Power Constraints: The Critical Bottleneck for AI Infrastructure ROI

Date: November 2025

Executive Summary

Power availability has emerged as the binding constraint on AI infrastructure deployment. Despite $2.8T projected capex through 2029, physical limits on electricity generation and distribution may cap achievable returns.

Key Metrics: - 55 GW new capacity required by 2030 (Citi, 2025) - $2.8T global power infrastructure investment needed ($1.4T U.S. alone) - 84% of data center operators cite power as top-3 site selection factor - 2-10 years approval timelines in constrained regions (Virginia, Netherlands, Ireland)

1. The Capacity Gap

1.1 Demand Projection

AI training and inference workloads require:

[ ^{AI}_t = ]

Typical Ratios: - 0.5-1.0 MW per $1M in data center capex (varies by PUE and GPU density) - H100 GPU: 700W peak power - A100 GPU: 400W peak power - Liquid cooling: Reduces power by 20-30% vs. air cooling

Example (Microsoft FY2024): - AI Invested Capital (ending): $111B - Implied Power: 111,000 × 0.75 MW/$M / 1,000 = 83.3 GW

1.2 Supply Reality

AWS Disclosure (2024): Added 3.8 GW in past 12 months Industry Need: 55 GW by 2030 (Citi) Current Annual Rate: ~4-5 GW/year → 45-50 GW over 10 years (shortfall of 5-10 GW)

Constraint: If demand outpaces supply, AI capex becomes stranded capital (underutilized assets).

2. Power Purchase Agreements (PPAs)

2.1 Recent Nuclear Deals

  1. Microsoft + Constellation Energy (September 2024)
    • 20-year PPA for Three Mile Island Unit 1 reactor (Pennsylvania)
    • Capacity: ~800 MW
    • Restart timeline: 2028
  2. AWS + Talen Energy (March 2024)
    • 960 MW from Susquehanna nuclear plant (Pennsylvania)
    • Direct connection to data center campus
  3. Google + Kairos Power (October 2024)
    • First corporate agreement for small modular reactors (SMRs)
    • 500 MW by 2030

2.2 Extraction from 10-K

Commitments & Contingencies Footnote:

Purchase Commitments (in millions):
  Power purchase agreements         $18,500  (over 15 years)
  Average annual commitment:          1,233
  Undiscounted total:                18,500
  Present value (disc. 5%):          12,800
  Capitalized ROU asset:              8,200  ← Add to IC^AI

Capacity Calculation: [ = ]

If $1,233M/year at $50/MWh: [ = = 2.82 ]

3. Utilization Rate Mystery

3.1 Industry Estimates

Reported: Hyperscalers claim “capacity constrained” (implying near-100% utilization)

Measured (academic studies, industry surveys): - 85%+ of GPU capacity sits idle - 10-15% typical utilization in hybrid research/production systems - 40% utilization considered good vs. industry average

3.2 Measurement Methods

nvidia-smi (NVML): Measures “percent of time kernel executing” — misleading - Can show 100% while doing zero computation (memory transfers)

DCGM (Data Center GPU Manager): Advanced metrics - SM (Streaming Multiprocessor) Activity: % time SM units active - SM Occupancy: How fully SMs occupied when active - SM Pipe Utilization: Utilization of specific computation pipelines

Key Insight: SM efficiency reveals true computational waste invisible to nvidia-smi.

3.3 Financial Impact

At 15% utilization, effective cost per useful GPU-hour:

[ = = 6.7 ]

Example: $1B in GPUs at 15% utilization = $6.7B effective capex for ROIC calculation.

Adjustment: [ IC^{AI, }_t = ]

[ ROIC^{AI, } = ]

4. Power-Constrained Growth Scenarios

4.1 Feasibility Check

Test: Can disclosed PPA capacity support implied AI power demand?

[ ^{}_t = ]

Decision Rules: - < 80%: Adequate headroom - 80-100%: Tight but feasible - > 100%: Power-constrained (must reduce growth or add capacity)

4.2 Growth Rate Adjustment

If power-constrained, terminal growth (g) must be capped:

[ g^{} = g^{} ]

Example: Analyst forecasts (g = 8%), but PPA covers only 60% of implied demand: [ g^{} = 0.08 = 4.8% ]

Impact on Terminal Value: [ = ]

Reducing (g) from 8% to 4.8% lowers terminal multiple by 15-25% (depending on WACC).

Bloom Energy 2025 Report: - 2024: 13% of facilities use some onsite generation - 2030 Projection: 38% use some onsite; 27% fully powered onsite (27× increase)

Drivers: 1. Grid connection delays (2-10 years in constrained regions) 2. Reliability concerns (baseload power for 99.99% uptime SLAs) 3. Cost arbitrage (natural gas onsite cheaper than grid in some markets)

Technologies: - Natural gas turbines: 10-50 MW per unit - Fuel cells (Bloom Energy): 300 kW per unit, modular - Small modular reactors (SMRs): 50-300 MW, timeline 2028-2032

6. Regional Constraints

6.1 U.S. Markets

Northern Virginia (Loudoun County): World’s largest data center concentration - Constraint: Dominion Energy substation capacity - Timeline: 3-5 years for new substations - Impact: Microsoft, AWS expanding to Ohio, Iowa as alternatives

Texas (Dallas, Austin): - Advantage: ERCOT deregulated market, faster interconnection - Risk: Grid stability (2021 blackouts), summer peak demand

6.2 Europe

Ireland (Dublin): Amazon, Google, Microsoft presence - Moratorium: No new data center connections until 2028 (EirGrid) - Reason: Data centers consume 18% of Ireland’s electricity

Netherlands (Amsterdam): Historical hyperscaler hub - Moratorium: New data center ban in Amsterdam until 2028 - Alternative: Frankfurt, Stockholm

7. Implementation: Power Constraint Checker

7.1 Algorithm

def check_power_constraint(
    ic_ai: float,  # AI invested capital ($M)
    ppa_capacity_gw: float,  # Disclosed PPA capacity (GW)
    power_ratio: float = 0.75,  # MW per $M capex
    tolerance: float = 0.80  # Utilization threshold
) -> Dict:
    implied_demand_gw = (ic_ai * power_ratio) / 1000
    utilization = implied_demand_gw / ppa_capacity_gw if ppa_capacity_gw > 0 else float('inf')

    if utilization <= tolerance:
        status = "adequate"
    elif utilization <= 1.0:
        status = "tight"
    else:
        status = "constrained"

    max_growth_factor = min(1.0, ppa_capacity_gw / implied_demand_gw)

    return {
        "implied_demand_gw": implied_demand_gw,
        "ppa_capacity_gw": ppa_capacity_gw,
        "utilization": utilization,
        "status": status,
        "max_growth_factor": max_growth_factor
    }

7.2 Integration with Terminal Value

If status == "constrained", apply max_growth_factor to terminal growth:

[ g^{} = g ]

Recompute EV/EBITDA with adjusted (g).

8. Data Sources

10-K/10-Q Disclosures: - Risk factors mentioning “power availability”, “electricity supply” - Commitments & contingencies → PPA details - Property, plant & equipment → Data center locations, capacity

Utility Filings: - Interconnection queue data (state public utility commissions) - EIA (Energy Information Administration) regional electricity rates

Industry Reports: - Bloom Energy: Data center power surveys - Uptime Institute: Data center trends - 451 Research / S&P Global Market Intelligence

Press Releases: - Nuclear PPA announcements (Microsoft, AWS, Google) - Data center groundbreaking / expansion announcements

9. Citations

Bloom Energy. (2025). 2025 Data Center Power Report. https://www.bloomenergy.com/

Citigroup Research. (2025). Big Tech AI Spending Forecast.

EirGrid. (2024). Dublin Data Center Grid Connection Moratorium.

Goldman Sachs Research. (2024). Gen AI: Too Much Spend, Too Little Benefit?

U.S. Energy Information Administration (EIA). State Electricity Profiles. https://www.eia.gov/

Module: src/ai_roi/power_analyzer.py Test Coverage: Regional constraint scenarios, PPA extraction accuracy