The AI Power Boom Has a Forecasting Problem: “Ghost” Data Center Demand

More than 700 gigawatts of large-load requests can sound like a forecast.

It is not.

It is closer to a giant list of projects asking the grid a question:

“Could you supply this much power here?”

Some of those projects will become operating data centers.

Some will move to another site. Some will shrink. Some will wait. Some will never be built.

That is the problem behind the phrase “ghost demand.”

Reuters reported on September 1 that large electricity-use requests across major U.S. regions—primarily data centers—had risen above 700 GW.^[1]

The useful lesson is not that 700 GW is fake.

The grid has several different demand numbers, and they do not mean the same thing.

Since this article was first published, Texas has moved from discussing that problem to actively verifying large-load projects, while PJM and federal regulators have also pushed toward better large-load registries, updated forecasts and clearer cost responsibility.^[2][6]

This rebuild gives you a way to read those numbers without choosing between two misleading extremes: “every project is real” and “the AI power boom is fake.”

Quick Answer

Ghost data-center demand means requested electricity load that may not become actual energized demand. It can include duplicate site options, projects without enough financial commitment, delayed projects, or plans that disappear before construction.

The key is to separate four stages:

Gross queue MW
→ substantiated MW
→ contracted MW
→ energized MW

Each is useful.

Each answers a different question.

A utility that treats all four as identical can build too much infrastructure—or wait too long to build what real customers need.

Why One Data Center Can Look Like Three

Imagine one developer wants a 500 MW AI campus.

It has three possible sites and does not yet know which one can get power first.

So it studies connections at all three.

One developer sees:

one 500 MW project with three options

Three utilities may temporarily see:

1.5 GW of requested future load

That does not require fraud.

It is a form of optionality: keeping several paths open while power, land, permits, financing and customers are still uncertain.

But when many developers do this, gross queue totals become poor forecasts of eventual operating load.

The Scarcity Feedback Loop

There is a deeper reason queue inflation can feed itself.

Power access looks scarce
→ developers apply earlier and at more sites
→ queue MW grows
→ scarcity looks worse
→ developers seek even more optionality

A queue is therefore not just a passive measure of demand.

The queue rules can change developer behavior.

Original Asset 1: The Demand Evidence Ladder

Instead of asking whether a project is simply “real” or “ghost,” place it on an evidence ladder.

Announcement
→ grid application
→ verified / substantiated
→ collateral or binding contract
→ financing / site control
→ construction
→ equipment installed
→ energized
→ measured load

Confidence generally rises as a project crosses milestones that are more expensive to reverse.

An announcement is cheap to change.

A binding power contract backed by collateral is harder to walk away from.

A building full of installed servers is stronger evidence still.

Actual meter data is no longer a forecast.

Original Asset 2: Four Demand Numbers You Should Not Mix

Number	What it tells you	What it does not prove
Gross queue MW	How much load has asked for study or access	That all projects will be built
Substantiated MW	Projects that passed a defined verification screen	That they are financed or operating
Contracted MW	Customers willing to accept binding commercial obligations	That all contracted load is already consuming power
Energized MW	Connected load capable of drawing electricity	That it runs at full contracted capacity every hour

This simple distinction prevents one of the biggest errors in AI-power coverage.

A megawatt in a queue, a megawatt under contract, and a megawatt on the meter are three different things.

Texas Is Now Testing the Queue

Texas provides the clearest current example.

In June, the Public Utility Commission of Texas approved ERCOT's Batch Zero process for large loads of 75 MW and above. Instead of evaluating every large request independently, ERCOT can study qualifying projects together and assess their combined effect on the transmission system.^[2]

After the September ghost-demand debate, ERCOT moved into project verification.

On September 9, ERCOT said it had begun sending verification requests for information to many projects conditionally included in Batch Zero.^[3]

By September 24, ERCOT was also referring to certain 25–75 MW data centers as “Substantiated Loads” and requesting additional state and community-impact information.^[4]

That is an important change in the story.

The grid is not merely counting requests anymore.

It is asking projects to produce evidence.

But as of October 4, this research did not find a final public total for verified Batch Zero MW.

So it would be premature to say exactly what share of Texas' earlier request total will survive verification.

Ohio Shows What Happens When a Queue Meets a Contract

AEP Ohio uses a different screening mechanism.

Its Data Center Tariff requires binding commitments and, for many customers, substantial collateral. It also includes minimum-demand charges designed to make customers pay for capacity reserved on their behalf.^[5]

AEP reported that as of February 12, 2026, data-center customers or developers had signed binding contracts for 5,642 MW under the newer tariff, in addition to 12,219 MW of contracts signed before it took effect.^[7]

That is much stronger evidence than an early application.

But it still does not mean 5,642 MW was already operating on February 12.

This is exactly why the evidence ladder matters.

Original Asset 3: Queue-to-Load Conversion

Eventually, planners can calculate a historical conversion rate for a defined group of projects:

Queue-to-Load Conversion
=
eventual energized MW ÷ original gross requested MW

This can be useful.

But it should be used carefully.

A conversion rate from one utility, one year or one screening regime should not be applied automatically to the whole country.

Rules change.

Customers change.

AI hardware changes.

The best use of the ratio is to ask:

How predictive was this particular queue after we know what happened?

Does Screening Ghost Demand Make the Power Problem Go Away?

No.

It can make the forecast smaller and more credible at the same time.

Lawrence Berkeley National Laboratory's June 2026 update estimates that U.S. data centers could use about 649 TWh in 2030 in its reference case, equal to roughly 11.8% of U.S. electricity use. Its uncertainty range is 521–843 TWh, or about 9.5%–15.3%.^[8]

That forecast is not built by simply adding utility queue requests.

LBNL uses a bottom-up model based on planned IT-equipment shipments, per-device energy use, cooling performance, facility types and locations.

So two statements can both be true:

• gross interconnection queues can overstate eventual project load;
• real data-center electricity demand can still grow very rapidly.

Original Asset 4: Two Lenses Are Better Than One

A better forecast uses two independent lenses.

Project lens:

applications → verification → contracts → construction → energization

Equipment lens:

chips / servers → installed capacity → utilization → cooling overhead → electricity

If both lenses point upward, confidence rises.

If queue MW explodes while contracts, construction and equipment evidence do not, uncertainty is much larger.

The Hard Question: Who Pays for a Project That Never Arrives?

A substation, transmission upgrade or new generation resource costs real money before a future data center begins using electricity.

If the project disappears, someone still has to absorb that cost.

This is why collateral and minimum-demand charges matter.

They move some of the risk back toward the customer asking the grid to reserve capacity.

The issue is now active at the regional level too.

PJM is developing a Large Load Registry and new mechanisms for integrating very large customers. In late September, FERC also directed changes around PJM's backstop procurement and emphasized updated load forecasts and cost allocation tied to new large-load growth.^[6][9]

These proceedings are still evolving.

The important point is the direction:

The more uncertain the demand, the more important it becomes to decide who carries the risk before infrastructure is built.

Verification Has Its Own Failure Mode

It would be easy to conclude that utilities should simply make every application more expensive and difficult.

That creates another problem.

A serious new customer may genuinely need a fast connection before every commercial detail is final.

If screening is too weak, utilities can plan around projects that disappear.

If screening is too strict or slow, viable projects can lose years waiting for power.

Lawrence Berkeley National Laboratory's 2026 Speed to Power work treats large-load interconnection as a broader problem involving forecasting, interconnection, procurement, operations and cost allocation—not just filtering bad applications.^[10]

Original Asset 5: The Two-Sided Forecast Error

The grid is trying to avoid two different mistakes.

False positive:

“Treat this uncertain project as real”
→ build too early
→ infrastructure is underused

False negative:

“Treat this serious project as speculative”
→ delay connection
→ real demand arrives before infrastructure

Good verification has to reduce both.

What to Watch Next

1. Texas Batch Zero results. How much conditionally included load survives verification?
2. Substantiated vs contracted MW. Do verified projects continue to financial commitment?
3. PJM's Large Load Registry. Can better regional data reduce double counting and improve forecasts?
4. Collateral and minimum-demand rules. Do they remove weak projects without blocking credible ones?
5. Construction evidence. Are substations, buildings and IT equipment actually arriving?
6. Energized load. How closely do today's queues predict actual metered demand several years later?
7. Bottom-up forecasts. Do equipment-shipment and utilization data continue to support rapid national demand growth after queue screening?

The Simple Idea to Remember

AI may create a genuine power shortage.

But a large queue is not proof of the size of that shortage.

The grid now has to solve two problems at once:

build enough for demand that will arrive
without building too much for demand that will not

The closer a project gets to irreversible commitment, the more useful its megawatts become as a forecast.

Key Vocabulary

interconnection queue
A pipeline of proposed projects requesting grid study, connection or service.

large load
A customer or project that requests unusually large amounts of electricity, such as a hyperscale data center.

ghost demand
Requested load that may be duplicative, optional, delayed or unlikely to become energized load. The term does not imply fraud.

substantiated load
A project that has passed a defined verification screen under the relevant utility/grid process.

collateral
Financial security posted by a customer to back its commitments and protect others if the project does not proceed.

minimum-demand charge
A billing rule requiring a large customer to pay for a minimum level of reserved demand even if actual use is lower.

energized load
A project that has been physically connected and is able to draw power from the grid.

Related Articles

• Why Power, Not Chips, May Limit the AI Data Center Boom
• Cheap Power Is Not Always Cheap: The Grid Delay Behind AI Data Center Economics
• What Must Be Built to Power the AI Data Center Boom?
• How to Understand a 100 MW Data Center Power Model

Sources

1. Reuters — Texas' halt on powering data centers reflects U.S. reckoning over “ghost” demand, September 1, 2026.
2. PUCT / ERCOT — Batch Zero process for connecting large electricity users, June 18, 2026.
3. ERCOT — Batch Zero Verification Requests for Information, September 9, 2026.
4. ERCOT — State and Community Impact RFI / Substantiated Medium Load Data Centers, September 24, 2026.
5. AEP Ohio — Data Center Tariff, checked October 4, 2026.
6. PJM — Large Load Registry and Interim Resource Adequacy Service, checked October 4, 2026.
7. AEP Ohio — Data Center Tariff load update, February 13, 2026.
8. Lawrence Berkeley National Laboratory — United States Data Center Energy Usage Report: 2025 Update, June 2026.
9. FERC — PJM Reliability Backstop Procurement concurrence, September 29, 2026.
10. Lawrence Berkeley National Laboratory — Speed to Power: solutions for accelerating large-load connections, June 11, 2026.

Update History

October 4, 2026 — Updated with ERCOT project verification, PJM/FERC large-load developments, and a new demand-evidence framework.

Sources checked through October 4, 2026. “Ghost demand” is used as a planning term for uncertain or duplicative requested load, not as an accusation of fraud. Gross queue MW should not be treated as a direct forecast of future electricity consumption.