Phantom Data Centers: How Utilities Sort Real Load

Phantom Data Centers: Telling Real Load From Speculation

Phantom data centers have become one of the strangest problems in the power sector. Utilities are receiving requests to connect data center load on a scale that, taken at face value, would reshape national electricity demand. Many of those requests will never be built. Developers file in several places for the same project, hold capacity as an option, or announce sites before securing land, customers, or financing. In October, reporting based on Wood Mackenzie analysis put fewer than a quarter of disclosed data center requests in the high-confidence category. Utilities are now tightening the rules. Exelon, for example, began requiring binding financial commitments and collateral, and its list of high-probability projects reportedly shrank from about 18,000 megawatts to 11,000. Separating real AI load from speculation is now a forecasting problem with billions of dollars riding on it.

Signals That Separate Real Load From Speculation

No single signal is decisive, but together they paint a fairly clear picture.

Why Phantom Data Centers Became a Problem

The logic from a developer's side is easy to understand. Power is now the scarcest input for AI infrastructure, connection timelines in some regions stretch for years, and the first to secure capacity wins. Filing early, and filing in several places, is a rational way to keep options open when the cost of filing is low. Some developers are speculators who hope to sell a site with a power commitment attached. Others are genuine builders hedging against delays.

The result, in aggregate, is a queue full of requests that cannot all be real. Wood Mackenzie has separately estimated that US grid operators are likely to serve only a minority of the total data center capacity requested, describing much of the rest as phantom load.

What Phantom Data Centers Cost Everyone Else

If utilities plan for load that never arrives, they risk building transmission, substations, and generation that sits underused, with costs that can end up in other customers' bills. If they discount requests too heavily, they risk being caught short when real projects arrive. Meanwhile, genuine projects, including non-data-center industrial load, wait behind speculative ones in the queue. Getting the forecast wrong in either direction is expensive.

How Utilities Are Responding

Utilities and regulators are converging on a set of tools that make speculation more expensive and commitment more visible:

• Collateral and deposits. Requiring money at risk before a request is studied or capacity is reserved.
• Minimum payment tariffs. Large load tariffs that require customers to pay for a high share of contracted capacity for many years, whether or not they use it.
• Readiness milestones. Requiring proof of land control, permits, and equipment orders at defined stages to stay in the queue.
• Application fees and credit thresholds. Raising the cost of filing and screening out weak counterparties.
• Queue reform. Moving from first-come, first-served toward first-ready, first-connected approaches, as several grid operators have done.
• Phased connections. Connecting load in stages tied to demonstrated progress rather than all at once.

Forecasting When Most Requests May Not Be Real

Traditional load forecasting assumed that a request to connect was a reasonable proxy for future demand. That assumption no longer holds for data centers. The more useful approach treats each request as a probability, weighted by the readiness signals in the table above, and builds forecasts as ranges rather than single numbers. Utilities that track how often requests at each stage actually convert to energised load can calibrate those probabilities over time.

AI tools can help sort and score large volumes of requests, but they inherit the same uncertainty, and a model trained on a few years of unusual demand growth may extrapolate badly. As we argued in AI in energy and utilities, grid-critical decisions belong with planners who can explain the assumptions, with AI in an advisory role.

What Developers Should Expect

1. Money at risk earlier. Expect deposits or collateral before a utility commits study resources or capacity.
2. Long minimum payment terms. Large load tariffs increasingly require paying for most contracted capacity over many years.
3. Proof of readiness. Be ready to show land control, permits, financing, and equipment orders at each stage.
4. Less tolerance for duplicates. Filing the same project in several territories may cost more and may be flagged.
5. Phased energisation. Plan for capacity delivered in steps tied to construction progress.
6. More disclosure. Regulators may ask for information on end customers and expected load profiles.
7. On-site and off-grid options. Some projects are pairing grid service with private generation to move faster.

Pros and Cons of Tougher Connection Rules

• Pro: cleaner queues. Real projects, including other industrial load, move faster when speculation is filtered out.
• Pro: fairer cost allocation. Collateral and minimum payments reduce the risk of stranded costs falling on other customers.
• Pro: better forecasts. Committed requests are a far more reliable basis for grid planning.
• Con: higher barriers for smaller developers. Collateral requirements favour well-capitalised players.
• Con: risk of overcorrection. Rules that are too strict could push real investment to other regions.
• Con: administrative burden. Verifying readiness across many requests takes utility and regulator time.

Real Scenarios Worth Thinking Through

These scenarios are illustrative, showing how phantom data centers play out in practice rather than presented as verified case studies.

A regional utility introduces collateral requirements and sees its data center request list fall by more than a third within months. The remaining projects have land, customers, and equipment on order. The planning team rebuilds its forecast around the smaller, more credible list and defers one expensive transmission project.

A developer files the same campus in three utility territories to secure the fastest connection. Two of the utilities share information through a regional working group and flag the duplication. The developer withdraws two requests and focuses on the site with the clearest path to power.

A utility underweights data center requests after a wave of cancellations, then finds itself short of capacity when a hyperscaler commits to a large campus. The episode pushes it toward probability-weighted forecasting rather than all-or-nothing assumptions.

“After testing multiple AI models on coding, research, and business prompts, combined outputs produced more reliable results than any single model.” Internal multi-model evaluation, Talkory research team.

Who Pays If the Load Never Arrives

Behind every phantom request is a cost allocation question. If a utility builds for load that never materialises, someone pays for the stranded infrastructure. Minimum payment tariffs and collateral are attempts to make sure that someone is the developer who asked, not households and small businesses. Regulators in several jurisdictions are actively debating where that line should sit, and the outcome will shape where AI infrastructure gets built.

We explored the broader fight over who funds grid upgrades in who pays for the grid upgrade. Phantom load makes that fight sharper, because the bill can arrive for capacity nobody ends up using.

Why Talkory Wins

Utility planners, regulators, and developers are all reading the same filings, tariffs, and market reports, and often drawing different conclusions. Talkory runs the same question across GPT, Claude, Gemini, Grok, Perplexity Sonar, and Kimi K3 in one pass. A team can ask how a tariff clause allocates risk, what readiness signals a request shows, or how comparable jurisdictions have handled large load, and see whether six independent models agree. Agreement suggests a mainstream interpretation. Disagreement flags exactly where to go back to the filing or the regulator before a decision is made.

Final Verdict

Phantom data centers are a symptom of power becoming the scarcest input for AI. Developers have every reason to file early and often, and utilities have every reason to doubt them. The response is converging: collateral, minimum payment tariffs, readiness milestones, queue reform, and phased connections, alongside probability-weighted forecasting that treats requests as possibilities rather than promises. Done well, those tools protect other customers and speed up real projects. Done badly, they either strand infrastructure or drive investment elsewhere. The forecast is only as good as the filter.