Wood Mackenzie: 72% of US AI Data Center Power Requests Are Phantom Load
Wood Mackenzie projects US grid operators will honor roughly 28% of 1,066 gigawatts in data center power requests, leaving ~768 GW of hyped AI energy demand that will never touch the grid.
US grid operators are on track to serve less than a third of the electricity demanded for the AI buildout. The energy scarcity narrative just got a major hole blown in it.
Key takeaways
- Wood Mackenzie projects US grid operators will commit to roughly 28% of the 1,066 gigawatts requested for data center projects, leaving an estimated 768 GW of "phantom load" that will never materialize.
- Data center developers have been submitting duplicative, non-binding requests to multiple utilities simultaneously, flooding interconnection queues with speculative bids that have no realistic path to construction.
- For Bitcoin miners, this deflates the primary political argument used to crowd them off the grid: AI's energy appetite was never as voracious as advertised.
Wood Mackenzie projects that US grid operators will commit to serving roughly 298 gigawatts of the 1,066 gigawatts requested for data center projects, first reported by Bloomberg on August 12, 2026. That leaves approximately 768 GW of requested load, by the arithmetic of the underlying Wood Mackenzie inputs, that will never be drawn from the grid. Against a current US data center capacity of roughly 24 GW, the gap between hype and physical reality is the story.
How 1,066 Gigawatts Became a Fiction
The mechanism is straightforward. Data center developers, facing uncertainty over which utility could deliver power fastest and cheapest, submitted requests to multiple utilities simultaneously. Each request counted as a real demand signal in interconnection queues. None required meaningful financial commitment upfront.
The result: a queue inflated with speculative, duplicative bids that utilities, grid planners, and policymakers then treated as real load.
Wood Mackenzie's Q4 2025 pipeline report put the disclosed US data center pipeline at 241 GW at end of 2025, with only 33% under active development. New capacity added to project pipelines in Q4 2025 came in around 25 GW, roughly half the Q3 pace. Projected capex from the largest developers was already set to decelerate in 2026 versus 2025, the first such slowdown since 2023.
An April 2026 Wood Mackenzie press release corroborates the directional picture: 600 GW was still searching for power agreements against 183 GW that had secured them, a firm-commitment rate of roughly 23%. That figure aligns closely with the new 28% projection.
Ben Hertz-Shargel, Global Head of Grid Edge at Wood Mackenzie, flagged the structural problem in the Q4 2025 report: "Much of the planned capacity belongs to new developers with a small number of massive, speculative projects, disproportionately targeting the South and Southwest." On the grid-side mismatch, Hertz-Shargel noted in Wood Mackenzie's "Up, up and away" report, via Data Center Dynamics, that PJM utilities have committed to serve twice as much new large-load demand as there is new power generation currently planned to support it.
Utilities are now fighting back. They are requiring substantial application deposits, demanding top-tier credit ratings, and collecting collateral to filter speculative bids. The 474 gigawatts of data center requests currently under scrutiny in Texas is a direct expression of the same dynamic playing out at the state level.
The Malinvestment Running Beneath the Hype
This is Austrian business cycle theory in megawatts. Cheap capital allowed frontier AI labs and speculative developers to submit non-binding grid requests at near-zero marginal cost, flooding interconnection queues with junk. Utilities, ratepayers, and grid planners built infrastructure assumptions on that junk.
The misallocation is also physical: transmission capacity studies, substation upgrades, and generation planning decisions were all shaped by a demand signal that was always mostly fiction.
Sightline Climate's February 2026 Data Center Outlook put the problem in similar terms, estimating 30 to 50% of 2026 pipeline capacity may not come online across 777 large projects and 190 GW tracked.
The realistic Wood Mackenzie buildout scenario has US data center capacity reaching roughly 110 GW by 2030, per its April 2026 public figures. That is real growth. It is also less than one-sixth of what the raw interconnection queue implies.
For Bitcoin miners, the second-order effect matters more than the headline. The "AI will eat Bitcoin's lunch on energy" argument has been the primary policy weapon used to crowd miners out of interconnection queues and power purchase agreements. Legislators, ESG campaigners, and hostile grid operators leaned on projected AI demand as cover.
If 72% of that demand is phantom, miners are not competing with real load. They are competing with press releases. Bitcoin mining's honest, interruptible load profile, which either pays for power or shuts off with no phantom bids, is the more rational grid resource by comparison. The DOE's selection of Amentum for a 1 GW AI data center at Savannah River and OpenAI's push into direct power trading for its Stargate portfolio show the frontier players are waking up to the same reality: securing firm power is the actual constraint, not filing a request.
What to Watch
The falsifiable trigger here is Wood Mackenzie's quarterly pipeline updates and FERC interconnection queue data. If the 28% commitment rate revises materially upward in future reports, or if a meaningful share of the 768 GW in phantom requests transitions to firm grid agreements within 24 months, the deflation thesis weakens. Until then, the signal from the data is consistent: the AI energy buildout is tracking far below the numbers that shaped the policy conversation.
Sources
- Wood Mackenzie Q4 2025 Data Center Pipeline Press Release
- Wood Mackenzie April 2026 US Electrical Equipment Market Press Release
- Sightline Climate Data Center Outlook, February 2026
- First reported by Bloomberg, August 12, 2026, citing Wood Mackenzie
Frequently Asked Questions
Developers submit interconnection requests to multiple utilities simultaneously, hedging on which can deliver power fastest. Each request enters the queue as a real demand signal, but most carry no binding financial commitment. Until recently, the cost of filing was low enough that speculative submissions were rational. Utilities are now requiring deposits, collateral, and credit-rating thresholds to filter the noise.
Forecasters typically model from gross interconnection queue data rather than net firm commitments, which inflates their demand baselines. Efficiency gains at the model and chip level also introduce uncertainty on realized consumption per workload. The distinction between announced pipeline capacity and actual power drawn from the grid is where the forecasts diverge from the Wood Mackenzie commitment-rate analysis.
Directly, yes. The political case for crowding Bitcoin miners out of interconnection queues rested on the premise that AI data centers represented urgent, unavoidable demand that had to be prioritized. With roughly three-quarters of that demand now characterized as phantom load by Wood Mackenzie, the scarcity argument loses its foundation. Bitcoin mining's interruptible, price-responsive load profile looks more valuable as a grid resource, not less, against this backdrop.
