The first 5 pages, shown as text so search engines and readers can find this document.
Page 1
1
Artificial Intelligence Meets Natural Stupidity: Managing the Risks
Amory B. Lovins
10 May 2025
Copyright © 2025 by Lovins Associates LLC. All rights reserved.
Abstract
Future electricity needs for artificial intelligence (AI) are wildly uncertain—shaped by unproven con-
cepts, disputed performance, limited trust, volatile markets, unpredictable adoption, and technical efficien-
cy that quadruples roughly each year. Yet a speculative surge is driving massive investment in data centers
and new electricity supplies, risking a 12-figure overbuild. Avoiding an electricity bubble requires clear-eyed
analysis, disciplined planning, and using markets to allocate risks fairly to potential beneficiaries.
A cautionary history: In 1999, the US coal industry claimed that information technology would need
half the nation’s electricity by 2020, so a strong economy required far more coal-fired power stations. Such
claims were spectacularly wrong but widely believed, even by top officials. Hundreds of unneeded power
plants were built, hurting investors. Despite that costly lesson, similar dynamics are now unfolding again.
In 2023–24, artificial intelligence (AI) was suddenly hailed as the vital foundation of the next econ-
omy. Big Tech giants, aided by 36 states’ subsidies, are investing a trillion dollars in data centers for AI ser-
vices. Forecasted electricity use soared, but actual national use hasn’t. In 2023, US grid electricity fell; in
2024 it rose 2%, the fourth-fastest rate in the past decade, as data centers’ share (little from AI) crept up from
4.4% to 4.5%. Electricity growth was real in a few hotspots, but was widely misreported as a national trend.
Worldwide, only about 5% of 2024 electricity growth powered new data centers; 5–10% is expected to 2030.
The known unknowables about markets for AI and AI’s need for power may create a new electricity
bubble with hundreds of billions of dollars of overbuilding. Hundreds of new data centers, each drawing the
power of a small city, have been proposed, but many will never be built, and not all built are certain to thrive.
Nonetheless, in the name of powering them, a US energy emergency was declared to speed, favor,
and further subsidize rapid expansion of gas, coal, and nuclear plants already rejected in the marketplace;
the emergency is apparently that they can’t compete. They’re also far too slow: gas turbines are sold out for
6+ years, and nuclear annual global net capacity additions now match just two days of renewable growth.
The market winners—solar, wind, and storage—are adding 93% of new US and 95% of world capac-
ity in 2025, but are officially dismissed (along with efficient and timely use) as unreliable, even as the Penta-
gon’s critical missions deploy them and private industry selects them for the most critical applications, in-
cluding data centers. New federal rules are also stalling queued-up renewables and storage totaling more
than total US generating capacity. They’re ready to go, and far faster and cheaper than the favored options.
A grounded view of AI’s status and prospects can help policymakers and investors avoid repeating
past mistakes. The International Energy Agency expects renewables to raise world electricity supply 10–20
times more than data centers raise demand. Even lower AI-driven demand would free more renewables to
displace fossil fuels sooner. However, an offsetting risk is little-noticed: AI’s most valuable technical uses
include finding and extracting more oil and gas at lower cost. If AI boosts fossil fuel production more than it
saves energy economy-wide, it could worsen emission threats. This hidden dynamic needs urgent scrutiny.
In the electricity sector, especially around emerging data hubs, AI growth poses risks including unre-
liable grids and load forecasts, stranded assets, weakened utility finances and credit, and cost burdens un-
fairly shifted onto other electricity users. State-level utility regulators can manage this by requiring develop-
ers to guarantee power payments, backed by bonds or insurance that can market-price the risk of default.
In the near term, smarter grid strategies can support data center growth without overbuilding. These
include improving demand flexibility, enabling other users’ efficiency gains, and co-locating new data cen-
ters with clean energy at underused gas plants. Just those “Power Couples,” combining clean supply with
flexible demand, can deliver fast, cheap, clean, reliable power to support AI progress on a solid foundation.
As AI evolves, the stakes for energy policy are high. Lost bets on electricity demand could waste ma-
jor investments, lock in unneeded fuel infrastructure, and derail health and environmental progress. But dis-
ciplined foresight can support innovation while avoiding previous errors. Smart regulation, accurate risk pric-
ing and allocation, and market-led investment in proven, least-cost solutions can ensure that the AI era ad-
vances not just technological progress but also resilience, economic stability, and global responsibility.
Page 2
2
Data centers are giant computing centers, built mainly in certain parts of the United States.
They now use about 4–5% of US and 1½% of global electricity. About a fourth of their usage
supports artificial intelligence (AI) tools and activities. In the rest of this decade, data cen-
ters could plausibly account for about a fourth of electricity’s total US growth, but that
fraction is often projected to reach half to three-fourths, and the International Energy
Agency in April 2025 projected 62% (2023–30). Such rapid growth could create big financial
risks for utilities and customers. If a data center is built, or planned but not built, or built
but can’t succeed and pay its bills, its power supply can impose huge costs on others.
Managing those financial risks needs special rules, including a new one suggested below.
Contradicting most media coverage, this synthesis supports three main findings:
• AI’s electricity needs have limited and highly uncertain growth prospects, and even
if large, are a much smaller part of potential total national or global electricity
growth than of its public and private financial risks, especially near data hubs; and
• AI data centers are best paired with renewables—especially quick-to-scale solar
power, which has proven quick, reliable, secure, competitive, and effective; but
• AI is being hyped and officially posited to justify more gas, coal, and nuclear power
—all unneeded, uncompetitive, and likely to raise everyone’s costs and risks.
Excessive electricity demand forecasts
Widespread forecasts that data centers will cause half or even three-fourths of electricity’s
demand growth to or past 2030 extrapolate from specific localities with extraordinarily
concentrated data centers, like Northern Virginia’s world-leading “Data Center Alley.”
Claims from 2019 that this area’s two counties originate or receive 70% of US or global In-
ternet traffic appear to be untrue; nobody knows. But encouraged by state tax breaks just
extended to 2035, the Alley does have 150+ large data centers, and it reports contracts for
new data centers drawing 40 GW (40 billion watts) of electricity—~8× its current data-cen-
ter load, and nearly half the total peak load of Texas. Data centers reportedly used a fourth
of Virginia’s total 2023 electricity, and over a tenth in six states. Unconstrained growth,
mainly for data centers, could nearly triple Virginia’s total electricity use by 2040. Other
hotspots in Georgia, Texas (with its own ~40-GW projection), Ohio, Oregon, North Dakota,
Nevada, Arizona, etc., seek to exploit some combination of state subsidies, cheap power,
discounted utility rates, and compliant regulation. The US Department of Energy has also
offered 16 AI sites on DOE land, and US Army land is likely; neither needs nuclear licensing,
so DOE is expected to propose such siting to bypass the Nuclear Regulatory Commission.
Much of AI-long localities’ data-center-driven electricity demand growth is self-inflicted by:
• 36 states’ taxpayer subsidies for new private data centers (though many of those in-
centives are now opposed and some are being reconsidered);
• opportunity-seeking, revenue-enhancing utility promotion, recruitment, and deals;
Page 3
3
• an apparently common tendency, as a Harvard Law study found, to try to transfer
part of data centers’ electricity costs to existing customers, often using undisclosed
“proprietary” data to sidestep legally required transparent utility regulation; and
• misunderstood complexities of estimating rapidly evolving data-center loads: for
example, forecasters rarely know when AI models will pivot from energy-intensive
training mode to far lighter, sporadic, but very protracted inference mode.
Trillion-dollar tech firms have proven as adept as many other industries in playing off their
suppliers and communities against each other to win billion-dollar tax breaks and other
concessions. Colorado’s biggest utility, for example, offers data centers a 3¢/kWh “eco-
nomic development” discount—lost revenue made up by other customers—for 10 years,
now proposed to stretch to 25 years for ≤40-MW projects (now 20). Tech firms don’t need
the money: the Magnificent Seven in March 2025 had nearly five times the market capitali-
zation of America’s hundred biggest shareholder-owned utilities combined. The Interna-
tional Energy Agency (IEA) estimates that AI-related companies, broadly defined, had
gained ~$12 trillion in market capitalization (November 2022 through 2024), equivalent to
about one-fourth of the total value or three-fourths of the rise in value of America’s 500
most valuable companies.
Data centers would come anyway; owners simply hope to be paid extra for promising local
jobs. Data-center jobs are proving costly and often illusory, but the reallocated costs to lo-
cal communities are real and persistent, and they’re not all financial. Common issues in-
clude water, land, noise, emissions, and public health, with some estimates of health
costs soon rivaling those of the largest states’ on-road emissions. States eager for tax reve-
nue from data centers often deflect or block bipartisan legislation to address constituents’
concerns—as in Virginia, where an independent state study found household electric bills
could rise $14–37/month by 2040, excluding inflation, for reasons related to AI; a 15% rise,
averaging $22/month, is already proposed, ostensibly for other reasons. (Market power is
also alleged to be potentially abused.) Adding huge new loads to antiquated and stressed
power grids may also be degrading power quality and reliability for all customers.
We’ll exclude here the volatile and unknowable potential power demand from mining
Bitcoin and similar cryptocurrency. This product has no significant legitimate use, seems
driven by financial speculation and political conflicts-of-interest, and has sound low-en-
ergy substitutes. The US Department of Energy estimates that cryptocurrency mining used
~0.6–2.3% of US electricity in 2024. For comparison, another load similarly sensitive to
electricity prices, indoor cannabis cultivation, used ~1%, displaceable by outdoor growing.
In Texas, which counts “prospective” loads and lacks crypto-mining restrictions like those
in Amsterdam and China, the ERCOT power pool by February 2025 had received requests
to serve a staggering 99 GW of new loads, more than doubling the 41 GW of requests pend-
ing a year earlier. Each GW is roughly the capacity of a very large power station. These re-
quests come from big new users like data centers (41 GW), cryptocurrency miners (8 GW),
and hydrogen and industrial (31 GW). ERCOT now predicts 73% growth from its record
Page 4
4
peak load of 85.5 GW in August 2023 to 148 GW in 2030. Such huge capacity additions are
a 12-figure bet—mostly on merchant generators’ voluntary behalf, sometimes on utility or
coop customers’ involuntary behalf—that those vast new loads will arrive, remain, thrive,
and pay, long after the power plant and grid have been financed and built. An obvious way
to free up capacity for data centers would be to fix other loads’ inefficiency, notably the un-
or poorly-insulated, drafty, resistance-heated buildings that crashed ERCOT’s grid in
2021’s Winter Storm Uri. However, Texas policy tends to favor expanded supply without
competing it against efficient use, so ERCOT’s forecast assumes only unchanged and un-
impressive 2018–22 rates of efficiency gain, despite the state’s vast potential savings.
Even in the US, as we’ll see below, signals of AI-driven electricity demand growth have
scarcely yet emerged from the noise. AI is estimated to use only 8% of 2023 or 11% of 2025
global data-center power, which in turn totaled only about 1.5% of global electricity use.
Since the world’s 2023 data centers used about 41 average GW but the AI part was only
about 3.3 GW, the new AI processors installed worldwide in 2023 added only ~0.04% to di-
rect electricity use (not counting perhaps an eighth as much in embodied energy to make
the equipment) and ~0.01% to carbon emissions. IEA estimated in April 2025 that training
all large AI models since 2020 had used on the order of 0.001% of global electricity or 0.1%
of global data-center electricity. However, dramatic ambition for future growth, like the
441% global increase in world data-center loads envisaged for 2023–30 by McKinsey or the
362-GW increase projected for 2035 by BNEF, brings the risk of incurring substantial im-
pacts and costs without gaining expected benefits. This risk is most acute for electricity
suppliers and users, especially in areas with many data centers, so we’ll focus there.
The three main US credit-rating agencies warned in 2024 that despite potentially “strong
benefit potential” to utilities building new supplies to power data centers, “the potential for
inaccurate load increase forecasts” could create
…substantial credit risk….Actual loads may turn out to be significantly lower than an-
nounced loads because of technological changes and efficiency improvements that
lower the power requirement for data centers….[exacerbating] long-standing chal-
lenges…in forecasting power supply needs….[T]he risk of overbuilding is most acute for
utilities that are short on power or at capacity to serve existing demand.
Omitting special ratemaking precautions for billion-watt (GW-scale) new loads thus cre-
ates “a considerable risk that residential [and other non-data-center] customers may end
up paying disproportionately” for possibly unneeded utility investments, such as the $3b of
new fossil-fueled power plants just greenlit in Georgia and the same in Louisiana, and
many utilities’ costs of data-center-linked transmission expansions. (Data centers are driv-
ing 80% of Georgia’s surge in projected power sales to 2028—a prospect that five tech
companies challenged in 2024.) So what is that power-plant-and-grid overbuilding risk, es-
pecially for proposed data centers, and how can it be managed thoughtfully, prudently,
and fairly?
Page 5
5
Unknowable electricity demand
Growth in AI data centers’ electricity use depends on growth in AI services—whose uncer-
tainties we’ll discuss below—times how much electricity each unit of service will need,
which in turn is the product of many highly uncertain terms. All these uncertainties multi-
ply into what Google’s head of clean energy calls “incredible uncertainty.” Thus the same
analyst who wrote that data centers account for about 75% of most forecasts’ electricity
growth to 2030–35 was asking within days “whether the data center boom is done before it
even truly started.” The scary part is that nobody knows, so speculation fills the void.
The many-layered uncertainties in data centers’ electricity needs may well compound to
severalfold just in this decade, while the major expansion is happening. (Amazon alone
plans to quintuple its data centers’ IT power, 13% of the US operating total, by adding 12
GW.) More than 2× uncertainty to 2030 is estimated by the utilities’ think-tank EPRI, 4× by
consultancy Wood Mackenzie, 5× (US, 2028) by the International Energy Agency (IEA) or
2.4× for the world in 2035, and 7× for third-party estimates (world, 2030) cited in that IEA
report. Over a few decades, compounding uncertainties could become far greater—per-
haps a millionfold. For example, resistive random-access-memory analog operations have
already demonstrated about a thousandfold lower energy use than modern graphic pro-
cessors. In 1999, IT pioneer Bill Joy estimated a trillionfold potential gain in AI problems
solved per joule of energy by ~2035, half each in hardware and software. (That approxi-
mates Koomey’s Law—computing energy efficiency doubles every ~18 months—and it’s
~10× less than the 1013× by which Hervé Touati estimates the cost of a megaflop of compu-
ting fell during 1965–2025.) In 2025, Joy’s estimate leaves about a millionfold still to go.
That’s ~4× or –75% per year.
Let that sink in. It means depreciation and amortization schedules—decades in the world
of atoms—may shrivel in the radically faster-moving world of bits. Some costly digital as-
sets can become obsolete about as fast as they can be built. Although data centers’ eco-
nomic life is longer—better algorithms can often run on older hardware, newer servers can
often plug into longer-lived data-center infrastructure, less-capable data centers can serve
lower-cost needs, and older servers can be resold into secondary markets—technical dy-
namism roils value projections. With 12-figure investments at stake under intense time
pressures, immediate uncertainties risk corporate failures and economic disruptions—es-
pecially for conventional AI power supplies that do take decades to recoup investments.
Anyone building a thermal power plant paying back in 20–30 years, to run an AI data center
whose energy use per inference may well drop ~75% per year, had better have a solid alter-
native market; otherwise it’s a bet on sustaining decades of rapid exponential growth in AI
services. Without that, you could have a stranded asset, akin to a 1976 Cray-1 computer
that you could later replace for less than the cost of a year’s electricity to run it ($100k at a
10¢/kWh tariff). A 2023 iPhone 15 is ~60,000× cheaper than a Cray-1, ~180,000× less en-
ergy-intensive, and ~5,000× more powerful, yet it depreciates in a few years. How about
the Cray? Sure enough, NVIDIA’s CEO suggests his firm’s graphic processor products less
than two years old should retire and their electricity be used instead in newer, faster ones.