The Intelligence Industrial Revolution — Part III
The AI Electricity Crisis: Intelligence Is Becoming an Energy Business
Artificial intelligence feels weightless.
A prompt is typed into a box. A response appears seconds later. No smoke, no engines, no visible machinery. From the user’s perspective, intelligence seems almost abstract — something that lives inside software.
But the physical reality is very different.
Every AI answer requires electricity. Every accelerator converts electricity into heat. Every rack needs cooling. Every data center needs substations, transformers, switchgear, backup power, fiber, permits, land, water or advanced cooling infrastructure, and a connection to an electrical grid capable of delivering enormous amounts of power continuously.
The AI revolution is therefore colliding with something Silicon Valley spent decades largely treating as background infrastructure:
the energy system.
And as AI scales from millions of chatbot queries into billions of agents, autonomous workflows, video generation systems, scientific models, robotics platforms and always-on digital workers, the question is becoming less theoretical.
The next great AI bottleneck may not be a shortage of chips.
It may be a shortage of megawatts.
“Artificial intelligence may be digital, but intelligence at scale is an industrial product made from silicon, water and electricity.”
Intelligence Has a Power Cord
For most of the software era, computing became increasingly abstracted from its physical requirements.
A SaaS founder didn’t need to think much about where the electricity came from. Cloud providers handled servers, networking, cooling and redundancy somewhere else. Software could scale globally without the founder ever seeing a data center.
AI is forcing the physical layer back into view.
The International Energy Agency estimates that global data-center electricity consumption could rise to around 945 terawatt-hours by 2030, roughly double current levels. The IEA expects data-center electricity consumption to grow around 15% annually from 2024 through 2030, more than four times the growth rate of electricity use in most other sectors. AI-optimized facilities are the largest force behind that acceleration.
That is an extraordinary growth rate for infrastructure that sits on top of electrical systems largely designed decades ago.
The IEA’s latest estimates also show electricity generation required to serve data centers increasing from around 460 TWh in 2024 to more than 1,000 TWh by 2030, and approximately 1,300 TWh by 2035 in its base case.
And the trend is already visible. Data-center electricity use reportedly rose about 17% in 2025, dramatically faster than the roughly 3% increase in global electricity demand overall.
This is no longer a hypothetical energy problem that might arrive in 2035.
It is happening now.
The Global Number Can Be Misleading
At first glance, roughly 3% of global electricity consumption by 2030 might not sound catastrophic.
And globally, it isn’t necessarily.
The problem is that electricity isn’t a giant interchangeable pool that can be moved instantly around the world.
AI demand is intensely local.
A hyperscale data center does not care that there is surplus generation thousands of miles away. It needs electricity delivered to one particular parcel of land, at one particular voltage, through one particular transmission and distribution network, continuously and reliably.
That is why aggregate electricity statistics can obscure the real constraint.
Imagine a city with enough water overall, but one neighborhood suddenly wants to build fifty skyscrapers. The city may have plenty of total water, yet the pipes serving that neighborhood cannot deliver enough volume.
AI faces a similar infrastructure problem.
The grid does not simply need:
more electricity.
It needs:
more deliverable electricity at specific places.
That distinction changes everything.
AI Data Centers Are Becoming Industrial Loads
Historically, data centers were significant electricity users, but many were still modest compared with large industrial facilities.
AI changes the density.
A modern AI campus can require hundreds of megawatts. Future clusters may move toward gigawatt-scale facilities.
At that point, the data center stops looking like a large office building filled with servers and starts looking more like an aluminum smelter, steel mill or chemical plant in terms of energy planning.
That matters because industrial-scale electricity consumption creates very different requirements.
A 10-megawatt facility can often connect to existing distribution infrastructure.
A 500-megawatt or 1-gigawatt AI campus may require:
new substations,
new high-voltage transmission,
new generation,
new transformers,
dedicated grid studies,
and years of permitting.
The U.S. Department of Energy’s recent grid planning work cites projections suggesting additional data-center load could add roughly 13 to 55 gigawatts of U.S. demand by 2030, depending on the scenario.
That is not simply a cloud-computing story.
It is a national infrastructure story.
The Hidden AI Bottleneck May Be a Transformer
The most expensive GPU in the world is useless if electricity cannot reach it.
This leads to one of the more surprising realities of the AI boom:
Some of the most important AI infrastructure may be equipment that has existed for more than a century.
Transformers.
Switchgear.
Circuit breakers.
Substations.
Transmission lines.
Power distribution equipment.
Backup generators.
Uninterruptible power supplies.
These aren’t glamorous products.
They don’t generate viral demonstrations.
They don’t write poetry.
Yet without them, the AI factory is dark.
The U.S. Department of Energy has explicitly identified transformers and broader grid infrastructure as strategic concerns as electricity demand rises. The agency’s current “Speed to Power” initiative is designed specifically to accelerate generation and transmission development as AI and other electricity-intensive industries expand.
The bottleneck therefore moves upstream.
The AI industry asks:
“How many GPUs can we deploy?”
The utility asks:
“Where will the power come from?”
The grid operator asks:
“Can the transmission system carry it?”
The substation operator asks:
“Can the transformer handle it?”
The municipality asks:
“Can we permit it?”
Only when every answer is yes does the AI cluster turn on.
The Real AI Stack Is Much Longer Than It Looks
Most people visualize the technology stack like this:
AI Model
↓
GPU
But the physical stack is considerably deeper:
AI Model
↓
GPU / Accelerator
↓
High-Bandwidth Memory
↓
Server
↓
Network
↓
Cooling
↓
UPS
↓
Power Distribution
↓
Transformer
↓
Substation
↓
Transmission Grid
↓
Power Plant
Remove one layer and everything above it becomes irrelevant.
A shortage of chips can delay deployment.
A shortage of power can stop deployment entirely.
That is why the AI investment story is becoming increasingly industrial.
The future of artificial intelligence may depend as much on electrical engineers, utility planners, turbine manufacturers and nuclear developers as on machine-learning researchers.
AI Is Turning Electricity Into Intelligence
There is a useful way to think about what a data center actually does.
A traditional factory takes physical inputs and transforms them into physical outputs.
Steel mills convert ore and energy into steel.
Refineries convert crude oil into usable fuels.
AI data centers convert:
electricity + data + silicon
into:
tokens + predictions + decisions + digital labor.
That makes electricity one of the fundamental raw materials of intelligence.
NVIDIA has popularized the phrase “AI factory” precisely because modern accelerated computing increasingly resembles an industrial production process.
The output is not automobiles or steel.
It is intelligence.
That suggests something profound:
Intelligence itself is becoming an energy-intensive industrial commodity.
And once intelligence becomes an industrial commodity, electricity becomes part of its cost structure.
Cost Per Token Eventually Becomes Cost Per Watt
Today, much of the AI industry talks about:
cost per million tokens,
inference cost,
GPU utilization,
model efficiency.
But underneath those figures sits energy consumption.
Better hardware can lower energy per token.
More efficient models can reduce compute requirements.
Quantization can lower memory demands.
Specialized accelerators can improve inference efficiency.
But if total AI usage rises dramatically, efficiency gains do not necessarily reduce overall electricity consumption.
They may do the opposite.
If AI becomes ten times cheaper, businesses may use one hundred times more AI.
That is the classic Jevons paradox applied to intelligence.
Cheaper inference could lead to:
persistent agents,
real-time AI assistants,
automated scientific research,
continuous video generation,
robotic fleets,
AI running inside every enterprise workflow.
Efficiency may reduce the cost of a unit of intelligence while dramatically increasing total intelligence consumed.
So the energy question isn’t simply:
How efficient will AI become?
It is:
How much intelligence will the world demand once intelligence becomes cheap?
The Agent Economy Could Multiply Power Demand
Chatbots are intermittent.
Humans ask questions periodically.
Agents are different.
An agent might operate continuously.
Imagine a marketing agent that monitors campaigns twenty-four hours a day.
A cybersecurity agent inspecting network activity constantly.
A trading agent analyzing markets continuously.
A coding agent testing software overnight.
A customer-service agent handling thousands of conversations simultaneously.
A robotics-control model making decisions multiple times per second.
The transition from:
human-triggered AI
to:
always-on autonomous AI
could fundamentally alter demand.
Instead of intelligence being used occasionally, intelligence becomes a continuous background service.
That could create an entirely different energy profile.
Millions of agents might never sleep.
AI Video Could Be an Energy Monster
Text generation is only one part of the story.
Generative video is substantially more computationally demanding.
As AI moves toward:
real-time video,
interactive virtual worlds,
simulation,
digital twins,
autonomous driving,
robotics,
scientific modeling,
the amount of computation required for each interaction increases.
That means AI power demand could become increasingly dependent on the mix of workloads.
A world dominated by occasional text queries looks dramatically different from one where billions of devices continuously interpret video, generate media and operate robots.
The second world is much closer to an industrial-scale energy system.
AI Doesn’t Want Cheap Electricity. It Wants Reliable Electricity.
Cost matters enormously.
But reliability may matter even more.
Training a giant AI model involves enormous coordinated clusters running across many thousands of accelerators.
Interruptions can be extremely expensive.
Inference systems serving millions of users require uptime.
Enterprise agents cannot simply disappear because electricity prices spiked.
Robotics systems require continuous operation.
This makes AI customers unusually sensitive to:
reliability,
redundancy,
power quality,
backup generation,
grid stability.
The ideal AI electricity supply therefore isn’t merely:
cheap.
It is:
cheap + abundant + reliable + scalable + increasingly low-carbon.
Few energy sources maximize all five simultaneously.
That is why the AI energy mix will almost certainly remain diverse.
Why Nuclear Is Suddenly Back in the Conversation
For years, nuclear energy occupied an awkward place in technology discussions.
It produces enormous amounts of low-carbon electricity with high capacity factors, but new plants can take years to permit and construct and may face enormous capital costs.
AI changes the equation because hyperscalers increasingly need something nuclear is particularly good at:
large quantities of reliable electricity around the clock.
A solar facility produces electricity when the sun shines.
Wind generation depends on weather.
Batteries can smooth intermittency, but storing huge amounts of electricity for extended periods remains expensive.
Nuclear plants can operate continuously.
For a hyperscaler trying to run a massive AI campus 24/7, that is extremely attractive.
The IEA expects nuclear to play an increasing role in supplying data-center electricity toward the end of the decade and beyond.
That doesn’t mean nuclear suddenly solves everything.
Conventional reactors take time.
Construction costs matter.
Regulation matters.
Fuel supply matters.
Political acceptance matters.
But AI may provide something the nuclear industry has struggled to find for years:
deep-pocketed customers willing to commit to long-term demand.
Small Modular Reactors and AI Campuses
This is where small modular reactors, or SMRs, become interesting.
The idea is appealing:
Instead of constructing a huge centralized reactor designed primarily for a utility, build smaller standardized reactors that can potentially be deployed incrementally near industrial demand.
In theory, a future AI campus could combine several reactors with renewable generation, battery storage and grid connections.
The campus becomes partially self-sufficient.
Whether SMRs achieve the economics promised by advocates remains an open question.
But the incentives have changed.
AI companies have extraordinary:
capital,
electricity demand,
technology expertise,
long planning horizons.
That combination makes them unusually plausible customers for new nuclear technologies.
Natural Gas May Be the Bridge Fuel AI Cannot Avoid
Nuclear attracts headlines.
Renewables attract climate commitments.
But natural gas could be one of the most important near-term beneficiaries of AI electricity demand.
Why?
Speed.
A new nuclear project can take years.
Large transmission projects can take years.
Gas turbines and associated generation can sometimes be deployed much faster.
Natural gas generation can also provide dispatchable power when renewables are unavailable.
The Department of Energy has already noted accelerating demand for natural-gas turbines associated with AI, manufacturing and broader electricity growth.
That doesn’t eliminate environmental concerns.
It highlights the tension.
AI companies may want clean electricity.
But they also want electricity now.
The gap between those objectives could produce substantial gas generation during the transition.
Renewables Still Matter Enormously
The rise of AI does not invalidate solar or wind.
Quite the opposite.
The IEA expects renewables to supply nearly half of the additional electricity required by data centers over the next several years in its base case. Natural gas and coal also contribute, while nuclear becomes more important later.
Renewables have several advantages.
They can often be built faster than conventional power plants.
Costs have fallen dramatically.
Large technology companies already understand renewable power purchase agreements.
Solar generation can be paired with batteries.
Wind can complement solar.
The challenge is matching variable electricity production with an AI campus that wants electricity continuously.
That pushes the opportunity toward integrated systems:
renewables + storage + grid + dispatchable generation.
The winner may not be one energy source.
It may be orchestration.
Batteries Become Part of the AI Stack
Batteries are normally discussed in relation to electric vehicles and renewable energy.
AI could become another enormous demand driver.
Data centers already use battery systems and UPS equipment to maintain power during outages or transitions to backup generators.
But as electricity demand becomes more volatile and data centers interact more actively with the grid, batteries could take on additional roles.
They can:
smooth short-term fluctuations,
provide backup,
reduce peak demand,
arbitrage electricity prices,
support microgrids,
and potentially provide grid services.
A future AI campus might operate batteries not only to protect servers but as part of an intelligent energy-management system.
The AI itself could decide when to:
charge,
discharge,
shift workloads,
purchase grid power,
run generators,
or temporarily reduce compute.
The data center becomes an energy trader.
The Overlooked Problem: AI Loads Can Move Fast
One of the more interesting emerging technical problems is that AI computing may not behave like conventional industrial demand.
Some data centers maintain highly stable loads.
But high-density AI clusters can also experience large and rapid power changes depending on workload scheduling and accelerator utilization.
The U.S. Department of Energy has recently been studying oscillations and electromagnetic-transient behavior associated with very large data centers specifically because these new loads can interact with the grid differently from older facilities.
This means the AI problem isn’t merely:
How many megawatt-hours do we generate?
It may also become:
How quickly can the grid respond?
Power systems have to maintain balance continuously.
Generation must match consumption.
Large abrupt changes can complicate:
frequency management,
voltage regulation,
equipment stress,
backup systems,
battery operation.
That creates demand for something beyond more electricity:
more intelligent electricity.
AI Needs a Responsive Grid
The grid of the future may need to communicate with AI campuses in real time.
Imagine a utility saying:
“We are approaching peak demand.”
The data center responds by:
delaying lower-priority training jobs,
reducing noncritical compute,
discharging batteries,
shifting workloads to another region.
Then electricity prices fall overnight.
The data center increases:
training,
batch inference,
synthetic-data creation,
simulation.
This is computational demand response.
Historically, factories could shift industrial processes based on electricity prices.
AI could do this at unprecedented precision.
Computer workloads are inherently programmable.
That means intelligence production itself can potentially become flexible.
Compute Could Follow Electricity
This leads to a fascinating possibility.
Today electricity travels to data centers.
Tomorrow some AI workloads might increasingly travel to electricity.
Suppose one region has abundant hydroelectric power overnight.
Another has surplus wind.
Another has extremely cheap solar during midday.
Non-time-sensitive AI jobs could migrate geographically.
Training.
Rendering.
Simulation.
Synthetic-data creation.
Large research workloads.
Instead of always moving electrons across thousands of miles of transmission lines, companies might move computation across fiber.
Fiber is extremely good at transporting information.
Transmission lines are expensive and difficult to build.
So perhaps the future AI grid is partly an electrical grid and partly a compute-routing network.
The system asks:
Where is intelligence cheapest to produce right now?
Then sends workloads there.
That’s an entirely new form of energy arbitrage.
The Data Center Could Become a Grid Resource
This reverses the traditional relationship between data centers and utilities.
Today the data center is viewed mainly as a load.
It consumes power.
In the future, a sophisticated campus could become an interactive grid participant.
It might contain:
battery storage,
backup generators,
renewable generation,
microgrid controls,
flexible compute workloads,
possibly nuclear or gas generation.
Then it could increase or decrease grid consumption based on system conditions.
In emergencies, some campuses might temporarily reduce nonessential workloads.
When generation is abundant, they could absorb surplus electricity.
The data center could become a giant programmable load.
And programmable loads can be extremely valuable to electrical systems.
Microgrids May Become AI Infrastructure
This is where microgrids become particularly interesting.
A microgrid combines local:
generation,
storage,
controls,
loads.
It can interact with the broader grid but may also operate partially or independently when necessary.
For AI campuses, the advantages are obvious.
Greater control.
Greater resilience.
Ability to integrate multiple generation sources.
Reduced dependence on constrained utility infrastructure.
Potentially faster deployment.
The Department of Energy has increasingly emphasized the role of microgrids as electricity demand grows and infrastructure becomes more distributed.
A future hyperscale facility may effectively operate like its own miniature utility.
AI Campuses Could Become Private Power Systems
Push this idea further.
Imagine a 2032 hyperscale AI campus containing:
2 GW of computing capacity,
dedicated gas turbines,
advanced nuclear reactors,
large solar arrays,
grid-scale batteries,
private high-voltage substations,
water recycling,
liquid cooling,
fiber connections to other AI campuses.
At that point, is this still simply a data center?
Or is it an industrial city built to manufacture intelligence?
The boundary begins to disappear.
The technology company owns:
compute,
power,
storage,
networking,
land,
possibly generation.
That resembles an industrial conglomerate far more than traditional software.
The Rise of the AI Utility
One possible future business model is the AI utility.
A company controls:
electricity supply,
data centers,
accelerators,
networking,
foundation models.
Customers simply purchase intelligence.
Instead of buying kilowatt-hours, they buy:
reasoning,
inference,
agent-hours,
robot control,
scientific computation.
Underneath the service, energy and intelligence become tightly integrated.
The cost of electricity becomes part of the price of cognition.
This could produce completely new financial metrics:
tokens per watt,
revenue per megawatt,
intelligence output per kilowatt-hour,
agent productivity per megawatt.
Energy efficiency stops being merely a climate objective.
It becomes a competitive advantage.
The New Metric: Intelligence per Watt
For years, AI labs competed on model quality.
Infrastructure companies competed on FLOPS.
Energy scarcity may make another metric increasingly important:
useful intelligence per watt.
If Model A delivers similar performance to Model B using half the electricity, that difference becomes enormously valuable at global scale.
If one accelerator architecture performs the same inference using 40% less electricity, that’s not simply an engineering achievement.
It directly affects:
operating costs,
data-center capacity,
grid requirements,
deployment speed.
Energy efficiency becomes strategic.
That might reward:
better chips,
better cooling,
better model architectures,
quantization,
sparsity,
smaller specialized models,
improved software.
In other words, the energy bottleneck may accelerate AI efficiency innovation.
The Water Question
Electricity isn’t the only resource.
Heat must be removed.
Some data-center cooling systems consume significant water.
That creates tension in regions already experiencing water stress.
The ideal AI site therefore depends not just on:
electricity cost,
but:
water availability,
climate,
cooling design,
local regulation.
Liquid cooling may reduce some constraints while introducing others.
Closed-loop systems can reduce water consumption.
Immersion cooling may change thermal economics.
Heat reuse could potentially redirect waste heat into:
district heating,
industrial processes,
greenhouses.
The physical design of data centers is becoming a surprisingly important technological frontier.
Heat Is the Waste Product of Intelligence
There is almost something philosophical about this.
Artificial intelligence appears to produce:
ideas,
images,
software,
predictions.
Physically, it also produces:
heat.
Every calculation eventually dissipates energy.
At sufficient scale, cooling becomes one of the defining engineering problems of artificial intelligence.
This is why liquid-cooling companies and thermal-management suppliers belong in the AI conversation.
The highest-performance AI chips cannot continue increasing power density indefinitely without corresponding improvements in heat removal.
AI performance therefore becomes partially constrained by thermodynamics.
You cannot prompt-engineer your way around physics.
Data-Center Geography Is About to Change
Historically, major cloud regions clustered around:
fiber connectivity,
customers,
tax incentives,
cheap land.
Energy increasingly becomes another dominant variable.
Future AI infrastructure may migrate toward places offering combinations of:
cheap electricity,
available grid capacity,
political support,
cool climates,
water,
renewable resources,
nuclear generation,
natural gas.
This could create entirely new economic geographies.
Regions once considered peripheral to technology could become strategic AI hubs because they possess abundant power.
The next Silicon Valley might not be defined by venture capital.
It might be defined by megawatts.
Energy Rich Could Become AI Rich
Countries with abundant reliable electricity may gain a new competitive advantage.
Historically, cheap energy attracted:
aluminum,
steel,
chemicals,
manufacturing.
Tomorrow it may attract:
AI.
That raises geopolitical implications.
Nations with:
hydroelectric power,
nuclear fleets,
natural gas,
solar resources,
geothermal potential,
strong grids
could become exporters of intelligence.
Perhaps future countries won’t export electricity directly.
They will export AI computation produced using electricity.
Energy becomes embedded inside digital output.
AI Could Reshape Energy Policy
Governments already worry about:
grid reliability,
industrial competitiveness,
electrification,
energy security.
AI adds another strategic priority.
If advanced AI becomes economically and militarily important, nations may view electricity infrastructure as part of national AI capability.
The U.S. Department of Energy’s 2026 strategic planning explicitly ties transmission and electricity-system expansion to supporting large AI data centers and other emerging loads.
That means AI policy can no longer remain confined to:
chips,
models,
data.
It increasingly includes:
generation,
transmission,
permitting,
natural gas,
nuclear,
renewables,
storage.
AI policy becomes energy policy.
The AI Arms Race May Become a Power-Plant Race
Suppose two countries possess similar AI researchers and access to similar chips.
Country A can deploy 50 gigawatts of new reliable electricity quickly.
Country B struggles to connect five gigawatts.
Who scales faster?
Compute leadership becomes partially determined by energy infrastructure.
That means national AI competition may increasingly involve:
reactor construction,
transmission expansion,
turbine manufacturing,
transformer supply,
energy permitting.
These aren’t traditionally considered high-tech geopolitical assets.
AI could make them strategic.
The Opportunity for Energy Companies
This transition creates extraordinary opportunities.
Utilities may see decades of previously stagnant load growth reverse.
Independent power producers may gain new customers.
Natural-gas companies may benefit from dispatchable-generation demand.
Nuclear companies gain hyperscalers as potential counterparties.
Renewable developers gain large creditworthy buyers.
Battery companies gain massive stationary-storage markets.
Grid-equipment suppliers benefit from expansion.
Cooling companies benefit from rack density.
Software providers benefit from energy optimization.
The AI boom begins spilling out across the industrial economy.
This is precisely what makes the AI story so much larger than semiconductors.
The New Picks and Shovels
During a gold rush, selling picks and shovels can be safer than searching for gold.
During the AI buildout, the picks and shovels increasingly include:
transformers,
gas turbines,
switchgear,
cooling systems,
batteries,
substations,
cables,
pumps,
fiber,
power electronics.
Some of the most profitable beneficiaries may look almost boring beside frontier AI laboratories.
But boring infrastructure often possesses something glamorous startups lack:
scarcity.
You can launch another AI application tomorrow.
You cannot manufacture a high-voltage transformer tomorrow.
You cannot permit a nuclear plant tomorrow.
You cannot build a transmission line overnight.
Scarcity creates pricing power.
AI’s Real Constraint May Be Time
This is another underappreciated point.
Capital can be raised quickly.
Software can be written quickly.
GPU orders can be placed quickly.
Physical infrastructure moves slowly.
Transformers have manufacturing schedules.
Utilities conduct interconnection studies.
Transmission lines need permits.
Power plants require construction.
Nuclear facilities require regulatory approval.
The conflict between:
AI speed
and:
infrastructure speed
may become one of the defining tensions of the next decade.
Silicon Valley operates in months.
Electric grids operate in years.
That mismatch is dangerous.
It also creates enormous opportunity for anyone capable of shortening infrastructure deployment times.
What If We Overbuild?
There is, however, another side.
The AI electricity boom could overshoot.
What if:
models become dramatically more efficient?
specialized chips reduce energy requirements?
inference costs collapse?
AI demand underwhelms?
capital spending slows?
Then some dedicated generation assets and data-center campuses could become overbuilt.
This is the same risk discussed in Part I.
A technology can be revolutionary while infrastructure investors still overpay.
Railroads changed the world and went bankrupt.
Fiber changed the internet and destroyed enormous capital.
AI power infrastructure could follow a similar cycle.
The critical question isn’t whether AI needs electricity.
It clearly does.
The question is:
How much, where, when and at what price?
Not Every Megawatt Is Equal
This makes location increasingly important.
One megawatt available today near existing fiber and data-center infrastructure may be worth more than a theoretically cheap megawatt requiring five years of transmission development.
Power markets may therefore begin assigning significant premiums to:
speed-to-power,
grid-ready sites,
existing substations,
firm interconnections.
Developers could increasingly value land based on electrical access rather than acreage.
In the AI era, the phrase:
“location, location, location”
may evolve into:
“power, power, power.”
The Carbon Contradiction
AI also faces an uncomfortable environmental contradiction.
Artificial intelligence could help:
optimize grids,
design better batteries,
model climate systems,
improve energy efficiency,
discover new materials.
But building and operating AI itself consumes significant electricity.
If additional demand is supplied primarily by fossil fuels, emissions rise.
If it comes from low-carbon generation, environmental costs fall.
So AI’s carbon impact depends heavily on the marginal generation used to support it.
The IEA expects a mixed answer: renewables supply much of the growth, while natural gas, coal and nuclear also contribute.
There is no single global AI electricity story.
Texas looks different from France.
Virginia looks different from Iceland.
China looks different from California.
Energy is fundamentally regional.
AI Could Actually Help Solve the Grid Problem It Creates
There is a nice irony here.
AI creates new stress on electricity systems.
AI may also improve them.
Advanced systems could help utilities:
forecast demand,
predict equipment failures,
optimize transmission,
manage storage,
detect faults,
integrate renewable generation,
schedule maintenance,
respond to extreme weather.
The IEA specifically notes that AI has substantial potential to improve energy-system efficiency and operations even as AI infrastructure increases electricity demand.
So the relationship becomes circular:
AI increases electricity demand.
Then:
AI helps optimize electricity supply.
The intelligence industry and energy industry begin evolving together.
Technology Companies May Become Energy Companies
This leads to perhaps the most important structural prediction.
For the last twenty years, technology companies mostly purchased electricity.
For the next twenty years, some may increasingly participate in creating it.
Hyperscalers have:
huge balance sheets,
long-term electricity demand,
excellent credit,
sophisticated engineering teams,
strong incentives to reduce energy constraints.
Those characteristics are unusually compatible with energy development.
They can sign long-term contracts.
Finance generation projects.
Support new reactor designs.
Build batteries.
Develop private transmission.
Locate facilities next to generation.
Potentially own generation outright.
The boundary between technology company and energy company begins to blur.
What an AI Hyperscaler Could Look Like in 2035
Imagine a future hyperscaler.
It owns:
AI models,
cloud infrastructure,
semiconductor designs,
robotics platforms,
data centers.
But it also finances:
small modular nuclear reactors,
natural-gas turbines,
solar farms,
battery storage,
high-voltage transmission,
private microgrids.
Its competitive advantage isn’t merely software.
It’s vertically integrated intelligence production.
That starts to look surprisingly similar to:
Standard Oil,
General Electric,
Samsung,
or another industrial conglomerate.
AI may push Big Tech toward vertical integration on a scale we haven’t seen in decades.
The Intelligence Refinery
Perhaps “data center” eventually becomes the wrong phrase.
A more accurate concept might be:
intelligence refinery.
Electricity enters.
Data enters.
Hardware transforms those inputs.
Intelligence exits.
Some of that intelligence becomes:
software.
Some becomes:
advertising.
Some becomes:
scientific discovery.
Some becomes:
robotic actions.
Some becomes:
financial decisions.
The data center becomes an industrial facility producing one of the most valuable commodities in the modern economy:
usable cognition.
That framing may help us understand why energy matters so much.
Refineries need oil.
Steel mills need electricity and ore.
AI factories need:
electricity + silicon + information.
The AI Energy Toll Booth
This connects directly back to the Toll Booth Theory introduced in Part I.
Suppose OpenAI wins.
It needs electricity.
Anthropic wins?
Electricity.
Google?
Electricity.
Meta?
Electricity.
Open source?
Electricity.
Agents?
More electricity.
Robots?
More electricity.
Video generation?
More electricity.
Artificial intelligence can change architectures, companies and models while still requiring power.
That makes energy one of the deepest toll booths in the entire AI economy.
You don’t need to know which AI model wins to know the machines must remain plugged in.
But Electricity Alone Isn’t the Moat
There is an important nuance.
Electricity is a commodity.
What may become scarce is:
electricity in the right place at the right time with the right infrastructure attached.
That means the valuable asset could be:
a grid interconnection,
a substation,
a transformer allocation,
a permitted data-center site,
a turbine order,
a nuclear power agreement.
This is why the AI energy story cannot be reduced simply to:
“Buy utilities.”
It is about bottlenecks across an interconnected physical system.
The New AI Energy Stack
The investable ecosystem increasingly looks like:
Generation
nuclear, gas, renewables, geothermal
Storage
batteries, long-duration storage
Grid
transmission, transformers, switchgear, substations
On-Site Power
microgrids, generators, fuel cells
Power Electronics
UPS, converters, distribution
Thermal Management
liquid cooling, pumps, heat exchangers
Software
grid optimization, energy management, workload scheduling
Data Centers
land, construction, operations
Compute
accelerators, memory, networking
This is the physical foundation of artificial intelligence.
The Ultimate Question: How Much Energy Does Intelligence Deserve?
There is also a philosophical question.
As AI consumes an increasing share of electricity, societies may eventually have to debate competing priorities.
Do we allocate scarce grid capacity to:
housing?
factories?
EV charging?
hospitals?
AI data centers?
Cryptocurrency?
If electricity becomes constrained, the economic value produced by different loads may come under scrutiny.
A data center producing life-saving scientific discoveries is easy to defend.
A facility generating billions of novelty videos might be viewed differently.
Utilities traditionally don’t judge the philosophical value of electricity consumption.
AI growth could make such debates politically unavoidable.
Intelligence Becomes Part of Energy Economics
For decades, economic growth required:
labor,
capital,
energy.
AI introduces something new:
machine intelligence.
But machine intelligence itself requires capital and energy.
So instead of replacing the old economy, AI folds back into it.
The intelligence economy may ultimately rest on very traditional foundations:
mines,
factories,
transmission towers,
power plants,
water systems,
industrial equipment.
This is the paradox of AI.
The more advanced the digital world becomes, the more dependent it becomes on physical infrastructure.
The Biggest AI Company May Need an Energy Division
Today an AI company’s org chart includes:
research,
engineering,
product,
sales.
Tomorrow it may include:
power procurement,
grid engineering,
nuclear partnerships,
energy trading,
microgrid operations.
That sounds bizarre until you remember that energy could determine whether billions of dollars of compute equipment can actually operate.
At sufficiently large scale, electricity stops being a utility bill.
It becomes strategy.
From Artificial Intelligence to Industrial Intelligence
Part I of this series argued that AI is becoming an industrial economy.
Part II argued that memory is becoming a central bottleneck.
Part III adds another layer.
Intelligence needs power.
Not metaphorical power.
Electrical power.
Enough of it to run millions of accelerators.
Stable enough to keep them operating.
Cheap enough to make inference economically viable.
Available quickly enough to keep pace with AI demand.
That puts artificial intelligence on a collision course with an energy industry that operates on completely different timelines.
The Final Inversion
For decades, technology companies built software on top of infrastructure supplied by someone else.
Electricity came from utilities.
Buildings came from landlords.
Networks came from telecom companies.
Cloud abstracted away the physical world.
AI reverses that abstraction.
At extreme scale, the physical infrastructure becomes too strategically important to ignore.
So perhaps the defining transformation isn’t merely:
software becomes intelligent.
It is:
technology becomes industrial again.
Google, Microsoft, Amazon, Meta and the next generation of AI infrastructure companies may increasingly finance or shape:
nuclear,
gas generation,
renewables,
battery systems,
transmission,
microgrids,
cooling,
industrial real estate.
And at that point the term “technology company” begins to feel incomplete.
They become builders of the infrastructure that manufactures intelligence.
The AI Electricity Crisis
Artificial intelligence sounds virtual.
Its constraints are not.
Every model depends on a chip.
Every chip depends on memory.
Every server depends on cooling.
Every rack depends on power distribution.
Every data center depends on a substation.
Every substation depends on a grid.
Every grid ultimately depends on generation.
There is no artificial intelligence without physical infrastructure.
There is no cloud without concrete.
There is no inference without electricity.
“The next AI bottleneck may not be a shortage of intelligence. It may be a shortage of megawatts.”
The world is rapidly learning that intelligence at scale is not simply software.
It is an industrial system.
And if artificial intelligence becomes one of the defining industries of the twenty-first century, then electricity may become one of its defining strategic resources.
The future AI race could ultimately be decided not only by:
who has the best researchers,
who has the best chips,
or who has the best model,
but by something far more fundamental:
Who can keep the machines powered?
The Intelligence Industrial Revolution
Part I — Money: The $1 Trillion AI Question: Who Actually Gets Paid?
Part II — Memory: AI Has a Memory Problem
Part III — Energy: The AI Electricity Crisis: Intelligence Is Becoming an Energy Business
Part IV — Economics: 50 Winners and Losers From the AI Revolution
Part V — Civilization: The AI Apocalypse Has Five Versions
The thesis connecting all three installments so far is becoming increasingly clear:
AI began as a software story. It became a semiconductor story. Now it is becoming an infrastructure story—and infrastructure ultimately means capital, memory, land, cooling and power.
For Edaptus, I’d make Part IV especially strong by going beyond a simple list of jobs. The more interesting story is where economic scarcity moves when intelligence becomes cheap: judgment, trust, physical execution, proprietary data, relationships, energy, ownership and distribution.
Leave a Reply
You must be logged in to post a comment.