The number that's been circulating in grid planning circles is staggering: 1,066 gigawatts of electricity requested by data center projects from U.S. grid operators and utilities. That's roughly a thousand nuclear reactors' worth of new demand queued up in interconnection applications. Wood Mackenzie's new projections, reported by Bloomberg, expect utilities to actually commit to about 28% of that — roughly 300 gigawatts — because the rest is phantom load: speculative projects, long-shot pitches, and developers holding multiple grid positions as options rather than commitments.
Here's the civilizational paradox: the skeptics who say "most of this demand will never materialize" are right about the queue. And the maximalists who say "the grid is nowhere near ready for what's coming" are also right. Both things are true simultaneously, and understanding why requires separating three distinct problems that keep getting collapsed into one argument.
The first problem is the phantom queue — speculative demand that inflates interconnection requests. The second is the real demand gap — the verified, contracted, under-construction load that genuinely outpaces grid capacity. The third is the physical bottleneck — not just generation, but transmission, equipment, and the timeline mismatch between what AI infrastructure needs and what the grid can deliver. Each of these has a different solution. Conflating them produces bad policy, bad investment, and a lot of wasted argument.
The Phantom Queue Is Real, But It Doesn't Mean the Problem Is Solved
When Wood Mackenzie says 72% of requested data center electricity "won't materialize," that's not a reassuring headline — it's a diagnostic. Interconnection queues have always been clogged with speculative applications, but the scale of AI-driven requests has pushed this to a new extreme. Developers file for grid connections the way venture capitalists spray seed checks: most won't convert, but you need the option.
Grid interconnection queues reached over 2,200 GW of planned capacity in mid-2026, according to a Reuters Events whitepaper — far exceeding the nation's total installed capacity of roughly 1,400 GW. That ratio alone tells you the queue is not a demand forecast. It's a land grab.
But here's what the "phantom demand" framing obscures: the 28% that does materialize is still a civilizational-scale infrastructure challenge. If Wood Mackenzie's projection holds, that's roughly 300 GW of new data center load — and it's landing on a grid that took decades to build the capacity it already has. The DOE's draft 2026 National Transmission Needs Study is unambiguous on this point: AI data centers have become the dominant driver of electricity demand growth, compelling planners to reconsider where and how the grid expands. Historically, transmission projects were driven by reliability concerns and aging infrastructure. Now, rapid load growth is the principal force — and it's concentrated in specific geographies rather than spread evenly across the system.
The DOE study cites projections from Lawrence Berkeley National Laboratory showing data center electricity demand growing at a 13% to 27% compound annual rate through 2028, when data centers could consume between 6.7% and 12% of all U.S. electricity. EPRI's separate estimate puts data centers at up to 9% of U.S. electricity generation by 2030, up from 4% in 2023. These ranges reflect genuine methodological differences between agencies — Lawrence Berkeley's higher-end scenario assumes faster AI chip deployment than EPRI's baseline — but both point in the same direction. The demand is real. The grid is not ready.
The Supply Gap Has a Specific Shape — and It's Not Just About Megawatts
Bank of America's semiconductor analysts, whose projections Utility Dive reported on, put the supply gap in concrete terms: data center demand will outpace planned utility capacity additions by more than 100 GW through 2030. That's not a rounding error. That's the equivalent of dozens of large power plants that need to exist but aren't yet built, permitted, or in some cases even proposed.
The shape of the gap matters as much as the size. Utilities have revised demand forecasts upward in each of the past three years as AI-related electricity demand materialized faster than their models predicted. The problem isn't that planners are incompetent — it's that AI infrastructure deployment is moving at software speed while grid infrastructure moves at geology speed.
Large gas turbines, the preferred technology for flexible dispatchable power, have manufacturing capacity largely committed through 2030. New units can take years to enter service after shipment. BofA's analysts note that this has increased interest in natural gas reciprocating engines from manufacturers including Caterpillar, INNIO, Rolls-Royce, and Wärtsilä — smaller, faster-to-deploy units that can respond rapidly to changing loads. That's a workaround, not a solution.
Meanwhile, the DOE study cites NERC forecasts showing total U.S. electricity consumption rising from 4,281 TWh in 2024 to 5,353 TWh in 2034 — a 25% increase in a decade. EPRI's range is wider: 30% to 46% growth between 2020 and 2035, depending on how quickly data center development accelerates. When agencies with different methodologies converge on "significant growth" while disagreeing on the magnitude, the honest answer is: plan for the high end and build in flexibility.
The DOE itself frames the historical context clearly: the last time the U.S. experienced rising electricity demand of this magnitude was before the early 2000s, when demand grew up to 30% driven by economic expansion and consumer adoption of electric products. That era was followed by two decades of flat demand. We are not in that era anymore. The efficiency gains from LED lighting and distributed solar that suppressed demand growth through the 2010s are now being overwhelmed by a single technology sector's appetite.
The Physical Bottleneck Is Transmission, Not Generation — and It's Worse Than the Numbers Suggest
Here's the constraint that doesn't get enough attention: even if you could conjure the generation capacity overnight, you couldn't deliver it. Transmission infrastructure takes six to seven years to build for long-distance lines, and some projects face permitting timelines stretching to 15 years. The DOE's transmission study identifies needs across 20 regions and points to some of the largest transmission portfolios in history being approved: MISO's $21.8 billion long-range transmission portfolio, SPP's $7.7 billion investment program, and Texas' $33 billion commitment to its grid. Investor-owned utilities are planning $1.1 trillion in grid investments through 2029. These are announced commitments, not operational capacity — the distinction matters enormously.
The McKinsey Global Institute, cited in a MindCast AI analysis tracking federal-state grid dynamics, reported equipment lead times more than doubling since 2019 and grid connection waits exceeding four years. FERC responded in June 2026 by ordering all six FERC-jurisdictional regional grid operators to justify or reform how large loads connect to the grid, naming cost shifting and transmission cost transparency explicitly. That's a regulatory acknowledgment that the interconnection process itself is broken — not just slow, but structurally misaligned with the load growth it's supposed to accommodate.
The physical stress isn't limited to the grid's edges. Bloomberg reported that rapid swings in AI data centers' power demands are straining vital equipment inside the facilities themselves — batteries, generators, and cooling systems malfunctioning or wearing out far sooner than expected. AI workloads at times see power usage spike as much as 50% above design capacity. This creates a feedback loop: volatile demand damages infrastructure, damaged infrastructure reduces reliability, reduced reliability forces more redundancy, more redundancy means more peak demand. The grid's problem and the data center's problem are the same problem.
The Workaround Economy Is Already Here — and It's Both the Solution and the Warning Sign
When the grid can't deliver, builders route around it. As of early 2026, data center developers were planning roughly 56 GW of on-site capacity, accounting for 30% of all planned data center builds nationwide. Texas leads with over 20.6 GW of planned behind-the-meter capacity, followed by New Mexico at 9.2 GW, Pennsylvania at 7.5 GW, and Utah at 6.0 GW. More than 7.5 GW of data center projects with on-site generation are already under construction, with another 60 GW-plus in pre-construction, per BofA's analysis.
SPP's response to this is instructive: the grid operator introduced its High Impact Large Load Generation Assessment framework, offering connection agreements in under 90 days for data centers that bring their own generation to support their load. That's a market signal dressed up as a policy — the grid is telling large loads: help yourself first, then connect.
Virtual power plants and demand flexibility tools are also scaling fast. A Brattle Group study cited in the Reuters Events whitepaper estimates that flexible distributed energy resources — smart thermostats, battery storage, demand response — could unlock up to 200 GW of demand response capacity across utilities and wholesale markets. I covered the 37.5 GW virtual power plant milestone [back in July](/p/the-grids-demand-problem-just-got-a-software-answer-and
