2025人工智能的崛起：对美国经济与能源影响的实证审视研究报告（英文）

NOVEMBER

13,2025TheRiseofAI:ARealityCheckonEnergyandEconomicImpactsMarkP.

MillsExecutiveDirector,NationalCenterforEnergy

AnalyticsDistinguishedSeniorFellow,

TexasPublicPolicyFoundation

DistinguishedFellow,HammInstitutefor

AmericanEnergyARESEARCHPARTNERSHIPBETWEEN:TheNationalCenterfor

Energy

Analyticshttps://energTheHammInstitutefor

American

Energyhttps://hamminstitute.orgTheRiseofAI:ARealityCheckonEnergyandEconomicImpactsExecutiveSummaryTotal

private

sector

spending

committed

to

artificial

intelligence

(AI)

is

nowasignificantshareof

U.S.GDP.Spendingon

building

data

centers

alone—

exceeding$50

billionannuallyand

rising—has

nowsurpassedspendingon

all

othercommercial

buildingscombined.The

evidence

ofan

emerging

structural

shift

intheU.S.

economy

canbe

seenin

thecombinationof

theepicspendingon

AI,rapid

adoption

of

AI

tools,

theimplicationsfornationalsecurity,thevigorousdebatesovertheimpactof

AI

insociety,and,as

well,

thestock

marketenthusiasms.Thecoreconsequenceof

AIdeployedatscaleisinits

potentialfor

boosting

productivity,

the

feature

of

every

economy

that

drives

growth

and

prosperity.

If

democratizingAIelevatesU.S.productivitygrowthonlytotheaverageof

the

past

half-century,

it

will

add

a

cumulative

$10

trillion

more

to

the

U.S.

GDP

than

is

now

forecastover

thecomingdecade.Often

ignored

inAI

forecasts:the

additionalwealth

createdby

using

the

newtechnologyleadstobehaviorsthatusemoreenergy.Anextra$10trillion

would

lead

to

increased

overall

energy

use

equivalent

to

about

five

billion

barrelsmoreenergyoverthenextdecade.Suchawealth-inducedincreasein

energy

consumptionwillbe

far

greater

than

the

quantity

ofenergy

needed

topowerthewealth-creatingAI.Butunleashingthiseconomicand

strategic

opportunity

requires

building,and

powering,

the

AI

infrastructure.AI

data

centers

are

not

unique

in

that

regard.

Energy

is

the“master

resource”

neededforoperatingeverypart

ofcivilization.As

Nvidia

CEOJensenHuang

recently

said:“AI

is

energy,AIis

chips,the

models,

and

the

applications

.

.

.

.Andwe

need

more

energy.”As

it

happens,

data

centers

are

measured

and

tracked

in

termsof

power

in

watts,

notdata

in

bits.What

is

unique

about

AI

data

centers

is

the

scale

and

velocity

ofpower

demandsnow

emerging.

Some

individual

datacentersnow

underconstruction

willhavecity-scalepowerdemands,andhundredsarebeingbuiltorplanned.

Therateof

construction—especiallyincombinationwithreshoringmanufac-

turingandreanimating

basicdomesticindustries—hasendedthetwo-decade

interregnumof

flatelectricsector

growth.Policymakers,

investors,

and

businesses

in

the

AI

supply

chain

are

all

interested

in

discerningjusthow

much

additional

electricitywillbe

needed

specifically

for

powering

data

centers

and

how

it

will

be

supplied.

For

this

analysis,we’verevieweddozensofdifferentindustryandtechnicalreportsto

lookforcluesand

consensus.Thefacts

andtrends

suggestAI

digital

demandswill

requirebuilding

noless

than

about

75

GW,

possibly

as

much

as

100

GW

of

generation

by

2030.

And

neitheroutcomeincludestheelectricitydemandsforexpandingtheancillary

butdirectlyrelatedtelecommunicationsnetworks,aswellasthat

needed

for

reshored

chip

fabrication

facilities

that

will

manufacture

the

logic

engines

inside

thedata

centers.Tomeetthatmuchnewdemandby2030,underlyingengineeringrealities

showthatmostoftheadditionalelectricitygenerationwillnecessarilycome

from

burning

natural

gas.

That

will,

in

turn,

require

about

a

10%

to

20%

increase

inoverall

U.S.gas

production.That

rise

in

gas

demandwill

occur

contemporaneouslywith

roughlythe

sameamountofnewdemandcomingfromadditionalLNG

exportterminals

thatarealready

under

construction.The

nation

is

capable,

technically,

of

meeting

such

a

level

of

growth

in

naturalgasproduction,pipelineinstallations,andpowerplant

construction.

The

primary

impedimentsare

institutionaland

regulatory.In

the

longer-term,

as

the

AI

structural

revolution

continues

to

play

out

past2030,evengreaterenergydemandswillemergetopowerthenextphase

of

growth.

Thosedemands

willlikely

bemetincreasinglyfromnuclearenergy

andadditionalsolarcapacity.Buteachof

thoserequireassociatedinfrastruc-

tureexpansionsthat,inherently,takefarlonger

than

the

current

torridpace

of

AIconstruction.

Nevertheless,

if

relevant

policiesand

projectsare

not

put

in

placetoday,electricityplannersintheforeseeablefuturewillbeagaincaught

flat-footed

in

failing

to

prepare

for

growth.Thechallengeisthatelectricity-relatedenergy

policies

now

in

place

were

framed

during

the

recent

period

of

low

or

no

growth,

combined

with

misguided

pursuitsofanenergytransitiontoreplace

conventional

energy

sources.The

AIboomhasilluminatedthefactthatnewperiodsof

growthareinevitable—

even

if

difficult

to

predict—and

that

such

periods,

predictably,

lead

to

increased

energydemandsrequiringadditionsto,ratherthanreplacementsof,existing

energysystems.Googleobservedearlierthisyearthat“AIpresentstheUnitedStateswitha

generationalopportunityforextraordinaryinnovationandgrowth”

but

thatitrequires“expedited

effort

to

increase

the

capacity

of...U.S.energy

systems.”A

July2025White

House

directive,Winning

the

Race:

America’s

AI

Action

Plan,

called

on

the

private

sector

to

build

the“vast

AI

infrastructure

and

the

energy

to

powerit.”

Thatisprecisely

whatisunderwayinanhistoricconvergenceof

the

technology,energy,andfinancialsectors.2KeyTakeaways

withImplications

forPolicymakersThe

great

AI

inflection

is

market-driven:•Therace

to

build

AIinfrastructuresandservicesis

beingfunded

by

theprivate

sector1i.e.1it

is

not

a

policy-driven

transformation.•Thepaceof

AIadoptiondependsonintegrationchallengesforspecificapplications1and

the

maturity

and

efficacy(power)ofthe

AI

tools1i.e.1

morepowerfulandusefulAItoolsarerapidlyemergingthatwilldrive

moreadoption.Three

key

metrics

illuminate

the

transformation:•The

rate

of

progress

in

the

underlying

AI

compute

capabilities

follows

a

well-established1

predictable

trendof

exponential

gains.•AIcomputationalperformancehasincreasedathousandfoldsince

2018

andis

continuing

at

that

pace.•Thescaleand

velocityof

privatecapitaldeployed

to

build

AIinfrastruc-

ture

is

setting

records

but

has

not

(yet)

exceeded

historic

highs

in

comparable

periods.•Over

$1

trillion

or

private

capital

is

planned

or

on

track

to

be

spent

to

build

AI

infrastructure.•The

scale

of

energy

demands

for

AI

infrastructures

will

require

substan-

tialexpansionof

bothelectric

powerand

natural

gas

production.•The

likely

pace

of

AI

will

require

75–100

GW

of

new

electricity

generat-

ing

capacity

to

supply

as

much

as

11000

terawatt-hours

ayearmore

electricity

for

digital

demands

by

the

early

2030s.•The

new

power

plants

will

drive

the

need

for

a

10%–15%

overall

increaseinU.Snaturalproduction,contemporaneouswitha

similarincrease

in

demand

for

LNG

exports.A

key1

enabling

requirement:

meet

the

unprecedentedscale

ofelectricitydemand

of

individual

facilities1

atthe

velocity

being

built.•Hundreds

ofplanned

data

centers

have

demand

exceeding

300

MW

each1manyover1GW1with

construction

completions

often

in

two

or

threeyears.Therearenoprecedentsinutilityhistoryforsuchscaleor

velocity.•The

pace

and

scale

are

in

tension

with

supply

chains

and

workforce

availability

in

a

sectorthat

is

accustomedto

flat

growthwith

regula-

tions1policies1

and

conventions

adopted

in

recentyearstopursue

an

energy

transition.•Pace

and

scale

are

also

in

tension

with

the

need

to

not

compromise

grid

reliabilityor

imposecostson

residential

rate

payers.Engineering

realitieswilldictateviablesolutions.

There

aretwo

timeframes

for

meeting

AI

power

demand:•Twotofive

years:Various

classes

ofnatural

gas

turbines

and

engines

dominate.•The

technical

capacity

exists

to

meet

the

demands

for

gas

production,pipelines,

and

power-producing

engines.

Both

grid-integration

andprivate

grid

approaches

to

projects

are

underway.•Five

to

10years:

Feasible

for

new

nuclear

generation

and

expanding

transmissiontoaccessutility-scalesolar/wind

facilities

(requiring

grid-scale

batteries).•Both

face

supply

chain

challenges;solar,wind,batterieswithforeignor

China

sourced

inputs;nuclear

with

restoring

atrophied

infrastruc-

tures.

Thereis

no

clear

path

to

significant

acceleration

of

transmission

construction(natural

gas

pipeline

construction

is

far

faster).TheRiseofAI:ARealityCheckonEnergyandEconomicImpacts3Ithasescapednoone’sattentionthateye-wateringamounts

ofcapital

are

beingdeployedin

the

privatesectoronartificialintelligence(AI).Similarly,it’s

nowcommonknowledgethatthegreatAIbuild-outisinducingincreasesin

electricdemand

notseen

fordecades.Aswithalltechnologies,whatappearstobeanovernightphenomenonin

factcomes

from

years

of

engineering

developments

and

incremental

progress.

History

showsthatwhentherelevant

capabilitiesbecome

good

enough

and

costs

are

low

enough,

a

tippingpoint

is

reached,

and

an

inflection

happens

with

rapidand

“surprising”

growth.Economic

historian

Joel

Mokyr,

theco-recipientof

this

year’s

Nobel

Prizein

Economic

Sciences,

borrowed

from

physics

the

term

“phase

change”

to

describe

such

inflections,

when

transformational

innovations

trigger

economic

and

socialrevolutions.1

Thechangeintheeconomyandindailylife—frompre-to

post-railroad,from

an

agrariantoindustrial

society—was

aphase

change

as

dramaticasa

liquid

becominga

solid.AsprofessorMokyrwrotein

one

ofhismany

seminalbooks,

TheLeverof

Riches:

“Technological

progress

has

been

one

of

the

most

potent

forces

in

historyinthatithasprovidedsociety

with

whateconomistscalla

‘freelunch,’

thatis,anincreasein

output

that

is

not

commensurate

with

the

increase

in

effortandcost

necessary

to

bring

itabout.”

2Buttechnologies’economic

and

societal

benefits

aren’t

unlocked

in

isolation.

Mokyr

is

not

naïve

inhis

“freelunch”observationbut

rather

uses

thataphorism

toframestructuralrevolutions,i.e.,howa

phasechange

occurs

in

the

real

world.In

Mokyr’s

framing,

such

revolutions

occur

at

the

intersection

of

three

forces:

newknowledge

instantiatedthroughtechnology,the

availability

and

sources

of

capital

to

deploy

technologies,

and

the

role

of

institutions

that

enableorconstrain

deployment.Few

doubtthatAIisconsequential,

even

ifthere

are

somewho

are

more

anxiousthanexcitedaboutthepossibilities.Thenature

oftheopportunities

andwhatitwilltaketoensurethattheUnitedStatescancapturethebenefits

fromAIcanberevealedbyansweringthreequestionsaboutthefuture,with

all

the

usualcaveatsabout

predictions,around

which

this

report

isorganized.1.Is

AI

a

structural

shift

or

a

bubble?2.

Whatare

the

upstream

power

implications?3.

Whataretheconstraintstounleashingthenecessaryenergy

to

fuel

an

AI

boom?IntroductionTheRiseofAI:ARealityCheckonEnergyandEconomicImpactsStructureofTechnologicalRevolutions4The

release

ofChatGPT

on

November

30,

2022,

led

towidespread

public

recognition

that

AI

was

now

good

enough

to

become

increasingly

useful—and

to

disrupt

businesses,

institutions(and

thus

policymaking),

and

stock

markets.

It

was

theclimaxof

decades-long

technologicalachievements.The

arrival

of

supercomputers

powerful

enough

to

execute

the

complex

mathematicsandcodingof“machinelearning”

arrived,

predictably,

because

ofthe

inexorable,

exponential

gains

in

compute

performance

(measured

in

FLOPS/second,

or

logic

operations

per

second,

wherein

conventional

computer

progressisoftenmeasuredinthemoregeneralMIPS,millioninstructions

per

second[MIPS]).

As

with

the

first

computer

revolution,

the

rate

of

improvement

continues

unabatedandexponentially.Theera

when

binarylogiccomputers

weredemocratized,circa1980,offers

an

analogy

forthe

state

and

future

of

the

emergence

of

practical

AI—a

technol-

ogythathasbeeninexistencefordecades,

justasdigitalcomputershadbeenby

1980.The

origin

ofdigital

computersbased

onbinarylogictracesto

1937,

with

ClaudeShannon’sseminalMITmaster’sthesis,theMagnaCartaof

thedigital

age.3

Alan

Turing,

a

colleague

of

Shannon,

observed

atthe

time

that

the

Colossus

computer(builtduringWorldWarII,

just

beforetheU.S.Eniac)couldperform

tasks

that

wouldotherwise

have

taken,as

Turing

issaid

to

haveobserved,“100

Britonsworkingeighthoursadayondeskcalculators100years”tocrackthe

Germancode.4

Somefourdecadeswouldpassbeforetheageof

thePCwould

arrive

becauseof

the

inexorable,exponential

progress

incompute

power.Similarly—nearlyfourdecades

before

thereleaseof

ChatGPT—theconcept

ofalearningalgorithm,of“machinelearning,”tracestoaseminal1986paper

co-authored

by

Geoffrey

Hinton,

credited

as

a

“godfather”

of

AI.

But

the

silicon

engines

that

could

realize

Hinton’s

vision

did

not

emerge

until

1993,

when

Jen-Hsun

“Jensen”

Huangco-founded

Nvidia.

5As

with

earlier,

conventional

logic,

it

took

decades

of

advancement

for

silicon

hardware

to

becomesufficiently

powerfuland

inexpensive

todemocra-

tizethemassivelyparallelfunctionsinherenttoAI.Thepasthalf-dozenyears

have

seen

athousandfold

increase

in

AI

computational

power,

unlocking

useful

AI—a

pace

that

is

notslowing;

indeed,evidence

points

toanacceleration.6Source:KonstantinF

.Pilzetal

.

,“TrendsinAI

Supercomputers,”arXiv,

Apr

.23,2025Economic

PerformanceGainsofConventionalComputers1.1Pillar

of

the

Boom:Compute

PowerPerformanceGainsinAISupercomputersSource:HansMoravec,“When

Will

Computer

Hardware

Match

the

HumanBrain?

”JournalofEvolutionandTechnology

,vol

.1(1998)

.1.

The

Structural

ShiftTheRiseofAI:ARealityCheckonEnergyandEconomicImpacts5The

July2025White

House

directive1Winning

the

Race:America’s

AI

ActionPlan1summarizedthestateofplayregarding

the

AI

imperative1

issued

a

call

to“buildandmaintainvastAIinfrastructureandtheenergytopowerit1”andoffered

a

directive

that

the

nation

should

“Build1

Baby1

Build!”7Private

sector

spending

on

data

centershadbeenrunning

atroughly

$10

billionperyearratebeforethe2022releaseofChatGPT.Sincethen1spending

tookoff—datacenterconstructionthisyearsurpassedconstructionspending

onall

otherU.S.commercialbuildings.The

spend

rate

exceeds

current

total

spending

to

buildeither

manufacturing

facilitiesor

power

plants.8Some

analysts

worry

that

the

spending1

and

the

associated

escalation

in

stockvaluations

ofall

the

companies

in

theAI

ecosystem1

upstream

and

downstream1pointstoaclassicbubble.9

However1BlackRock

recently

made

a

beton

thefuture

witha$40

billionacquisitionof

AlignedDataCenters—the

”largestdatacenterdeal

in

history”—in

aclearsignalof

that

firm1s

expectation

of

a

long-run

trend1

not

a

bubble.10

Of

course

there

is

a

correlation

between

the

twodomainsof

whatinvestorsthinkastockisworthandwhatisbeingbuilt

and

planned1andexpectationsabout

the

future

usesof

AI.Total

capital

spending

on

AI

infrastructures

is

on

track

to

exceed

$1

trillion

a

year1

withsome

analysts

predicting

acumulative$5

trilliondeployed

by2030.11

While

only

hindsight

will

reveal

whether

or

not

private

and

public

market

investorsentimentswereoverexuberant1thereare

anumber

of

basic

indica-

tors

that

can

reveal

whether

the

AI

boom

is

structural1

i.e.1

a

secular

shift

in

theeconomy1or

a

stock

fad.Note:

Spending

recordedin

the

yearconstruction

starts

.Source:LydiaDePillis,

“TheA

.I

.SpendingFrenzyIsProppingUp

theRealEconomy,Too,”

TheNewYork

Times

,

August27,2025.BigTechCapitalSpending1.2The

Spending:Data

CentersSpendingonConstruction:Officesvs

DataCentersSource:IanHarnett,TheAICapex

Endgame

Is

Approaching

,Financial

Times,October

3,2025.TheRiseofAI:ARealityCheckonEnergyandEconomicImpacts6Overall

U.S.

business

spending

on

information

infrastructures

has1

since

the

start

ofthe

21st

century1

dominated1

dominated

overall

spending

on

the

four

core

areas

ofprivate

sector

investment1

substantially

exceeding

capital

deployedforindustrialequipment1transportationequipment1andstructures.

Theshareofthatinformationinfrastructurespending

hasbeen

increasingly

shifting

from

purchasesof

on-sitecompute

toenterprisecloud-basedservices.

The

arrival

of

AI

accelerates

the

trend

because

the

dominant

locus

of

AI

spending

isoncloud

infrastructure.The

current

level

of

data

center

capital

deployed

(annual

spending

on

construction1ratherthanstatedtotalcommitments)canbeputinthecontext

of

the

overall

scale

of

nearly

$1

trillion

in

annual

spending

by

business

to

use

cloud

services.

Roughly

halfofglobal

cloud

services

is

providedbyU.S.

businesses.12Of

course1allinfrastructuresatscalenecessarilyconsumesignificant

amountsof

energy

tooperate.

Attention

to

thatreality

was

triggered

bya2024

FederalEnergyRegulatoryCommission

(FERC)forecastthat

drew

on

a

range

of

estimatessuggesting

the

U.S.

would

need

between

50GWand

130GW

more

generating

capacitythan

hadbeen

earlier

imagined1

endingthetwo-decade

interregnumof

nearly

flatelectricsector

load

growth.13Forcontext1theannualenergyused

by

a

single

1-GW

data

center

is

more

thantenfoldtheannualenergyusedbycarson11000milesofsuperhighway.It

costs

about$10billion

to

build

both

a1-GW

data

center

and11000

miles

ofhighways.

And

the

cost

ofthe

energy-using

GPUs

inside

the

data

center

is

another$20

billion1essentiallythesameasthe$20

billioncostof

thequantityof

cars

that

11000

miles

of

highway

supports(at

peak).14Thus1thebiggestquestionforthe21stcentury:Areweattheequivalentof

1958in

buildingoutanew

AI-infused“informationsuperhighway”infrastruc-

ture1

or

are

we

approaching

the

equivalent

of

the

year

1992

when

the

last

section

of

the

401000

milesof

interstate

network

wascompleted?1.3The

Spending:Current

ContextU.S.

Business

Sector

Capital

ExpendituresSource:FederalReserveBankof

St.LouisCloudServices:Global

RevenuesTheRiseofAI:ARealityCheckonEnergyandEconomicImpactsSource:Statista7In

the

face

of

disruptions,

analystsand

journalists

seek

illustrative

analogies

hoping

some

are

useful

to

gauge

trends.

The

AI

race

is

often

framed

as

a

ManhattanProjectoramoonshot.15

Botharecategoryerrorson

two

counts;

both

were

anchored

in

government

spending

and

programs,

and

both

involved

single-purpose

goals,

not

infrastructures.Analogizing

theAIbuild-outwith

the

construction

ofthe

U.S.

Interstate

Highway

System

does

illuminate

the

energy

features

of

nation-scale

infrastruc-

tures,

but,alsounlike

AI,thehighwaysystem

wasfinanced

withpublicfunds.

The

privately

financed

continental

railroad

system

has

also

been

offered;spending

tripled

from

the1840s

to

1870s,

peaking

at

over

5%

of

the

nation’s

GDP.16

However,

a

comparisonwithAIwould

be

more

analogous

to

consid-

eringthe

overall

spending

onthegreattransportationtransformation

made

possible

by

the

underlying

technologyof

thecombustionengine.Theemergenceof

theinternetinfrastructureoffersamorerecentanalogy,

though

it

too

would

be

more

relevant

in

terms

of

the

contemporaneous

expansion

of

the

internet

along

with

cellular

networks

and

the

PC.

While

that

build-out

was

associated

with

the

infamous

dot-com

bubble

of1999,

it

isobviousinhindsightthatallthe

enthusiasms—and

the

collateral

need

for

physical

infrastructures—were

all

realized

and

exceeded

in

the

subsequent

years.

That

boom

was

privately

financed,

wherein

the

capital

deployed

reached

1.25%of

GDP—a

level

that

the

AI/cloud

build-out

is

about

to

reach.17Perhaps

the

most

suitable

analogy

is

the

revolutionary

development

of

chemicalscience

atthe

end

of

the

19th

century

and

thesubsequentemergence

ofthechemicalindustryintheearly20thcentury.

Chemical

science

allowed

humanitytomanipulatethebasicbuildingblocksofmoleculesandatomsto

invent

and

fabricate

entirely

new

products

andservices,

from

pharmaceuticals

tothepolymersthatareessentialto

nearly

every

modern

product.

Similarly,

the

knowledge

underlying

the

invention

ofAI,

machine

learning,

and

large

language

modelsenables

manipulationof

basicdata

tocreate—hence

the

new

locution

of“AI

factories”—entirely

new

classes

ofproducts

and

services

for

everysectorof

the

economy.Notably,

over

the

two

decades

prior

to

1929,

private

sector

investment

to

build

what

wasthenrevolutionarychemicalfactoriesandinfrastructuresrosetoover5%of

GDP.18To

match

that

share

of

GDP

by

2030,

AI

spending

would

have

to

reach

nearly

$2

trilliona

year,a

levelin

therange

of

some

forecasts.19Source:GoldmanSachs,“Why

We

Are

Notina

Bubble…Yet,”

Global

StrategyPaperNo

.73,October8,2025ConstructionSpending

byChemical

Industries:1900to19301.4

The

Spending:Historical

ContextStock

MarketConcentration

bySectorSource:EstimateinterpolatedfromavailableCensusbenchmarks

(seeendnote18)TheRiseofAI:ARealityCheckonEnergyandEconomicImpacts8In

his

acceptance

speech

for

the

1987

Nobel

Prize

in

Economic

Sciences1

Robert

Solow

observed

that

“technology

remains

the

dominant

engine

of

growth1

with

human

capital

investment

in

second

place.”20

The

collective

effect

oftechnological

progress—enabled

by

capital

markets

and

facilitated

byinstitutions—yieldedanhistoricjumpinU.S.laborproductivityfollowing

WorldWar

II

which1

in

turn1

producedanenormous

wealthexpansion.Technology

has1

over

various

periods

in

history1

yielded

similar

gains

inproductivity.From1910to19301the

labor-hours

needed

per

car

manufactured

droppednearlyfivefold1asdidthelabor-hourspertonof

steel.21

Therailroad-

era

productivity

gains

in

shipping

were

visible

not

just

from

greater

speed

thanhorse-and-wagontransportbutalsobecauseofa25-foldcollapse

inthe

ton-milecostof

shipping

goods1