitoring categories)
This is a monitoring framework, not a list of recommendations. This is not investment advice.
1) Independent power / nuclear-heavy portfolios (firm power story)
Constellation Energy (CEG)
Vistra (VST)
AES (AES)
2) Geothermal incumbent operator/developer (conventional geothermal exposure)
Ormat Technologies (ORA)
3) Oilfield services / drilling transfer (EGS toolchain optionality)
SLB (SLB)
Halliburton (HAL)
Baker Hughes (BKR)
4) Grid equipment and electrification (the bottleneck merchants)
Eaton (ETN)
Quanta Services (PWR)
Hubbell (HUBB)
Siemens Energy (ENR / “Siemens Energy”)
GE Vernova (GEV)
Vertiv (VRT)
Schneider Electric (SU / “Schneider Electric”)
5) Utilities in data-center-heavy regions (rate base + customer concentration risk)
Dominion Energy (D)
American Electric Power (AEP)
Southern Company (SO)
Duke Energy (DUK)
Entergy (ETR)
How to use this list: track capex plans, interconnection commentary, transformer lead times, turbine order books, regulatory filings, and load-growth disclosures. Don’t turn categories into certainties.
What I’m watching next (the real signal)
If you strip away hype and look for falsifiable checkpoints, the next 18–36 months come down to a handful of observable realities:
Do U.S. data-center load forecasts keep getting revised up?
Do transformer and switchgear lead times come down, or stay sticky?
Do gas turbine backlogs worsen?
Do nuclear restarts/uprates move from press release to timeline execution?
Does EGS produce sustained multi-year performance, not just short tests?
Do hyperscalers sign more 24/7 firm clean contracts, or quietly settle for annual offsets?
If those answers lean one way, the grid becomes the story of the decade.
If they lean another way, the AI boom migrates toward available power instead of fashionable zip codes.
Either way, electricity has stopped being wallpaper.
It’s now a map of leverage.
The skeptic’s corner: what could break this thesis
A good story becomes a dangerous story when you forget what could falsify it.
Here’s what could break the “AI makes electricity strategic” narrative—or at least shrink it:
1) Efficiency snaps back. If model architectures, chips, and inference optimization deliver another era of compounding efficiency (the way data centers did from 2015–2019), the headline TWh numbers could disappoint.
2) AI demand shifts geographically. Power is uneven. If new compute migrates to regions with surplus hydro, nuclear, gas deliverability, or friendlier permitting, the U.S. bottleneck might become a regional story rather than a national one.
3) Utilities and regulators slow the machine. Large-load tariffs, moratoria, or stricter reliability rules can turn “demand” into “requests that never connect.”
4) Geothermal executes—then stalls. EGS can work technically and still fail commercially if drilling costs don’t fall, if reservoir performance decays, or if seismic and permitting risks become political.
5) The capital cycle flips. Grid equipment and construction are cyclical. A wave of over-ordering can turn scarcity into surplus, and the market will re-price the whole story fast.
None of these cancels the strategic reality. They change the slope.
Here’s the uncomfortable part: the next AI map may look less like a software leaderboard and more like an electrical atlas.
Northern Virginia matters because of fiber, customers, and inertia. Now it has to bargain with substations and generation.
Texas matters because it can build. ERCOT still has weather, congestion, and reserve-margin politics.
The Pacific Northwest has hydro and fiber. Land, transmission, and community resistance can still slow the clean story.
Georgia, Ohio, Indiana, Arizona, and the Carolinas used to be “secondary markets.” They’re becoming test sites for whether America can add load faster than it adds excuses.
Read every press release through that lens. A hyperscaler announcing a new campus makes a power claim, not just a cloud-capex claim. A utility load forecast becomes a strategic document. A transformer backlog turns into a throttle.
The boring nouns turn into verbs: connect, permit, drill, finance, cool, dispatch, interconnect.
Watch who can do those verbs repeatedly, under pressure, with regulators watching and capital costs still biting. That is the dividing line. The next phase of AI will not be won only inside model labs. It will be won at substations, gas interconnects, drilling pads, turbine factories, transmission hearings, cooling loops, and project-finance desks. The model may be digital. The constraint is physical.
That’s where the real leverage moved.
Source Framework
U.S. EIA — Annual Energy Outlook (AEO): https://www.eia.gov/outlooks/aeo/
U.S. EIA — AEO narrative: https://www.eia.gov/outlooks/aeo/narrative/index.php
U.S. EIA — Electric Power Monthly: https://www.eia.gov/electricity/monthly/
U.S. EIA — Electric Power Annual: https://www.eia.gov/electricity/annual/
Our World in Data (Ember) — electricity generation by source (CSV): https://ourworldindata.org/grapher/electricity-prod-source-stacked.csv
Ember — Electricity Data Explorer: https://ember-energy.org/data/electricity-data-explorer/
DOE — Data center electricity demand summary (Dec 2024): https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers
Goldman Sachs — AI poised to drive 160% increase in data center power demand: https://www.goldmansachs.com/insights/articles/AI-poised-to-drive-160-increase-in-power-demand
DOE — Enhanced Geothermal Systems overview: https://www.energy.gov/hgeo/geothermal/enhanced-geothermal-systems
NREL — Annual Technology Baseline 2024, geothermal: https://atb.nrel.gov/electricity/2024/geothermal
DOE — GeoVision report (PDF): https://www.energy.gov/sites/default/files/2019/06/f63/GeoVision-full-report-opt.pdf
DOE — FORGE (Frontier Observatory for Research in Geothermal Energy): https://www.energy.gov/eere/forge/forge-home
Fervo Energy — Technology overview: https://fervoenergy.com/technology/
Fervo Energy — Cape Station record production results: https://fervoenergy.com/fervo-energys-record-breaking-production-results-showcase-rapid-scale-up-of-enhanced-geothermal/
Fervo Energy — Cape Station groundbreaking: https://fervoenergy.com/fervo-energy-breaks-ground-on-the-worlds-largest-next-gen-geothermal-project/
Fervo Energy — 320 MW SCE PPAs: https://fervoenergy.com/fervo-energy-announces-320-mw-power-purchase-agreements-with-southern-california-edison/
Fervo Energy — 31 MW Shell Energy PPA: https://fervoenergy.com/fervo-energy-announces-31-mw-power-purchase-agreement-with-shell-energy/
Constellation — Crane Clean Energy Center / Microsoft PPA announcement: https://www.constellationenergy.com/news/2024/Constellation-to-Launch-Crane-Clean-Energy-Center-Restoring-Jobs-and-Carbon-Free-Power-to-The-Grid.html