Phantom AI Data Center Demand: Why Only 28% of Requested Power Will Materialize
Wood Mackenzie projects that US grid operators and utilities will commit to roughly 28 percent of the 1,066 gigawatts requested for AI data center projects, according to analysis the firm published this week. More than two-thirds of the electricity sought for the AI buildout will never reach the grid. The binding constraint on AI infrastructure is not a shortage of demand; it is a surplus of phantom AI data center demand inflating interconnection queues, capex plans and power prices.
The gap between requests and commitments is the story. Wood Mackenzie's pipeline tracking counts about 600 GW of US data center projects still searching for power capacity, against 183 GW that have signed construction or electricity supply agreements with utilities. Interconnection requests outnumber firm commitments by a wide margin, and the firm expects most of the excess to dissolve during queue processing.
Phantom AI Data Center Demand Is Clogging the Grid Queue
Interconnection queues are the funnel where the gap opens up. Developers routinely file multiple requests for the same site, submit speculative proposals and treat queue positions as options rather than commitments. Grid operators cannot distinguish real projects from phantom AI data center demand, so every request enters the system as potential load. Wood Mackenzie attributes the expected 72 percent shortfall to long-shot interconnection pitches that fall away as processing advances.
The forecasting problem is structural, not incidental. Every interconnection request enters the queue as candidate load, so official queue data behaves like a registry of ambition rather than a demand forecast. Grid operators are left working with numbers they know are inflated, and load forecasts swing with each round of filings.
The distortion cuts both directions. Load forecasts built on inflated queues push utilities toward oversized transmission and generation plans, while operators that discount the queue risk underbuilding against demand that actually arrives. The result is a planning environment where the industry's headline demand figures carry little information about real load.
The pattern extends Wood Mackenzie's warnings from earlier this year, when the firm's "Breaking the speed limit" report described the AI data center race outpacing grid development and creating risk for projects, markets and consumers. The new projection quantifies that risk: roughly seven in ten gigawatts requested are unlikely to convert into firm commitments.
The Collocation Bet Has Limits
Faced with grid constraints, the industry has shifted toward generating power on site. More than 90 GW of collocated generation now sits in US interconnection pipelines, a measure of how large the bet has grown. Wood Mackenzie finds the model works only for the most sophisticated and well-capitalized hyperscalers, and that for most developers such projects will not materialize, making the approach unscalable.
For developers, the choice between grid power and collocation is a trade-off between cost and control. Grid power is cheaper and simpler but subject to queue delays measured in years. Collocation bypasses the queue but concentrates risk with hyperscalers that can absorb construction and fuel costs, and Wood Mackenzie's analysis suggests it does not scale beyond that group.
Physical bottlenecks reinforce the financial ones. Average power transformer lead times run about 128 weeks, generator step-up units take roughly 144 weeks, and some buyers face waits up to four years. Domestic manufacturing cannot keep pace with the surge, and Wood Mackenzie senior analyst Ben Boucher expects the supply constraints to persist well into the 2030s. Transmission buildouts remain five to ten years away, while the firm sees small modular reactors and fusion contributing little over the next decade. AI companies will run on the grid and the generation technology that exists today.
Power Prices Send a Contradictory Signal
Market pricing points the same way. Power prices currently sit below the level needed to incentivize major new gas generation, even as forecasts call for substantial capacity additions to support AI-led load growth. Chris Seiple, Wood Mackenzie's vice chairman for energy transition and power, has described load growth and affordability as in direct opposition in deregulated markets. Utilities face a choice between building speculatively against phantom AI data center demand and underbuilding against demand that does arrive, and neither path prices cleanly.
Consumers sit on the other side of that trade-off. Network investment has to be recovered through rates, so every speculative build carries a price tag for ratepayers when load growth and affordability collide. The firm's analysis treats that collision as the central risk of the current buildout.
Despite the bottlenecks, most data centers still prefer grid power, which typically prices more competitively than on-site alternatives, according to Wood Mackenzie analysts Ben Hertz-Shargel and Chris Seiple. The preference keeps pressure on the queue even as its conversion rate stays low.
| Power pipeline metric | Value |
|---|---|
| Power requested for US data center projects | 1,066 GW |
| Share expected to become firm commitments | ~28% |
| Projects with signed construction or power agreements | 183 GW |
| Projects still searching for power capacity | ~600 GW |
| Collocated generation in interconnection pipelines | 90+ GW |
| Average transformer lead time | ~128 weeks |
| Generator step-up unit lead time | ~144 weeks |
What the Mismatch Means for Markets
The implications reach beyond the grid. Hyperscaler power contracts, data center valuations and the $65 billion US electrical equipment market are all pricing a demand curve the grid will not deliver at the scale requested. Developers that treat queue requests as confirmed load will overpay for land, equipment and power purchase agreements. Grid planners who build for the full 1,066 GW risk stranding assets, while planners who ignore requests underserve the projects that do convert. McKinsey has separately examined the overbuilding risk, concluding that whether the projected growth materializes depends on whether demand for AI compute lasts.
Financing carries the same exposure. Power purchase agreements signed against projected load face renegotiation when that load fails to appear, and project finance tied to queue positions is vulnerable to attrition that has nothing to do with credit quality.
Equipment suppliers sit closest to the risk. Transformer orders are placed years ahead of energization, and the $65 billion electrical equipment market Wood Mackenzie now tracks is partly built on requests that will never convert.
The slowdown is already visible in construction data: developers added less data center capacity in the fourth quarter of 2025 than in the prior quarter, an early sign that the steepest growth projections may not materialize.
What Decision-Makers Should Watch
For strategists and investors, the practical read is straightforward. Treat 1,066 GW as a ceiling rather than a forecast. The 183 GW of signed agreements is committed load; the roughly 600 GW still searching for power is inventory rather than demand. Benchmark power procurement against committed load rather than queue size, and watch interconnection queue exits with the same attention as capacity announcements.
Smaller developers carry the most exposure. Collocation and dedicated power deals are open to hyperscalers with balance sheets and negotiating leverage, while speculative developers holding queue positions face stranded assets as more requests fall away. The projects that survive the queue will command premium pricing for the power they can secure.
Why this matters
The distance between requested and committed power is the most consequential number in AI infrastructure right now. Capex, financing and power contracts built on the full request pipeline will be repriced as the queue clears, and the projects that convert will capture the real scarcity value. Plan against committed load, not phantom AI data center demand.
Sources
US AI data center power demand outlook | McKinsey
Related Articles
- Texas Data Center Grid Freeze: 474 GW of AI Power Requests Hit the Approval Wall
- US Data Center Electricity Demand Rises as AI Infrastructure Spending Hits $725 Billion
- Data Center Power Coalition Launches to Tackle AI Grid Delays
✔Human Verified
Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.