AI Is Running Out of Power — And the Next Chip Shortage May Not Be About Chips at All
For the past several years, one bottleneck has dominated the artificial intelligence boom:
chips.
Companies could not get enough Nvidia GPUs.
Cloud providers raced to secure accelerators.
Startups waited months for computing capacity.
Governments worried about semiconductor supply chains.
But the next constraint on AI may be something much older and much less glamorous.
Electricity.
Not because the world is literally running out of power.
The problem is more specific — and potentially more difficult to solve.
AI companies are building enormous data centers faster than power grids can connect and supply them.
And that means the next phase of the AI boom may depend less on who can manufacture the most advanced chip and more on who can actually find enough electricity to turn those chips on.
That changes the AI race completely.
The AI Boom Has Reached the Power Grid
An AI chip sitting in a warehouse has enormous theoretical computing power.
It also has almost no economic value until someone installs it inside a data center, connects that facility to a grid, cools the equipment and supplies electricity continuously.
That last part is becoming harder.
Morgan Stanley recently estimated that U.S. data-center developers could face a 32-gigawatt net power shortfall through 2028, equivalent to roughly 34% of projected requirements even after accounting for measures such as on-site power generation and fuel cells.
Thirty-two gigawatts is not a minor infrastructure mismatch.
It is roughly the output of dozens of large power plants.
And unlike ordering additional servers, building the infrastructure necessary to deliver that electricity can take years.
AI Moves Faster Than the Electricity Grid
This is the fundamental problem.
A technology company can order thousands of GPUs relatively quickly.
A power utility cannot necessarily build:
new transmission lines,
substations,
gas turbines,
nuclear plants,
or grid connections
on the same timeline.
A large data center might be designed and constructed within a few years.
Major transmission projects can take considerably longer.
That creates a strange situation.
Companies may have:
the land,
the building,
the servers,
the chips,
and the customers
but still be waiting for electricity.
The New Question Is No Longer “Can We Get the GPUs?”
It is becoming:
“Can we power them?”
That sounds almost absurd considering how much attention the semiconductor shortage received.
But computing infrastructure has always depended on physical infrastructure.
The AI boom simply reached a scale where that dependency can no longer be ignored.
For years, electricity was treated almost like an unlimited background resource.
Plug in another server.
Add another rack.
Expand another cloud region.
At AI scale, that assumption stops working.
AI Servers Consume Much More Power Than Traditional Computing
Modern AI infrastructure is unusually power-intensive.
The latest accelerated computing systems pack enormous amounts of processing power into relatively small spaces.
That means higher power density.
Higher cooling requirements.
And much larger electricity loads.
The International Energy Agency estimates that data centers consumed roughly 415 terawatt-hours of electricity globally in 2024, around 1.5% of global electricity use.
Their consumption had already been growing at roughly 12% annually over the previous five years.
AI is accelerating that trend.
By 2030, electricity generation required to supply global data centers could exceed 1,000 terawatt-hours.
That is more than double today's level.
This Isn't Just About ChatGPT
When people hear “AI electricity demand,” they often imagine millions of people asking chatbots questions.
That is only one part of the picture.
Modern AI infrastructure supports:
model training,
AI search,
image generation,
video generation,
coding assistants,
recommendation systems,
enterprise automation,
autonomous agents,
scientific computing,
advertising,
and increasingly robotics.
Each new AI service increases demand for inference — the process of actually running models after they have been trained.
And inference is persistent.
A giant training run eventually ends.
Millions of users asking AI systems to do things every day do not.
Inference Could Become the Bigger Electricity Problem
Early AI infrastructure discussions focused heavily on training.
Training frontier models requires enormous clusters of GPUs.
But as AI becomes embedded into everyday products, inference demand can grow continuously.
Imagine AI operating inside:
search engines,
email,
office software,
shopping,
phones,
cars,
customer service,
healthcare systems,
and financial platforms.
Each individual request may not consume an extraordinary amount of electricity.
Multiply it by billions of interactions.
Then repeat every day.
That becomes infrastructure.
This Is Why Data Centers Are Getting So Large
The newest AI data-center campuses are not simply buildings filled with computers.
They increasingly resemble industrial projects.
They require:
huge electrical connections,
cooling systems,
backup generation,
water infrastructure,
fiber connectivity,
power substations,
and sometimes dedicated generation.
Some planned campuses are measured in gigawatts, not megawatts.
At that scale, a data center is no longer simply another corporate building.
It begins to look like a major industrial consumer.
And America's Grid Wasn't Built for This
Much of the U.S. electricity grid was designed around relatively predictable demand growth.
Factories.
Homes.
Commercial buildings.
Air conditioning.
Population growth.
Then data centers arrived with something unusual:
enormous blocks of electricity demand appearing quickly in concentrated locations.
That makes planning difficult.
A utility might expect gradual demand growth.
Then a hyperscaler proposes a campus requiring hundreds of megawatts.
Or multiple campuses appear in the same region.
The grid suddenly has to catch up.
Virginia Is a Preview of What Could Happen Elsewhere
Northern Virginia became one of the world's biggest data-center hubs because it had:
fiber connectivity,
available land,
business-friendly policies,
and access to major internet infrastructure.
But the concentration also created pressure on the electricity network.
Similar issues are now appearing elsewhere as AI data centers expand across the United States.
The lesson is simple:
the ideal place for a data center is not necessarily the place where enough spare electricity exists.
PJM Is Already Facing the Problem
PJM Interconnection operates the largest regional electricity market in the United States, covering parts of the Mid-Atlantic and Midwest.
It has been dealing with rapidly rising demand forecasts linked partly to data-center growth.
Regulators recently addressed a projected 6,800-megawatt power shortfall in the region.
That number is important because it shows this is no longer a theoretical forecast about 2035.
Grid planners are already trying to decide:
who builds the required generation,
who pays for it,
and whether existing households should bear some of the cost.
Those questions are going to become increasingly controversial.
The AI Boom Could Eventually Show Up on Your Electricity Bill
This is where the story moves beyond Silicon Valley.
If a data center requires billions of dollars of:
transmission,
generation,
substations,
and grid upgrades,
someone has to pay.
The tech company?
The utility?
Government?
Existing electricity customers?
Regulators are increasingly asking whether giant new data-center customers should pay a larger share of the infrastructure required to serve them.
That matters.
Because households will understandably object if AI companies consume enormous amounts of electricity while ordinary customers see higher bills.
This Could Become a Political Problem
The AI infrastructure boom was initially framed mostly around:
innovation,
jobs,
national competitiveness,
and technology leadership.
But enormous data centers affect local communities.
They can require:
land,
electricity,
water,
transmission lines,
and new generation.
Residents may ask:
Why is my community supporting infrastructure for a trillion-dollar technology company?
Will my electricity bill rise?
Will the project create enough local jobs?
Who gets the economic benefit?
Those questions are already becoming part of data-center permitting debates.
The Biggest AI Companies Know This
Look at what technology companies are doing.
Microsoft.
Amazon.
Google.
Meta.
OpenAI and its infrastructure partners.
They are no longer thinking only about computers.
They are increasingly thinking about:
power contracts,
nuclear energy,
renewable projects,
natural gas,
fuel cells,
and dedicated generation.
That tells you something important.
Energy is becoming part of the AI stack.
The AI Stack Used to Look Like This
At a simplified level:
models
↓
software
↓
GPUs
↓
data centers
Now we need to add another layer.
models
↓
software
↓
GPUs
↓
data centers
↓
electricity
↓
the grid
The bottom layers may determine how fast everything above them can grow.
This Is Why Nuclear Power Suddenly Looks Interesting Again
AI companies need something unusual.
Large quantities of electricity.
Available continuously.
Preferably low-carbon.
That makes nuclear power attractive.
Nuclear plants produce enormous amounts of steady electricity without the intermittency associated with wind and solar.
The U.S. government is now supporting efforts to increase nuclear output, while major technology companies have signed agreements connected with existing and future nuclear generation.
AI did not single-handedly cause the nuclear revival.
But it significantly strengthened the economic case.
Restarting Old Nuclear Plants Was Almost Unthinkable a Few Years Ago
Now it is being seriously considered.
Why?
Because a functioning nuclear facility near a data-center region suddenly becomes extremely valuable.
An AI boom that requires constant gigawatt-scale electricity demand changes the economics of existing energy assets.
Infrastructure that once looked outdated can become strategically important again.
Natural Gas May Benefit Too
There is an uncomfortable reality here.
Renewables are expanding rapidly.
But AI companies need electricity now.
Solar and wind can be built relatively quickly, but their output varies.
Large data centers operate continuously.
That means reliable generation still matters.
Natural gas plants can provide dispatchable power much faster than new nuclear plants.
As a result, the AI boom could create additional gas demand even as technology companies maintain ambitious climate targets.
This is one of the contradictions of the AI transition.
Renewable Energy Alone Doesn't Solve the Problem
Imagine a data center with enormous solar contracts.
Great.
But what happens at night?
Or during periods of low renewable output?
Batteries can help.
Transmission can connect different regions.
Other generation can fill the gap.
But the challenge is not simply generating the same number of kilowatt-hours over a year.
The data center needs reliable electricity every second.
That requires an entire system.
The Grid Is the Forgotten AI Technology
People talk endlessly about:
transformers,
large language models,
GPU architecture,
memory bandwidth,
and chip fabrication.
But literally powering the AI revolution requires another kind of transformer.
The electrical kind.
And those are not infinitely available either.
Substations.
Transformers.
Switchgear.
Transmission equipment.
Generators.
Cooling systems.
Electrical engineering is becoming part of the AI investment story.
This Is Why Companies Like Schneider Electric Became AI Winners
One of the less obvious beneficiaries of AI has been the electrical equipment industry.
Data centers need far more than servers.
They need systems that:
distribute electricity,
manage power,
protect equipment,
cool racks,
monitor energy use,
and keep everything operating reliably.
Companies supplying that infrastructure can benefit even if they never manufacture a single AI chip.
This broadens the AI economy enormously.
Nvidia Still Sits at the Center — But It Doesn't Control the Grid
Nvidia can design faster GPUs.
TSMC can manufacture more advanced chips.
Memory companies can produce more HBM.
Cloud companies can build more facilities.
None of them can instantly create gigawatts of grid capacity.
That is why electricity represents a fundamentally different kind of bottleneck.
Semiconductor capacity can be expanded through enormous investment.
Power infrastructure also requires:
permits,
public approval,
land,
regulators,
utilities,
and long construction timelines.
Money helps.
It does not remove every constraint.
Interestingly, Nvidia May Be Better Protected Than Some Suppliers
Morgan Stanley argues that Nvidia and Broadcom may be relatively insulated from the immediate effects of delayed AI data-center projects.
Why?
Their most important chips are often ordered early.
Customers may secure GPUs long before an entire facility is fully ready.
Those chips could potentially be redirected to other projects.
But suppliers deeper in the infrastructure chain may face different risks.
Memory and Optical Suppliers Could Feel Delays More Quickly
An AI server contains much more than a GPU.
It needs:
high-bandwidth memory,
networking,
optical components,
storage,
power-management hardware,
and other electronics.
If data-center deployment slows because power connections are delayed, demand for some of those surrounding components could also shift.
That means the power shortage can propagate backward through the semiconductor supply chain.
The AI boom is interconnected.
A problem at the grid connection can eventually affect a memory supplier thousands of miles away.
This Is the Part Investors Should Understand
“AI demand is strong” does not automatically mean every AI-related company benefits equally.
Imagine that technology companies want 100 new data centers.
Power is available for only 70 on schedule.
The highest-value components may still sell.
But some infrastructure orders could be postponed.
That creates winners and losers even inside an overwhelmingly strong AI market.
The bottleneck determines where revenue actually flows.
Yesterday's Bottleneck Was GPUs
Then came:
HBM memory.
Advanced packaging.
Networking.
Cooling.
Now power.
The pattern is important.
AI infrastructure is expanding so rapidly that it keeps discovering the next constraint.
Solve one bottleneck and another appears.
That is normal in massive industrial buildouts.
AI Is Becoming an Industrial Revolution, Not Just a Software Boom
This may be the bigger lesson.
The first generative-AI wave looked like software.
ChatGPT opened in a browser.
Midjourney generated images.
Coding assistants appeared inside development tools.
It felt digital.
But underneath that software sits an enormous physical economy.
Semiconductor fabs.
Data centers.
Power plants.
Transmission lines.
Cooling equipment.
Fiber networks.
Construction.
Land.
Water.
AI is increasingly behaving like industrial infrastructure.
That Makes the AI Boom Much More Expensive
Buying GPUs is only the beginning.
A fully operational AI campus can require billions of dollars of infrastructure beyond the chips themselves.
That helps explain why projected AI investment figures have become so enormous.
Companies are not simply buying computers.
They are effectively constructing a new layer of industrial infrastructure.
It Also Explains the $3 Trillion Question
We previously looked at how enormous amounts of capital are being poured into AI infrastructure.
The power crunch explains where some of the next wave may go.
Not just into:
Nvidia GPUs,
but into:
power generation,
grid equipment,
transformers,
cooling,
nuclear,
gas,
renewables,
storage,
transmission,
and data-center construction.
The AI trade is becoming an energy trade.
Electricity Could Determine Where AI Gets Built
Suppose two states both want a giant AI data center.
One offers tax incentives.
The other offers abundant reliable electricity.
Which matters more?
Increasingly, power availability may win.
That could shift data-center investment toward regions with:
surplus generation,
faster grid connections,
available natural gas,
nuclear capacity,
or strong renewable resources.
Electricity becomes an economic-development advantage.
Countries Could Face the Same Competition
This applies internationally too.
AI sovereignty is often discussed in terms of:
chips,
models,
data,
and cloud infrastructure.
But countries trying to build domestic AI capacity also need electricity.
A nation may gain access to advanced GPUs and still struggle to deploy them at scale if its grid cannot support giant data centers.
That gives energy-rich regions a new strategic advantage.
The Middle East Understands This
Countries with abundant energy resources are increasingly investing heavily in AI infrastructure.
Their advantage is not only capital.
It is energy.
If AI computing becomes one of the world's most electricity-intensive industries, access to cheap reliable power becomes strategically valuable.
The geography of technology could shift as a result.
AI Could Become an Energy Arbitrage Business
Imagine the cost of computing the same AI workload in two regions.
Same GPUs.
Same software.
But electricity costs:
$30 per megawatt-hour in one region
and
$90 in another.
At massive scale, the difference becomes enormous.
That encourages companies to move computational workloads toward cheaper power where latency and regulation allow.
In other words:
future AI infrastructure may follow electricity the way heavy industry historically followed coal, hydroelectric power or natural gas.
That Would Be a Full-Circle Moment
The digital economy was supposed to reduce the importance of geography.
Software could exist anywhere.
Cloud computing made physical location almost invisible to users.
AI is reminding us that the cloud was always physical.
Behind every AI answer is:
a computer,
inside a building,
connected to a grid,
somewhere on Earth.
The physical world never disappeared.
We simply stopped noticing it.
Could Better Chips Solve the Problem?
Partly.
More efficient AI accelerators can perform more calculations per unit of energy.
Better cooling improves efficiency.
Better model architecture can reduce computational requirements.
Smaller models can handle many tasks without using frontier-scale systems.
All of that matters.
Efficiency may become one of the most important competitive metrics in AI.
But there is a problem.
Efficiency Does Not Necessarily Reduce Total Electricity Use
This is sometimes called the rebound effect.
Suppose a new GPU performs twice as much AI work using the same electricity.
That sounds like electricity demand should fall.
But if cheaper computing causes companies to run five times more AI workloads, total demand still rises.
AI may become more efficient per task while consuming more electricity overall.
We have seen versions of this phenomenon throughout technology history.
The Better AI Gets, the More We May Use It
Consider search.
If every Google search becomes AI-generated, computing demand rises.
Now imagine:
every email has AI,
every spreadsheet has AI,
every phone has AI,
every customer-service interaction has AI,
every car has AI,
and millions of software agents work continuously in the background.
Efficiency improves.
Usage explodes.
That is why electricity forecasts remain so large.
The Real Race May Become Compute Per Watt
For years AI companies competed on:
model size,
benchmark scores,
GPU count,
and training compute.
Power constraints could change the priority.
The winning systems may increasingly be those that achieve more intelligence with less electricity.
That favors:
better chips,
better cooling,
better models,
better software,
and more specialized hardware.
Energy efficiency becomes a competitive advantage.
Apple May Have Been Early to This Idea
Consumer-device companies have spent years optimizing performance per watt because battery life matters.
Data-center AI had less pressure.
Plug into the grid and scale.
That mindset is changing.
At gigawatt scale, even small efficiency improvements can save enormous amounts of money and infrastructure.
The data center is beginning to care about energy efficiency the way a smartphone does.
This Could Accelerate Smaller AI Models
Not every task requires the largest model available.
A customer-service bot.
A summarization tool.
A basic coding assistant.
A recommendation engine.
Many tasks can run on smaller specialized models.
If electricity and GPU capacity remain constrained, companies will have stronger incentives to match the model to the task.
That could make the AI ecosystem more diverse.
Bigger will not always mean better.
AI Could Also Move Closer to the User
Another possibility is edge AI.
Instead of sending every request to a massive data center, some AI processing can happen on:
phones,
PCs,
cars,
factory equipment,
and local servers.
That distributes electricity demand.
It also reduces latency and potentially improves privacy.
The cloud will remain essential.
But a power-constrained AI future makes local computing more attractive.
Then There Is Cooling
Electricity enters the server.
Most of it eventually becomes heat.
That heat has to go somewhere.
Traditional air cooling becomes less practical as rack power density rises.
That is why liquid cooling has become increasingly important for advanced AI systems.
Again, solve one problem and another infrastructure market grows.
GPUs create electricity demand.
Electricity creates heat.
Heat creates cooling demand.
AI spending spreads through the entire industrial supply chain.
Water Could Become the Next Controversy
Some data-center cooling systems use substantial amounts of water.
That creates tension in regions already facing water scarcity.
Communities may increasingly evaluate data-center projects based on both:
electricity consumption
and
water consumption.
The environmental footprint of AI infrastructure is therefore not a single issue.
Power.
Water.
Land.
Emissions.
Transmission.
They are connected.
Could Data Centers Simply Build Their Own Power Plants?
Increasingly, yes.
Or at least some form of dedicated generation.
“Behind-the-meter” generation allows large facilities to produce part or all of their electricity without relying completely on the existing grid.
Options can include:
natural gas,
fuel cells,
renewables,
batteries,
and eventually small modular nuclear reactors.
This could dramatically change the relationship between technology companies and energy companies.
A Future Data Center Could Look Like a Small Utility
Imagine an AI campus with:
its own gas turbines,
solar arrays,
battery storage,
fuel cells,
substation,
backup generators,
and dedicated transmission.
At some point, describing it as simply a “data center” almost understates what it is.
It becomes an energy-and-compute complex.
That may be where hyperscale infrastructure is heading.
This Creates an Entire New Investment Cycle
The first AI winners were obvious:
Nvidia.
Then investors looked toward:
AMD,
Broadcom,
memory,
networking,
and semiconductor equipment.
The infrastructure bottleneck broadens the opportunity again.
Potential beneficiaries can include industries involved in:
electrical equipment,
power generation,
gas turbines,
nuclear,
fuel cells,
transmission,
cooling,
construction,
and grid software.
AI demand escapes the technology sector.
But This Also Raises the Risk of Overbuilding
There is another side to the story.
AI demand forecasts are enormous.
Companies are planning infrastructure around extremely optimistic assumptions.
What if AI becomes dramatically more efficient?
What if demand grows more slowly?
What if customers refuse to pay enough for AI services?
What if new architectures reduce computing requirements?
Then some infrastructure could be underutilized.
That is the classic risk of every investment boom.
Railroads Had It
Telecommunications had it.
The internet had it.
Renewable energy had it.
Semiconductors had it.
Infrastructure gets built in anticipation of future demand.
Sometimes demand arrives.
Sometimes investors get ahead of reality.
AI will not be exempt from economics.
That's Why the “AI Is Running Out of Power” Headline Needs Context
No, AI is not about to shut down because America suddenly has no electricity.
Existing data centers will continue operating.
New generation is being built.
Grid upgrades are underway.
Technology will become more efficient.
Some projects will find alternative power sources.
The issue is timing.
AI infrastructure can grow faster than the electrical system supporting it.
That mismatch can delay projects and raise costs.
That is the real story.
And It May Last Longer Than the GPU Shortage
Chip fabrication capacity can expand relatively quickly when enough money is invested.
Electric grids move differently.
Transmission projects may face:
years of permitting,
local opposition,
environmental review,
equipment shortages,
and construction.
Power plants can take years.
Nuclear can take considerably longer.
That means electricity could remain a structural constraint even after chip supply improves.
This Is Why 2027 and 2028 Matter
The next few years will reveal whether planned AI infrastructure can actually be connected fast enough.
The crucial metric will not simply be:
How many GPUs did Nvidia ship?
It may be:
How many megawatts of usable AI data-center capacity actually came online?
That is a very different way of measuring the AI boom.
Consumers May Eventually Notice
Most people never think about the electricity behind ChatGPT or Gemini.
That could change.
If AI contributes to:
higher regional electricity demand,
new transmission projects,
energy debates,
or utility-rate disputes,
the physical cost of AI becomes more visible.
That could influence public opinion.
People love convenient technology.
They are less enthusiastic about paying someone else's infrastructure bill.
Tech Companies Will Need a Better Argument
“We need more power because AI is the future” may not be enough.
Communities will want tangible benefits:
jobs,
tax revenue,
grid investment,
reliable electricity,
and perhaps lower rather than higher costs.
The companies that handle those relationships well may build faster.
Those that do not could face delays.
Infrastructure politics becomes AI strategy.
The Most Important AI Engineer May Eventually Be an Electrical Engineer
That sounds like a joke.
It isn't.
AI's first phase was dominated by:
machine-learning researchers,
software engineers,
and chip designers.
Its next phase increasingly requires:
electrical engineers,
power-system planners,
civil engineers,
mechanical engineers,
construction teams,
and energy specialists.
The AI revolution is becoming much more physical.
What Should We Watch Next?
There are five signals that matter.
1. Data-center connection delays
If grid waiting times continue increasing, power is becoming a genuine deployment constraint.
2. Electricity prices
Rising prices in major data-center regions could affect both operators and consumers.
3. Nuclear and gas deals
Large technology companies signing more direct energy agreements would show how seriously they view the problem.
4. AI efficiency
Rapid improvements in performance per watt could soften the power shortage.
5. Location changes
If data centers increasingly move toward energy-rich regions, electricity has officially become one of AI's primary competitive advantages.
My Take: The AI Boom Has Entered Its Industrial Phase
The first generative AI race was easy to understand.
Who has the best model?
Then:
Who has the most GPUs?
Now the questions are becoming much more physical.
Who has the land?
Who has the cooling?
Who has the transmission?
Who has the generation?
Who has the electricity?
That is not a software race anymore.
It is an infrastructure race.
The Next Nvidia Bottleneck May Be Outside Nvidia's Control
Nvidia can keep making faster chips.
Broadcom can keep building networking hardware.
Memory manufacturers can expand HBM production.
But none of those companies can instantly rebuild the U.S. power grid.
That's why this bottleneck feels different.
The AI industry is reaching the limits of systems that were never designed for this pace of growth.
The Bigger Lesson
The cloud was never actually in the cloud.
It was always:
steel,
silicon,
water,
fiber,
concrete,
and electricity.
AI is simply making that reality impossible to ignore.
The technology may feel almost magical on your screen.
The infrastructure underneath it is anything but.
The Bottom Line
America's AI boom is beginning to collide with a surprisingly old constraint:
the power grid.
Morgan Stanley estimates that U.S. data-center developers could face roughly a 32-gigawatt net power shortfall through 2028, creating the possibility that some planned AI capacity arrives later than expected.
Nvidia and Broadcom may be relatively protected because their highest-value components are often secured early, but delayed data-center deployments could ripple into areas such as memory, optical networking, power-management components and other parts of the AI supply chain.
Meanwhile, global data-center electricity demand continues climbing rapidly, with the IEA expecting the electricity required to supply data centers to more than double by the end of the decade.
None of this means AI is literally about to run out of electricity.
It means something more interesting:
AI companies are learning that buying the world's most powerful chips is useless if you cannot get enough power to run them.
The first AI shortage was GPUs.
The next one may be megawatts.
And if that happens, the companies that control the future of AI may not only be chipmakers.
They may also be the companies that can build the power system underneath them.
Related Reading
The $3 Trillion Question: What’s Actually Being Built for the AI Boom
China Is Quietly Building Its Own Nvidia — And Nvidia’s Moat Is Starting to Crack
The U.S. and China Are About to Talk About AI Losing Control — That Should Get Your Attention
Comments
Post a Comment