AI’s “Ghost Demand” Is Becoming Real Economic Risk—and Texas Just Hit the Brakes

U.S. data-center developers have requested more than 700 gigawatts of electricity—over ten times estimated current data-center use in the country. Now utilities are discovering that some demand disappears when companies have to prove they can actually fund their projects. Combined with supplier-backed AI financing and weak ROI visibility, the problem is bigger than overbuilding: speculative demand can create real costs long before the customers arrive.

he AI investment boom has been built around one enormous assumption:

Demand is coming.

More models will need more compute.

More agents will need more inference.

More companies will deploy AI.

More consumers will use it.

More workloads will move into data centers.

So build.

Buy the GPUs.

Secure the land.

Lease the data centers.

Reserve the electricity.

Raise the financing.

Build some more.

But Texas has just introduced a very inconvenient question into that narrative:

How much of the demand is actually real?

Reuters found that requests from very large electricity users—mostly data centers—have surpassed 700 gigawatts across portions of the Midwest, Mid-Atlantic and South.

That is more than ten times estimates of current U.S. data-center electricity consumption.

Texas alone has seen requests surge from roughly 48 GW in 2023 to more than 474 GW.

The number became so difficult to trust that Texas froze new data-center grid connections while regulators audit which proposed projects are legitimate.

The industry has a name for the problem:

ghost demand.

Projects request enormous quantities of electricity.

Utilities prepare for the load.

Infrastructure planning responds.

Capital reacts.

But some projects may be:

duplicated,

speculative,

underfunded,

or unlikely ever to be built.

Here is what makes that important beyond Texas:

Ghost demand can still produce very real economic consequences.

The Ghost Disappears When Someone Asks for Money

One of the most revealing details in the Reuters investigation is what happened when utilities started requiring stronger financial commitments.

Exelon tightened collateral requirements.

Its estimate of high-probability data-center demand fell by roughly 40%, to about 11 GW.

Ohio imposed requirements that included connection-study fees as high as $100,000.

AEP Ohio's data-center demand pipeline fell by more than half.

Nothing happened to AI capability overnight.

No model suddenly became less intelligent.

No data-center technology disappeared.

The requirement changed.

Companies had to put more financial credibility behind the electricity they claimed they would eventually consume.

And a significant amount of projected demand disappeared.

That deserves attention.

Because it reveals a distinction the AI investment boom increasingly needs:

expressed demand

versus

financeable demand

versus

economically productive demand.

Those are not the same thing.

Reserving Power Is Not the Same Thing as Having Customers

A developer can request 1 GW of electricity.

That does not prove:

the data center will be financed,

the GPUs will be purchased,

the facility will be completed,

customers will lease the capacity,

AI workloads will materialize,

or those workloads will generate enough economic value to support the entire investment.

Yet electricity requests enter infrastructure planning long before all of those questions have been answered.

That creates an extraordinary inversion.

The physical economy begins preparing for future AI demand before the AI economy has proven that demand economically.

Utilities have to think about:

generation,

transmission,

substations,

grid upgrades,

land,

capacity auctions,

and reliability.

Those investments cannot be conjured instantly after demand appears.

So infrastructure must anticipate.

But anticipation creates vulnerability to exaggeration.

This Is the AI Ghost Demand Multiplier

Call it the:

AI Ghost Demand Multiplier.

It occurs when an expectation of future AI demand creates real economic activity before the underlying end-market demand has been validated.

The sequence looks like this:

Expected AI demand

→ developer announces data center

→ enormous electricity request

→ grid forecast rises

→ utilities plan new infrastructure

→ financing expectations increase

→ land and power become more valuable

→ AI infrastructure valuations rise

→ more developers enter

→ more power gets reserved

→ forecasts rise again.

The original demand may still be uncertain.

The economic consequences no longer are.

That is the multiplier.

A speculative request can create very real signals elsewhere in the economy.

The Public Can Pay for Demand That Never Arrives

Reuters identifies one particularly serious consequence.

Utilities have to build for expected future demand.

If they underbuild and AI demand really arrives, electricity reliability suffers.

If they overbuild and the projects never materialize, somebody still has to pay for infrastructure that was constructed in anticipation of them.

That somebody can include ordinary electricity customers.

Consumer advocates have warned that speculative data-center demand can result in households and businesses absorbing unnecessary grid costs.

In PJM, the country's largest electricity market, growth in existing and forecast data-center demand contributed to an estimated $29.4 billion increase in capacity costs across roughly four recent auctions.

That changes the nature of the AI investment story.

The risk is no longer confined to venture capitalists.

Or Nvidia shareholders.

Or AI startups.

Or hyperscalers.

Expectation itself can migrate into:

electricity bills,

grid investment,

public infrastructure,

land prices,

energy markets,

and regional economic planning.

The AI boom develops externalities before the AI revenue arrives.

This Connects Directly to AI Demand Circularity Risk

This matters because the electricity story is not occurring independently of what is happening in AI finance.

Nvidia recently paused parts of a program designed to help smaller AI cloud companies finance purchases of Nvidia hardware.

The proposed model could:

support the customer's financing,

allow Nvidia to rent capacity back if the cloud company could not sell all of it,

and give Nvidia 50% of cloud revenue above an agreed threshold generated from Nvidia-powered capacity.

Nvidia has also helped assemble approximately $500 billion in financing for AI customers and agreed to guarantee up to $105 billion associated with OpenAI data-center infrastructure.

Those arrangements can have legitimate commercial purposes.

They can accelerate infrastructure.

They can help smaller companies obtain capital.

They can expand access to compute.

But they create a question that becomes much more important after the Texas findings:

What happens when anticipated demand, supplier-supported financing and speculative infrastructure requests begin validating one another?

The Demand Can Start Proving Itself

Consider the feedback loop.

Investors see extraordinary AI demand forecasts.

That makes data centers attractive.

Developers request enormous amounts of electricity.

Utilities report huge future demand.

Those forecasts appear to validate the AI boom.

Higher expectations attract more financing.

Financing allows more data centers to be proposed.

More data centers place more electricity requests.

Those requests strengthen electricity-demand forecasts.

The forecasts support additional infrastructure investment.

And the expanding infrastructure appears to prove the original assumption:

AI demand must be enormous. Look at everything being built for it.

But that reasoning can become reflexive.

Infrastructure being planned because people expect demand cannot simultaneously be treated as independent proof that the demand exists.

Demand Forecasts Can Become Valuation Inputs

This matters enormously on Wall Street.

AI valuations depend partly on expectations.

Future revenue.

Future customers.

Future capacity.

Future compute.

Future market share.

Future productivity.

A company does not need to generate all of tomorrow's profits today.

The market values what investors believe it can generate later.

That is normal capitalism.

The problem appears when one speculative indicator begins validating another.

Data-center plans demonstrate demand.

GPU orders demonstrate demand.

Electricity requests demonstrate demand.

AI-cloud financing demonstrates demand.

Infrastructure leases demonstrate demand.

Model revenue projections demonstrate demand.

Then each metric reinforces valuations elsewhere in the ecosystem.

But several of those indicators may derive from the same underlying expectation of future AI usage.

They are not necessarily five independent signals.

They may be five manifestations of one bet.

One Dollar of Expected Demand Can Cast Multiple Shadows

Imagine an AI developer expects one future customer to require enormous compute.

That expectation can generate:

a data-center proposal,

a power reservation,

a GPU order,

a financing agreement,

a cloud contract,

a construction commitment,

and a supplier revenue forecast.

Seven economic signals.

One expected source of final demand.

If that customer materializes and creates enough value, the infrastructure was justified.

If it does not, the economy is left with multiple obligations created around demand that existed primarily as expectation.

That is why counting activity is not the same as measuring demand.

Then Ask Who the Final Customer Is

This remains the most important question in the entire AI investment boom.

Who ultimately pays?

Not Nvidia.

Not the cloud provider.

Not the data-center developer.

Not the bank.

Not the utility.

Not the venture-capital fund.

Eventually somebody outside the AI financing loop has to create enough economic value to support:

the chips,

the data center,

the electricity,

the financing,

the cloud margin,

the model,

the agent,

and the investor return.

That customer might be:

a hospital,

a manufacturer,

a bank,

a pharmaceutical company,

a government,

a retailer,

a software company,

or a consumer.

Eventually the economic chain has to terminate in productive demand.

Without that, the infrastructure is simply passing financial obligations upstream.

AI Adoption Does Not Automatically Mean AI ROI

This is where the infrastructure boom can diverge from enterprise reality.

A company can integrate AI.

Employees can use it.

Tokens can be consumed.

Models can run.

Cloud providers can report usage.

None of that automatically proves the company generated economic returns exceeding the complete cost of deployment.

The actual equation includes:

model costs,

compute,

integration,

human review,

security,

monitoring,

workflow redesign,

training,

remediation,

vendor dependency,

and failure exposure.

AI usage is measurable immediately.

AI ROI can take far longer to establish.

That temporal gap matters when infrastructure is being built today based on assumptions about economic value tomorrow.

Society Cannot Reconfigure as Quickly as Compute

There is another dimension to this problem.

Technology can scale extremely quickly.

Society cannot.

A new model can deploy in weeks.

A data center can consume the electricity of a city.

Automation can remove thousands of jobs.

But displaced workers do not instantaneously move into equally productive new roles.

Organizations do not instantly redesign themselves around AI.

Education systems do not instantly retrain populations.

New industries do not instantly absorb displaced labor.

Consumer demand does not automatically expand because corporate productivity increased.

Capital moves faster than institutions.

That creates a timing problem.

Productivity Can Rise While Demand Weakens Somewhere Else

Suppose AI allows a company to perform the work of 10,000 people with 7,000.

The company's productivity may improve.

Margins may improve.

AI spending appears justified.

But 3,000 people now have less income.

Multiply that across industries.

Automation can simultaneously:

improve corporate efficiency

and

reduce labor income elsewhere.

Eventually that matters because workers are also consumers.

They buy:

homes,

cars,

food,

healthcare,

software,

travel,

retail products,

and services.

Economic systems need both productivity and purchasing power.

If capital capture moves much faster than income replacement, AI can create a temporary—and potentially significant—demand imbalance.

That does not mean automation is bad.

Technological revolutions have repeatedly created enormous new prosperity.

It means economic adaptation has a speed limit.

AI Infrastructure Does Not Wait for Society to Catch Up

The investment machine is operating at extraordinary speed.

Reuters reports that Big Tech's planned AI data-center spending now exceeds $700 billion this year.

Meanwhile:

data-center developers secure power,

AI companies sign enormous cloud contracts,

GPU providers support customer financing,

banks finance infrastructure,

and public markets price future AI growth.

The infrastructure is being committed now.

The productivity transformation takes longer.

The employment adjustment takes longer.

The organizational redesign takes longer.

The creation of entirely new AI-native industries takes longer.

That timing difference is risk.

This Does Not Mean the Demand Is Fake

This distinction matters.

There is unquestionably enormous real AI demand.

PJM itself says that even after utilities reduce questionable forecasts, substantiated data-center demand remains large enough to overwhelm electricity systems where generation is not expanding quickly enough.

The problem is therefore not:

AI demand doesn't exist.

It does.

The question is:

How much?

At what price?

For how long?

Supported by what final customer economics?

And how much infrastructure should be built before those answers become clearer?

That is a much harder investment problem.

“Ghost Demand” Is Not Necessarily Fraud

Nor does ghost demand automatically mean somebody is lying.

Several rational behaviors can produce inflated aggregate requests.

Developers may submit applications in multiple places while deciding where to build.

Companies may reserve more power than they ultimately expect to use.

Landowners may secure grid positions because those positions themselves become valuable.

Projects may be genuine when proposed but later fail to secure financing.

Market conditions change.

Technology changes.

Customers change.

Demand forecasts change.

Individually rational behavior can still produce collectively distorted signals.

That distinction is important.

Market distortion does not always require market manipulation.

Sometimes incentives are sufficient.

This Is Why the Systemic Failure Is More Interesting Than a Conspiracy

The more important question is not whether several companies secretly coordinated to inflate AI.

It is whether the structure of the boom itself rewards everyone for forecasting enormous future demand.

Data-center developers benefit from securing power.

Utilities respond to connection requests.

Chip companies benefit from expected infrastructure growth.

AI labs benefit from confidence in future adoption.

Cloud companies benefit from capacity expansion.

Banks benefit from financing.

Investors benefit when valuations rise.

Local governments expect jobs and investment.

Every participant can make a rational decision.

And collectively they can still build too much.

That is a classic systemic-risk problem.

No conspiracy required.

The Ghost Demand Multiplier Can Reach the Real Economy

The sequence can become:

AI growth expectation

→ larger infrastructure forecast

→ greater electricity reservation

→ more grid construction

→ higher capacity costs

→ higher electricity bills

→ more financing

→ more data centers

→ more GPU purchases

→ stronger AI revenue

→ higher valuations

→ greater confidence in AI demand

→ additional infrastructure commitments.

The original expectation now affects:

consumers,

businesses,

utilities,

interest rates,

energy infrastructure,

public policy,

and capital markets.

That is why this is not simply an AI bubble question.

It is a multiplier-risk question.

The Small AI Cloud Companies Sit in a Difficult Position

Now return to the Nvidia financing model.

A smaller AI cloud company sees enormous AI growth forecasts.

It fears being left behind.

It needs expensive GPUs to compete.

Nvidia can help make financing easier.

The cloud company buys the equipment.

It assumes financing obligations.

It builds capacity.

Then it has to find customers.

Under Nvidia's proposed revenue-sharing structure, Nvidia could potentially rent unused capacity back and take 50% of revenue above specified thresholds.

That creates a difficult business model if end-market demand disappoints.

The infrastructure exists.

The financing exists.

The GPU obligation exists.

But utilization determines whether the economics work.

The company cannot pay debt with projected demand.

Eventually somebody has to actually rent the compute.

FOMO Can Become Capital Misallocation

The AI boom contains an especially powerful commercial message:

If you don't build now, you will be left behind.

That message may be correct for some organizations.

It can also encourage overinvestment.

Buy capacity.

Secure chips.

Reserve power.

Sign long contracts.

Build before competitors do.

The risk is that fear of missing future demand can create current demand for infrastructure.

Those are different things.

A company purchasing GPUs because it has profitable customers is one signal.

A company purchasing GPUs because it fears that future customers may punish it for not having them is another.

Both generate the same hardware sale.

They do not represent the same economic certainty.

This Is Where AI Demand Circularity Risk Becomes Physical

My earlier concern around AI Demand Circularity Risk was financial:

How much demand exists independently of the supplier:

financing,

guaranteeing,

investing in,

or purchasing capacity back from

the customer buying its products?

The Texas story adds another layer.

Now speculative demand can manifest physically as:

reserved electricity,

planned transmission,

generation requirements,

substations,

land development,

and capacity-market costs.

AI Demand Circularity Risk can therefore become infrastructure circularity risk.

Financial expectations become physical assets.

The physical assets then appear to validate the financial expectations.

The Market Can Be Right About AI and Wrong About Timing

This may ultimately be the most important distinction.

AI may transform the global economy.

AI may create enormous productivity.

AI may justify trillions of dollars of infrastructure over time.

And the current buildout can still overshoot.

Those propositions are completely compatible.

The internet changed everything.

Telecommunications companies still massively overbuilt fiber during the dot-com era.

Railroads transformed economies.

Railroad investment still produced speculative booms and bankruptcies.

Transformative technology does not eliminate capital cycles.

Sometimes it intensifies them.

The question is not whether AI matters.

It clearly does.

The question is whether capital is attempting to price decades of transformation into infrastructure being built in a few years.

When Expectations Become Liabilities

There is another useful distinction:

AI expectation → AI obligation.

A forecast is flexible.

Debt is not.

A prediction can change.

A data-center lease still has payments.

A projection can be revised.

A GPU financing obligation remains.

A power plant still has to be paid for.

A transmission line still exists.

A cloud contract still has minimum commitments.

A company can change its mind about future demand.

Infrastructure cannot always change with it.

That is where optimism becomes exposure.

The Balance Sheet Eventually Wins

During a technology boom, narrative can outrun economics for a long time.

Eventually:

interest gets paid,

leases come due,

utilization gets measured,

customers renew or leave,

power gets consumed or remains unused,

and revenue either materializes or does not.

That is where infrastructure investment becomes unforgiving.

You cannot pay a data-center lease with future enthusiasm.

You cannot service debt with projected tokens.

You cannot finance an electrical grid with hypothetical customers forever.

Eventually the balance sheet asks:

Who is paying?

Wall Street Is Not Outside the Loop

Wall Street plays an important role because infrastructure this large requires capital.

Banks.

Private credit.

Private equity.

Institutional investors.

Public markets.

Bond markets.

All participate.

That does not make Wall Street complicit in some proven scheme.

It means financial institutions are part of the transmission mechanism between AI expectations and physical investment.

If AI projections are right, the returns could be extraordinary.

If projections overshoot, losses will not remain confined to technology companies.

They can migrate into:

credit,

utilities,

real estate,

energy,

public markets,

and eventually government policy.

That is how sector risk becomes macroeconomic risk.

Then Comes the Bailout Question

There is no evidence today that AI companies will require a taxpayer bailout.

But concentration and dependency make the question legitimate as a future risk scenario.

What happens if society becomes dependent on:

AI infrastructure,

AI healthcare systems,

AI cybersecurity,

AI financial systems,

and AI cloud capacity

while the companies providing them become financially distressed?

Governments have historically intervened when strategically important infrastructure becomes too consequential to fail abruptly.

Banks.

Airlines.

Automakers.

Energy systems.

Financial markets.

The more AI becomes embedded in essential economic functions, the more the phrase:

too important to fail

could eventually enter the conversation.

That possibility makes infrastructure discipline more important now.

Reliance Changes Political Risk

The technology industry wants rapid adoption.

Commercially, that makes sense.

But rapid adoption can create dependency before the underlying business model has matured.

Once:

hospitals,

governments,

manufacturers,

banks,

schools,

and critical infrastructure

depend on AI systems continuously, the failure of a major provider becomes more than a shareholder problem.

It becomes a continuity problem.

That can transfer private risk toward the public sphere.

The greater the dependency, the harder it becomes to allow failure to resolve itself normally.

That is another reason why inflated or speculative infrastructure growth matters.

The Real Threat Is Misallocation

Capitalism depends on risk.

Companies make bets.

Some succeed.

Some fail.

Capital moves toward better uses.

That process is essential.

The danger appears when distorted demand signals cause enormous quantities of capital to move toward the wrong place simultaneously.

Then the consequences become systemic.

Workers train for the wrong jobs.

Utilities build the wrong infrastructure.

Banks finance the wrong capacity.

Cities plan around projects that never arrive.

Companies borrow against customers that never materialize.

Public markets assign valuations based on growth that takes longer than expected.

The AI itself can still be revolutionary.

The allocation around it can still fail.

That Is an AI Failure Too

AI failure intelligence cannot stop at:

hallucination,

bias,

cybersecurity,

model drift,

or physical harm.

There is another category:

macroeconomic AI failure.

The technology works.

But the expectations surrounding it create:

overcapacity,

capital misallocation,

energy distortion,

labor disruption,

financial fragility,

and infrastructure that arrives faster than economically productive demand.

The failure is not inside the model.

It exists in the economic system built around the model.

The Questions Texas Is Forcing Everyone Else to Ask

The Texas freeze introduces questions that should travel well beyond electricity regulation.

How much AI infrastructure demand is backed by actual customers?

How much is backed by financing?

How much exists because developers submitted requests in multiple markets?

How much disappears when deposits are required?

How much GPU demand depends on supplier-supported financing?

How much cloud capacity is supported by guarantees or buybacks?

How much electricity infrastructure is being planned around projects that may never be built?

Who pays when demand projections are wrong?

How much of AI valuation depends on infrastructure growth being interpreted as independent evidence of end-user demand?

How much productive economic value exists at the end of the capital chain?

And the hardest question:

Are we building infrastructure because customers need it—or assuming customers must need it because we are building so much infrastructure?

The Strategic Conclusion

Texas did not stop AI.

It did something much more consequential.

It asked for proof.

Who owns the proposed data center?

Is the project actually financed?

Is the electricity request credible?

Is the developer financially capable of building what it claims it will build?

And when stronger financial requirements appeared elsewhere, significant portions of projected demand disappeared.

That should matter to everyone following the AI investment boom.

Because the industry's problem may not be that AI demand is fictitious.

There is clearly enormous real demand.

The problem may be that real demand, anticipated demand, duplicated demand, financially supported demand, speculative demand and economically productive demand are repeatedly being counted as though they mean the same thing.

They do not.

And every layer built around those assumptions creates obligations.

The chip is real.

The debt is real.

The lease is real.

The data center is real.

The transmission line is real.

The electricity cost is real.

The consumer's bill is real.

Only the customer at the end may still be hypothetical.

That is the danger of the AI Ghost Demand Multiplier.

Expectation can become infrastructure.

Infrastructure can become revenue.

Revenue can become valuation.

Valuation can attract financing.

Financing can create more infrastructure.

And eventually the existence of all that infrastructure is offered back to the market as evidence that the original expectation must have been correct.

That loop can produce enormous prosperity if real economic demand catches up.

If it does not, the ghost leaves behind very real liabilities.

This is not evidence of a Ponzi scheme.

It is not proof of coordinated market manipulation.

It is something markets have encountered many times before:

a transformative technology, extraordinary capital, enormous expectations, and a financial system struggling to distinguish the scale of the eventual revolution from the speed at which it should be funded today.

AI may transform everything.

Society may simply need longer to absorb that transformation than capital markets are currently pricing.

Workers have to transition.

Organizations have to restructure.

New revenue models have to develop.

Consumers need purchasing power.

Infrastructure has to find productive users.

ROI eventually has to reach the end customer.

Technology can move exponentially.

Economies usually do not.

And that mismatch may become one of the defining AI failure risks of this investment cycle.

Because the most dangerous demand in a capital boom is not necessarily demand that is completely fake.

It is demand that is real someday, financed as though it were real today.

I write about AI failure intelligence, ROI exposure, high-stakes decision architecture, and the hidden pathways through which AI incidents become financial and institutional consequences.

Follow me and subscribe to my work if you are responsible for investing in, acquiring, governing, insuring, or protecting strategically important AI systems and need to understand what technical failure can become after it leaves the engineering team.