Energy-AI Siting Score

Problem solved

Hyperscaler real estate teams, neocloud build-out leads, and utility load-planning groups need a single comparable score across candidate jurisdictions that captures the five binding constraints that actually decide siting in 2026: firm-power availability, water cost, network latency to user populations, AI talent depth, and policy stability over a 10-to-15-year asset life.

Inputs

• Available firm capacity and interconnection queue depth (FERC, NERC, ISO/RTO interconnection reports)
• Industrial electricity tariff and behind-the-meter PPA pricing (EIA 861, ICE forwards)
• Water cost and stress score (USGS, WRI Aqueduct, regional utility tariffs)
• Latency to top metro user populations (RIPE Atlas, CAIDA, peering exchange data)
• Talent depth: AI engineering employment density, university CS program output (BLS OEWS, IPEDS)
• Policy stability: tax stability, permitting predictability, sovereign or sub-sovereign legal risk
• Site-readiness: brownfield versus greenfield, transmission proximity, gas-pipeline access

Outputs

• 0-to-100 composite Energy-AI Siting Score by jurisdiction
• Five-pillar breakdown: firm power, water, latency, talent, policy
• Sensitivity table: score under each pillar weighted plus or minus 10 percentage points
• Pareto frontier across the top 20 candidate jurisdictions
• Constraint flags: jurisdictions failing one binding criterion regardless of composite score

Method

1. Step 1. Define the candidate set. Filter to jurisdictions with at least 500 MW of available interconnection capacity within a 36 month interconnect timeline.
2. Step 2. Score firm power (weight 0.30 default). Combine available MW, interconnect queue position, and the ratio of firm to intermittent generation. Score 0 to 100.
3. Step 3. Score water (weight 0.15 default). Combine all-in water cost per gigawatt-year of cooling, the WRI Aqueduct stress index, and the existence of a water-rights regime that prices stress correctly. Penalize jurisdictions with WRI Aqueduct extreme stress flags regardless of price.
4. Step 4. Score latency (weight 0.20 default). Measure RTT to a defined basket of top-N metros relevant to the workload (consumer inference, enterprise inference, or training-only). Training-only deployments downweight latency.
5. Step 5. Score talent (weight 0.15 default). Combine AI engineering employment density, CS program output, and the existence of major training labs or hyperscaler campuses within 200 km. Talent matters more for build-and-operate than for build-and-lease.
6. Step 6. Score policy stability (weight 0.20 default). Combine tax-incentive stability (ten-year horizon), permitting predictability, sub-sovereign legal risk, and energy-policy reversal risk. Apply a sovereign-stability cap.
7. Step 7. Combine into the composite. Normalize each pillar to 0-to-100, weight, sum. Run the Pareto-frontier check across the top 20.
8. Step 8. Apply constraint flags. Any jurisdiction that fails water-stress, sovereign-stability, or interconnect-availability gates regardless of composite score is flagged separately.

Assumptions

• Default pillar weights reflect the 2026 binding constraints for North American training and inference deployments. Engagement-specific weights (Asia, Europe, sovereign-AI) are calibrated upfront.
• Firm power means dispatchable power on demand. Solar plus storage is treated as firm only when storage capacity supports the workload duration profile.
• Latency thresholds depend on workload class. The framework uses 50 ms RTT for consumer inference, 100 ms for enterprise inference, and ignores latency for training-only.
• Policy-stability scoring is a structured judgment exercise, not a regression output. It is documented and contestable.
• The composite is a comparative scoring tool, not a financial NPV. NPV ranking requires a separate capex-and-PPA model.

Limitations

• Interconnect queue data is published with a lag and reflects intentions rather than commitments. The framework discounts speculative queue positions.
• Talent density is observed at MSA resolution, which misses sub-metro clustering. The framework supplements with named-employer proximity.
• Policy stability is the most contested pillar. Two analysts running the same inputs can produce a 15-point divergence on this pillar alone. The framework reports both and explains the divergence.
• Cross-border deployments (US-Canada, US-Mexico) carry currency, tariff, and trade-policy risks that the score does not capture and that must be modeled separately.

Example application

Applied to a 2026 hyperscaler decision between Quebec hydropower siting and ERCOT siting for a multi-gigawatt training campus. The Siting Score runs the eight steps on Quebec, Ontario, ERCOT West, ERCOT South, and PJM-West. Quebec wins on firm power and water, but carries a Hydro-Quebec tariff-risk flag and a Bill 69 policy reversal risk that the score surfaces explicitly. The brief uses the breakdown to argue that Quebec dominance is conditional on a long-form PPA structure that locks the energy price for the asset life. See Quebec hydropower and the new gating of AI compute.