NextFin News - The artificial-intelligence infrastructure boom is running into a wall it cannot buy its way out of: electricity. Kimmeridge Capital, the energy-focused private equity firm founded by Ben Dell, warns that as many as half of the data centers currently planned across the United States could face delays, as power-grid constraints and shortages of critical electrical equipment replace chip supply and capital as the binding constraint on AI expansion. The warning reframes a question that has moved quietly to the center of the AI investment thesis: the bottleneck is no longer in the server rack. It is at the substation.
Kimmeridge is an unusual messenger for a data-center warning. Founded in 2012 by Dell, Neil McMahon and Henry Makansi, the firm has raised more than $6.5 billion of limited-partner commitments and built its reputation in upstream oil and gas, not technology infrastructure. Dell chairs Commonwealth LNG, an export facility in Louisiana, and his firm's research division has published two studies since early 2024 modeling how AI-driven data-center expansion reshapes US power consumption. That energy-native vantage point is precisely what makes the warning worth weighing: the firm that profits from understanding physical energy constraints is saying the AI buildout has outrun them.
The Numbers Behind the Warning
The scale of the at-risk pipeline is concrete. Sightline Climate's 2026 data-center outlook, which underpins the warning, tracks at least 16 gigawatts of US data-center capacity slated to come online in 2026 across roughly 140 projects. Of that, only about 5 GW is visibly under construction. The remaining 11 GW sits in the announced stage with no visible construction progress, despite typical build timelines of 12 to 18 months. Sightline projects that 30% to 50% of the 2026 pipeline will be delayed or canceled.
The firm's own track record justifies caution. Of the capacity expected to come online in 2025, 26% slipped, and a further 10% quietly pushed back commercial operation dates.
"The 2025 projects were planned likely two to three years ago, predating the absolute acceleration in AI demand and today's labor and equipment shortages," Olivia Wang, a research analyst at Sightline and co-author of the report, said. "We think those slated for this year are likely to face even steeper challenges."
The binding constraints are physical, not financial. High-voltage transformer delivery times now stretch three to five years, and switchgear is sold out through 2028. Grid interconnection waits of four to ten years are becoming common in many regions.
"Grid connectivity is increasingly becoming the factor that determines which AI projects move forward and which remain on paper," Ditlev Engel, chief executive of Energy at DNV, wrote in May.
Why This Is a Structural Shift, Not a Cyclical Dip
The central judgment here is that power has become a structural constraint on AI infrastructure, not a cyclical bottleneck that will clear once supply chains catch up. Three pieces of evidence support that call.
First, the demand trajectory itself has shifted regimes. The International Energy Agency projects global data-center electricity consumption rising from 415 terawatt-hours in 2024 to roughly 945 TWh by 2030, an annual growth rate of about 15%. Goldman Sachs Research forecasts US data-center power demand jumping from 31 GW in 2025 to 66 GW by 2027. EPRI estimates data centers could consume 9% to 17% of US electricity generation by 2030. These are not demand fluctuations around a stable mean; they are a step-change in the load profile of the entire grid. For context, the US grid was engineered for load growth in the low single digits; a system that adds capacity at 1% to 2% a year cannot absorb a load category doubling inside three years without breaking something.
Second, the constraint is on the supply side of a network that cannot be expanded quickly. Transmission and distribution infrastructure has multi-year permitting, multi-year equipment lead times, and a manufacturing base that takes two to three years to expand. Hitachi Energy's new South Boston transformer plant, announced in September 2025, is not expected to be operational until 2028. A constraint with a three-to-five-year relief horizon is structural by definition. The same dynamic shows up in the interconnection queue: the average time for a data-center project to navigate the US grid interconnection process has grown from under two years in 2008 to nearly five years today, with California projects stretching beyond nine years.
Third, the industry is responding with structural adaptation rather than temporary workarounds. Developers are pursuing "bring your own power" strategies, onsite generation, microgrids, and fuel cells at a pace that signals a permanent change in operating model. The GW Ranch project in Pecos County, Texas, illustrates the scale of the bypass: Pacifico Energy secured a Texas Commission on Environmental Quality air permit for 7.65 GW of gas-fired generation, the largest such permit in the United States, to serve hyperscale data centers on a private grid with first power targeted for the first half of 2027. Public reporting indicates Amazon is behind the project. When the largest AI spenders start building their own power plants, the grid is no longer a utility relationship. It is a competitive threat.
Regulators are already reacting to the same pressure. The Decentralized Access to Technology Alternatives Act, introduced in January 2026, would exempt data centers that build fully off-grid power infrastructure from Federal Energy Regulatory Commission oversight, including interconnection rules that currently add years to project timelines. A regulatory carve-out for customers large enough to build their own generation is a marker of structural strain, not a cyclical hiccup.
The Counter-Thesis: Is the Denominator Broken?
The strongest challenge to the "half delayed" framing comes from SemiAnalysis, which argues that the headline number rests on a flawed denominator. Its case has three parts.
First, Sightline's under-construction figure of 5 GW is too low. SemiAnalysis contends that the top two hyperscalers alone exceed 5 GW of capacity under construction on a self-build basis, before counting the many gigawatts being built by third-party developers. If true, the construction rate is far higher than the 31% implied by the 5-of-16 GW arithmetic.
Second, most of what is flagged as "at risk" sits in the early-stage, pre-construction announced bucket. These are speculative megawatts that were never going to land on a 2026 timeline under rigorous analysis. In SemiAnalysis's model, they simply show up in 2028 and beyond rather than being canceled. Over the past six months, its year-end 2026 North American hyperscaler self-build forecast moved only about 1%, and its colocation forecast less than 5%.
Third, the equipment backlog that matters is locked in. Major manufacturers including GE Vernova, Hitachi Energy, and Mitsubishi Electric are booked out three to four years on their main equipment lines, and prepayments of 10% to 15% are now standard to secure a queue position. A speculative announcement that dies in a county commission hearing removes zero orders from anyone's books, and when a real order does fall away, a queue running years deep means the slot is reallocated rather than vanishing.
This counter-thesis is substantial and partially persuasive. It correctly identifies that the "half delayed" figure describes the slice of the pipeline most prone to slipping — large, publicly announced, early-stage projects — rather than the entire US data-center buildout. But it does not defeat the core point. Even if the denominator is inflated, the direction of travel is unchanged: capacity additions are slowing. Wood Mackenzie found that newly added US data-center pipeline capacity fell to 25 GW in the fourth quarter of 2025, half the third-quarter volume, with the total disclosed pipeline at 241 GW and only 33% under active development. Under-construction capacity across primary markets fell 6% year over year to 5,994 MW, the first decline since 2020. A re-timing from 2026 to 2028 is still a delay, and a delay in AI infrastructure is a delay in the revenue that depends on it.
There is also a subtler point in SemiAnalysis's own data that cuts against its conclusion. The firm notes that its year-end 2026 forecast barely moved over six months — but that is a forecast built by analysts who already strip out speculative announcements. The fact that the market's public-facing pipeline number still relies on unfiltered press releases is itself evidence of an information gap. Investors who cannot distinguish a permitted, equipment-ordered project from a land-bank announcement are pricing the two as one.
The Second-Order Trade: Delays Do Not Kill Demand, They Re-Price It
The market's first-order reading of delay warnings is bearish for the AI buildout chain: fewer data centers means less demand for equipment, construction, and power. That reading mistakes a supply constraint for a demand collapse. The second-order effect is the opposite. When capacity is constrained, value migrates to whoever controls the scarce input.
In this case, the scarce input is power-ready capacity. The beneficiaries are not the speculative developers with announcements and no grid slots. They are the operators who already hold interconnection positions, the equipment makers with booked-out backlogs and pricing power, and the energy producers sitting on the gas and generation assets that can deliver electrons on a usable timeline. The delay is not a reduction in AI's power appetite. It is a transfer of pricing power from the buyers of electricity to the sellers.
The earnings already support this. Digital Realty reported a record $1.9 billion backlog in the second quarter of 2026, a development pipeline of 1.4 GW under construction at a total cost of $20 billion, and two additional US hyperscale leases adding $410 million of annualized GAAP rent after quarter end. Vertiv posted net sales of $3.274 billion, up 24% year over year, and raised full-year sales guidance to roughly $14 billion. Equinix added 9,700 net interconnections in the quarter, its highest-ever quarterly addition, and raised full-year revenue guidance to 11% to 12% growth. These are not the numbers of a chain bracing for cancellation. They are the numbers of a chain with more demand than it can serve.
The asymmetry runs deeper. A delay hurts the marginal, under-capitalized developer who cannot post the 10% to 15% prepayment that now holds an equipment queue slot. It does not hurt the incumbent with balance-sheet depth. The same bottleneck that kills speculative announcements concentrates the industry toward players who can wire money upfront and wait out a three-year transformer lead time. Scarcity is a consolidation machine.
There is a third-order implication that the consensus has not fully priced. If delays push energized capacity into 2027 and 2028 while AI model deployment continues on its current trajectory, the industry faces a period where compute demand grows faster than the facilities that house it. That gap does not reduce AI spending; it redirects it. Capital that would have gone into new buildings goes instead into extending the life and density of existing ones, into liquid cooling retrofits, and into software that squeezes more workloads per watt. The delay is a headwind for the greenfield construction chain and a tailwind for the installed-base efficiency chain.
The warning lands into a market that has already begun repricing the buildout chain. Listed data-center operators have fallen out of favor during 2026 as investors treat them increasingly as bond proxies rather than direct AI beneficiaries, and Bank of America downgraded Digital Realty while keeping a buy rating on Equinix, which it called a "self help story." That backdrop cuts both ways: it means part of the delay risk is already reflected in valuations, but it also means the market is pricing the sector as if the growth story is intact rather than as if the grid is the binding constraint.
What to Watch
The falsifying signal for the delay thesis is specific: if more than roughly 14 GW of the 16 GW Sightline tracks actually achieves commercial operation during 2026 — that is, if commercial operation dates land within 10% of the announced target — the "up to half delayed" framing is wrong. A secondary signal would be transformer lead times compressing back below 24 months, which would indicate the equipment bottleneck is clearing faster than the manufacturing base suggests.
Into the second half of 2026, three checkpoints matter. First, whether the 11 GW of announced-but-not-under-construction capacity begins breaking ground or quietly migrates to later years. Second, whether equipment makers maintain raised guidance and order growth through the rest of the year. Third, whether hyperscaler capital commitments translate into paid revenue and energized capacity rather than remaining as press releases.
The short-term read is mixed: delays create headline risk for the buildout chain and can pressure speculative names. The medium-term read is constructive for those who control power-ready assets and equipment backlogs. The long-term read is that AI's growth ceiling is now set by the grid, and grids do not scale at the speed of software.
The AI boom was supposed to be constrained by chips, then by capital, then by talent. It turns out the constraint was electrons all along — and electrons answer to physics, not to funding rounds.
Explore more exclusive insights at nextfin.ai.
