How we use electricity data to inform analysis
Data centers are land-light, capital-heavy, power-defined assets. The binding constraint on AI capex isn't dollars or land — it's electrons reaching the rack. We use four lenses of public electricity data to evaluate every site's risk and every region's headroom:
- Capacity supply — generators by fuel + MW + location (this page's map below). Tells us what's actually available, where, and what kind of dispatchable firm capacity exists vs. variable renewables. A 1 GW DC inside the service territory of a marginal-coal-retirement grid is a different bet than one next to a 2.5 GW nuclear plant.
- Reserve margin vs. required minimum (per NERC) — how close each region is to running out of firm capacity during peak demand. PJM & NYISO are already below their required margins.
- Capacity auction prices — the cleanest market signal of scarcity. PJM jumped 9× from $28 → $269 → $329/MW-day across two consecutive auctions. MISO went $30 → $217 in one year (+650%).
- DC share of peak-load growth — for each region, what % of new MW demand is data centers vs. industrial/residential. The higher this share, the more the strain is structural rather than weather-driven.
The composite below ranks each Balancing Authority on a 0–3 strain index. Red = at risk of capacity shortfall under normal or extreme conditions through 2030. We use the same BA tag on every DC marker in the main map, so you can see at a glance which sites are in stressed-grid regions.
Data center buildout signal — off-peak load growth, current week vs 2y ago
DCs run 24/7 at near-constant load. Residential drops at night. So overnight-minimum load growth YoY is the cleanest DC signal — most of it is new industrial/DC capacity that's been turned on. Multiplied by each region's DC share-of-growth (from utility IRPs) gives an implied DC attribution.
Per-site implied utilization (vs stated capacity)
Cross-referencing BA-level off-peak growth with each marquee site's stated capacity + public energization status. Shows how much of each project's "stated GW" is actually drawing power today vs. still under construction. Validated against known-utilized sites (xAI Colossus at 100% ✓, AWS Susquehanna BTM at 63% ✓).
EIA live — hourly load across 14 Balancing Authorities (last 72 h)
Live via EIA Form 930 API. Lines = current demand. Right-most = right now. All 14 BAs covered, hourly cadence.
ERCOT live — capacity vs demand (today, 5-min interval)
No-auth real-time JSON endpoint with 5-min granularity. Blue = available capacity. Amber = demand. Gap = reserve margin.
Strain by region
Capacity supply — major US power plants + AI data centers
≥1 GW operating plants (curated subset) + 100 DC facilities. Color = fuel type · size = MW · DCs in green ring. Toggle layers (bottom right) for satellite view.
Capacity auction prices ($/MW-day)
PJM RPM and MISO PRA — the most legible price signals of capacity scarcity.
Reserve margin vs. required (2026)
Green bars = comfortable. Red bars = below NERC's required reserve margin.
Methodology & sources
Strain index (0–3 composite) per region uses four weighted inputs:
- NERC LTRA risk rating (40%): LOW=0, ELEVATED=1, HIGH=2
- Reserve margin shortfall vs. required (30%): negative shortfall = strained
- DC share of peak load growth (20%): higher = more DC-driven structural strain
- Capacity auction $/MW-day (10%): log-scaled; $300+/MW-day = severe
Data sources: NERC LTRA Dec 2025 · ISO IRP and capacity auction filings (PJM RPM, MISO PRA, ERCOT CDR, CAISO RA, ISO-NE FCA, NYISO ICAP) · individual utility IRPs (Dominion, Duke, AEP, Entergy, SRP, OPPD).
DC share of growth is sourced from each utility's IRP forecast assumptions where available, otherwise estimated from the gap between historical load growth and recent forecast. For PJM, Dominion's 2024 IRP attributes ~85% of forecast load growth in its service territory to data centers. For ERCOT, the large-load interconnection queue (380 GW, 70%+ DC) is the basis.