On the interconnection side, average queue times across major U.S. grids have extended to roughly eight years, up from under two years in the early 2000s
The Breaking Point
For most of the last decade, energy was a cost line. Today it is a rate limiter. Global data centre electricity consumption reached approximately 415 TWh in 2024, representing around 1.5% of world electricity output. That figure is projected to more than double to roughly 945 TWh by 2030, driven almost entirely by AI-intensive workloads and cloud compute expansion. This is not gradual drift — it is a step-function shift that grid planning cycles, interconnection queues, and utility investment timelines were never designed to absorb at this pace.
The hardware composition driving that demand has also changed structurally. AI-optimised servers are on track to represent 44% of total data centre power consumption by 2030, up from a meaningful but smaller share today. That single number — nearly half of all facility power flowing through compute hardware designed for AI — reframes the energy challenge from operational to architectural. Efficiency gains at the cooling or facility layer will not close the gap if the hardware profile continues its current trajectory.
Where the Shift Accelerated
Two forces arrived simultaneously and compounded each other: demand growth outran grid capacity, and the cost of that imbalance became visible in capacity markets.
The PJM capacity auction for the 2026/2027 delivery year cleared at the federal price cap — approximately $329 per megawatt-day — driven materially by data centre load forecasts. That result is not a pricing anomaly to be hedged away. It is a market signal that one of the most important grid regions for U.S. data centre operations has structurally tightened and is being priced at the ceiling. Capacity costs feed directly into utility rates, moving financial pressure downstream to every customer on the grid and generating political and regulatory attention that did not exist at lower clearing prices.
On the interconnection side, average queue times across major U.S. grids have extended to roughly eight years, up from under two years in the early 2000s. Northern Virginia — still the largest data centre market in the world — sits at the extreme end, with interconnection timelines running four to ten years in some cases. The arithmetic is unworkable for AI compute deployments operating on 18-to-36-month project horizons. Projects that cannot solve the power timeline problem either stall, relocate, or absorb the cost and risk of behind-the-meter generation.
That third option is accelerating. Natural gas turbines, fuel cell systems, and hybrid microgrids are being contracted at multi-gigawatt scale to bypass interconnection delays. The shift toward on-site generation introduces a new layer of operational complexity: commodity price exposure, emissions planning, and evolving state-level regulatory obligations that vary sharply by jurisdiction.
Where This Hits Global Heads of Data Center Energy
The operational exposure is distributed across three dimensions simultaneously.
Procurement cost and contract structure. The PJM auction result is a leading indicator, not an isolated data point. Any capacity market serving a major data centre cluster is now subject to similar dynamics as load growth forecasts are revised upward. Long-term PPAs became the primary instrument for managing this exposure: data centre operators contracted more than 17 GW of new renewable energy capacity through long-term agreements globally in 2024, a record for the sector. Physical PPAs that align delivery location with load remain the most direct hedge against basis risk; VPPAs provide financial offset but leave the operator exposed to local capacity and energy pricing swings. The procurement function must now model capacity cost pass-throughs — not just energy price — as a core variable in financial planning.
Interconnection queue position and site strategy. A queue position held today in a constrained market is an asset; one not yet secured is a multi-year liability. Power availability must precede land availability and network connectivity in the site evaluation sequence, not follow them. Markets with shorter queue times and available substation capacity — even if they carry a connectivity or cost premium — are undervalued relative to constrained hubs where stranded capacity risk is real. Geographic diversification across jurisdictions, including Nordic markets with hydropower and cooler climates, is now a risk management tool, not just a cost optimisation.
Nuclear and baseload contracting. Microsoft’s 20-year agreement with Constellation Energy for output from the restarted Three Mile Island Unit 1 reactor, expected online in the late 2020s, marked a shift in how hyperscalers think about baseload reliability. Google’s partnership with Kairos Power targets up to 500 MW from advanced reactors by 2035. Meta has structured agreements spanning existing nuclear plants and SMR development. These deals set a reference frame for what long-duration, carbon-free baseload contracting looks like at scale. For operators without a nuclear strategy, the question is not whether to engage but at what point in the SMR development pipeline a commitment is worth the development risk.
What Could Still Change the Read
The demand projections underpinning this analysis carry meaningful uncertainty. The 945 TWh figure for 2030 reflects current AI workload trajectories; material improvements in model efficiency — smaller architectures, inference optimisation, hardware-software co-design — could compress that number. The industry has historically underestimated efficiency gains even as it has underestimated demand growth. Both can be true simultaneously, but the net direction for power demand remains upward under any credible scenario.
Regulatory outcomes are genuinely open. Tariff redesigns in multiple jurisdictions could shift how grid upgrade costs are allocated between large loads and general ratepayers. FERC’s proposed rulemaking to accelerate interconnection study timelines — targeting 60-day windows with co-located generation commitments — could, if implemented effectively, begin to reduce queue friction by the late 2020s. Texas Senate Bill 6’s curtailment mandates for large customers add compliance obligations that are still being operationalised. None of these regulatory trajectories is confirmed as resolved; each represents a live variable in the cost and risk model.
The behind-the-meter generation buildout also carries regulatory tail risk. Permitting conditions, emissions standards, and state-level interconnection rules for on-site generation are evolving in ways that are not yet uniform across jurisdictions. An operator who contracts modular gas turbine capacity at scale today faces regulatory assumptions that may not hold across all deployment geographies by the time projects reach commissioning.
The Question This Leaves Your Team
If interconnection timelines in your priority markets exceed your planned deployment horizon, and behind-the-meter generation introduces the regulatory and commodity risk you would otherwise avoid, which sites in your current pipeline carry stranded capacity exposure — and does your board understand that the power constraint is the binding constraint, not capital or land?
Sources
- Meer — When AI destroys value before creating it (Link)
