In 1965, the northeast United States went dark, and the lesson utilities drew from that blackout was not about relays but about visibility: you cannot manage what you cannot see, and you cannot trust a signal you have not verified. The lesson utilities drew from that blackout was not about relays. It was about visibility: you cannot manage what you cannot see, and you cannot trust a signal you have not verified. Decades later, the same lesson is being relearned in a domain where the signals are fluent, confident, and frequently wrong. Artificial intelligence is being wired into the machinery of electricity supply at precisely the moment demand is turning upward for the first time in a generation, and the gap between what these systems appear to know and what they actually know is becoming a structural feature of the power business rather than a bug in it.
A Flat Line That Bent
For roughly twenty years, U.S. electricity consumption grew at about half a percent a year — so close to flat that planners treated stagnation as the baseline. Rabobank research projects that U.S. power consumption could climb 20 percent by 2030, a dramatic shift for a grid accustomed to nearly flat annual growth of around 0.5 percent, as Yahoo Finance reports The forces behind the bend are familiar: hyperscale data centers built for AI workloads, the electrification of transport, and the reshoring of manufacturing. What matters for the deception angle is not the size of the number but the nature of the demand. A data center does not consume power like a neighborhood. It draws in bursts, it signs contracts years ahead, and it arrives with its own forecasts — forecasts increasingly generated by the same class of models the facility exists to run.
That circularity is where the first illusion takes hold. A hyperscaler’s projection of its future load feeds a utility’s projection of peak demand, which feeds an investment plan for substations and high-voltage lines, which feeds a rate case. At no point in that chain is there a physical measurement of what the building will actually draw when it is running. The number is a model output wearing the costume of a fact, and every downstream decision inherits its uncertainty while shedding its label.
The Forecast That Forgot It Was a Guess
Utilities have always forecast. The difference now is the confidence interval. Machine learning models trained on historical load patterns perform impressively on the past and poorly on a future that contains no precedent — and the current demand curve has no precedent in the working memory of anyone now employed in the sector. A model asked to predict consumption during a reindustrialization and an AI buildout is extrapolating from a period of near-zero growth. It will produce a number. It will produce that number with the same clean formatting whether its inputs are representative or not.
This is the deception in its purest form: not a lie, but an answer delivered without the epistemic humility that a human analyst might attach. The system does not say “I have never seen anything like this.” It says 20 percent, or whatever the regression yields, and the presentation is indistinguishable from a case where the model has genuine grounds. Grid planners are not naive, but they are busy, and a dashboard that renders uncertainty as a decimal invites treatment as measurement.
The Merchant’s Long Contract and the Price It Hides
Independent power producers have moved fastest to monetize the boom, and their deals illustrate a second variety of gap. Constellation Energy signed a 20-year power purchase agreement with Alphabet and also has deals with Amazon, Microsoft, and Walmart; Amazon also has power purchase agreements with Vistra and Talen Energy. On paper, these are triumphs of foresight — long-term revenue locked against volatile spot markets.
Look closer at what the contract claims versus what it secures. A 20-year PPA fixes a price and a volume, but the counterparty’s obligation depends on the data center being built, powered, and kept running at the projected scale. The agreement is a promise about a future that neither party controls, and the AI-driven demand forecast sitting behind it is the same forecast that has already been wrong in both directions, as the source’s own framing of the demand inflection suggests. The market reads the headline number — 20 years, a hyperscaler’s name — and treats the risk as retired. It is not retired. It is transferred to a document whose assumptions will not be tested for a decade, by which time the people who signed it will have moved on.
Constellation and Vistra are down 30 percent and 22 percent over the past year, which suggests investors have begun to price some of this. But the decline is read in the market as a valuation story, not as a signal that the underlying contracts may be less firm than their duration implies. The appearance of security and the substance of it have drifted apart, and the drift is invisible on a term sheet.
Equipment That Ships Before the Load Arrives
The equipment makers sit at the far end of the same chain. GE Vernova carries a backlog above $176 billion for gas turbines and related gear; Bloom Energy has seen demand for its solid-oxide fuel cells surge as data center operators look for power that can be delivered in months rather than years. The pitch is speed, and the speed is real.
The deception here is subtler and lives in the word “backlog.” A backlog is an accounting construct that converts signed orders into an appearance of future revenue, and it assumes the orders convert. When the orders are driven by hyperscaler timelines that depend on AI demand forecasts, the backlog is a stack of conditional promises. GE Vernova will build the turbines. Whether the data centers that ordered them are running at projected capacity when the turbines arrive is a separate question, and it is not one the backlog answers. The equipment gets delivered; the demand that justified it remains a projection until the meters spin.
The Boundary Between Correlation and Causation
The deeper problem with AI in energy planning is not that the models are bad. It is that they are good at the wrong thing. A model trained on historical load and weather can find correlations that hold in the training data and will continue to hold only as long as the world resembles that data. The current moment — reindustrialization, AI buildout, electrification, and climate volatility arriving together — is precisely the world that does not resemble the training set.
When such a model is asked to attribute demand growth to a cause, it will produce an attribution. It will say data centers account for this share, electrification for that share, and the numbers will look like a decomposition. But correlation in the inputs does not establish causation in the world, and the model has no mechanism for knowing which of its correlated features is doing the work. A planner who reads the attribution as a causal statement will size infrastructure for a future that the model has described but not explained. The output is a map with no legend, and the legend is what the decision actually requires.
What the Machine Cannot Tell You It Doesn’t Know
There is a technical name for the property at stake: calibration. A well-calibrated system knows the difference between what it can predict and what it cannot, and it communicates that difference. Most deployed forecasting systems are not calibrated in this sense. They are optimized for accuracy on average and for fluency in presentation, and those two objectives can actively conflict with honesty about uncertainty.
The consequence for the grid is that The people making large infrastructure decisions are receiving inputs that look more certain than they are. A human analyst who had never seen a demand curve like this one might say so. The system does not. It produces a smooth line, a confidence band that reflects only the variance it was trained on, and a recommendation. The recommendation is then cited in a rate case, a press release, or an investment thesis, and the caveat that never existed in the output never appears in the record.
Responsibility When the System Is Wrong
The question that follows is not technical. When a forecast-driven investment turns out to be mispriced, who answers for it? The model’s developers will point to the data. The utility will point to the forecast. The hyperscaler will point to its own projections, which were themselves model outputs. The regulator will point to the filing. At each link in the chain, the decision was made on the basis of an input whose uncertainty had been stripped away somewhere upstream, and no single actor can be shown to have known that the number was softer than it looked.
This diffusion is not an accident of the technology. It is what happens when a system that cannot represent its own ignorance is inserted into a decision process that depends on someone representing it. The 1965 blackout produced a clear lesson because the failure was visible and the cause could be traced. The failures now being seeded into the power system will be slower, distributed across contracts and forecasts and rate bases, and they will arrive years after the models that produced them have been retrained out of existence. The grid will still be standing. The question is whether anyone will be able to say, with evidence, why it was built the way it was — or whether the answer will be a number that looked like a fact at the time.