Data Centers: 18–36 Months. Power Grid Upgrades: 5–7 Years.
AI demand is growing faster than physical infrastructure can be built. The bottleneck isn't compute — it's concrete, steel, copper wire, and permits. The physical world can't keep up with the digital one.
18–36moData center construction timeline
5–7yrPower grid upgrade timeline
40GWNew power capacity needed for AI by 2030 (US alone)
3–4yrPower transformer lead time (currently)
Choose your depth. The data doesn't change — just the explanation.
AI needs huge buildings full of computers, and those computers need enormous amounts of electricity. Building these computer buildings takes about 2-3 years. But making the electricity system bigger — adding new power lines and transformers — takes 5-7 years. So AI companies can announce plans today, but they can't actually run all their computers for many years.
The AI compute buildout faces "atoms vs. bits" constraints. Software can scale at nearly infinite speed — physical infrastructure cannot. Data centers require land acquisition, permits, construction (18-36 months), and most critically: power connections. Connecting to the power grid requires utility upgrades, new substations, and sometimes new transmission lines — typically 5-7 years due to permitting, materials, and labor. Power transformers currently have 3-4 year lead times. The IEA projects US data centers will need 40GW of new power capacity by 2030 — roughly equivalent to adding 40 new nuclear plants in capacity.
Data center construction: CBRE "Data Center Trends H2 2024" (18-36 months for greenfield; 12-18 for brownfield expansion). Power grid: NERC "Long-Term Reliability Assessment 2024" — 5-7 years for new transmission line permitting in most US states due to NEPA review, state PUC approval, and right-of-way acquisition. Transformer bottleneck: DOE "Transforming the Grid" report — large power transformer (LPT) lead times 3-4 years; US manufactures ~70% domestically. IEA "Electricity 2024": data centers projected to consume 1,000 TWh by 2026 (double 2022). Implication: Stargate's $500B commitment requires solving permitting and power that takes longer than the 4-year timeline.
CBRE Data Center Trends: cbre.com/insights/books/h2-2024-north-america-data-center-trends. IEA Electricity 2024: iea.org/reports/electricity-2024. NERC Reliability Assessment: nerc.com/pa/RAPA/ra/Reliability%20Assessments%20DL/NERC_LTRA_2024.pdf. DOE Transformer Report: energy.gov/oe/articles/doe-releases-report-large-power-transformer-supply-chain. Power capacity: track FERC interconnection queue (ferc.gov/industries-data/electric/electric-power-markets/electric-queue) — 1,700GW of generation capacity waiting for grid connection. Monitor: Data center construction starts (CBRE, JLL), utility interconnection agreements, power purchase agreement (PPA) activity for hyperscalers.
The Timeline Mismatch
AI demand is growing exponentially. Physical infrastructure grows linearly. The gap between what AI companies want to build and what they can physically power is widening.
AI Infrastructure Build Timeline vs. Demand Growth
Time required for each infrastructure component vs. AI power demand growth trajectory.
Data Center (greenfield)
18–36mo
Land, permits, construction, fit-out. Fastest AI constraint to solve.
Power Substation
2–3yr
Utility upgrade to deliver power to new data center. Often the limiting factor.
Large Power Transformer
3–4yr
Lead time for LPTs due to manufacturing constraints. US/Europe limited supply.
Transmission Line
5–7yr
New HV transmission requires FERC, state, and local permits. Longest bottleneck.
New Power Plant (gas)
3–5yr
Natural gas peaker plants fastest option. Nuclear: 15+ years.
Cooling Water Rights
1–3yr
Permitting for water withdrawal at scale. Often overlooked constraint.
US Data Center Power Consumption (GW)
IEA projections. 2023 baseline to 2030 demand.
FERC Interconnection Queue (GW waiting)
Generation capacity waiting to connect to US grid. Bottleneck indicator.
The Power Paradox
Hyperscalers are announcing AI data center buildouts faster than the power grid can possibly support them. The Stargate project's 100,000 MW of planned capacity would require more power than the entire state of Texas currently consumes. The physical infrastructure to deliver that power takes 5-7 years to build — longer than the AI hype cycle has existed.
Sources
• CBRE. "H2 2024 North America Data Center Trends." cbre.com.