AI Infrastructure Stress Index (AISI) 2026
"We don't forecast AI power demand. We audit the stress between demand narratives and grid reality."
Measuring structural tension in the AI-energy nexus. AISI synthesizes demand-supply gap, forecaster disagreement, interconnection queue delays, and wholesale price stress into a single composite score — revealing how vulnerable the current "AI buildout is on track" consensus is to grid bottlenecks, permitting delays, and energy price shocks.
AISI 46/100 — The "AI buildout is on track" consensus shows moderate structural tension due to persistent interconnection delays and wide forecaster disagreement.
Demand Gap
13.5%
Forecast Spread
31 GW
Queue Wait
5yr
QUICK ANSWER · AS OF Jan 15, 2025
What is the AI infrastructure stress level in 2026?
The AI Infrastructure Stress Index (AISI) reads 46/100 (moderate). The demand-supply gap is 13.5% (7 GW shortfall). Forecaster disagreement spans 31 GW (CV 26.7%). Average interconnection queue wait: 5 years. Methodology: v0.1-beta.
AISI Composite
46/100
Demand Gap
13.5%
Forecast Spread
31 GW
Queue Wait
5 yr
AISI measures structural tension in the AI-energy nexus — not a forecast of power shortages.
AISI Score
46/100
moderate
Demand Gap
13.5%
7 GW shortfall
Forecast Spread
31 GW
CV 26.7%
Queue Wait
5yr
Avg interconnection
AI Infrastructure Stress Index (AISI) — Beta
AISI measures the structural tension between AI compute demand narratives and energy infrastructure reality. Unlike a simple demand forecast, AISI quantifies how fragile the "AI buildout is on track" consensus is — how vulnerable it is to a single grid interconnection delay, NERC reliability warning, or wholesale price spike that could trigger rapid repricing of AI infrastructure timelines.
AI Infrastructure Stress Index (AISI)
4/4 components live · Methodology v0.1-beta
46/100
🟡 moderate
Projected AI demand 52 GW vs 45 GW committed capacity — 13.5% gap.
4 forecasters, spread 31 GW, CV 26.7% — moderate disagreement.
Avg queue wait 5 years. 2,300 GW total backlog (generation + storage).
PJM $42.5/MWh, ERCOT $38.2/MWh. YoY +18.5%. DC premium 12%.
📐 Methodology & Data Sources
AISI synthesizes 4 independent signals into a composite measure of structural tension in the AI-energy nexus. A higher score indicates greater stress between AI compute demand growth and energy infrastructure delivery capacity.
Demand-Supply Gap (40%): score = gapPct × 2.5. Source: EIA AEO + NERC LTRA.
Consensus Dispersion (30%): score = CV% × 1.5. Source: IEA, EIA, NERC, Goldman, McKinsey, Bernstein.
Infrastructure Lag (20%): score = queueYears × 15 × 0.6 + queueRatio × 50 × 0.4. Source: DOE/LBNL.
Price & Risk Signal (10%): score = yoyChange × 3 + premium bonus. Source: PJM + ERCOT.
Composite = Σ(weight_i × score_i) / Σ(weight_i). Signal: low <31, moderate 31–55, elevated 56–75, high ≥76.
📌 Why the AI-energy nexus matters
Hyperscalers have committed $200B+ annually to AI infrastructure, but the US grid adds only ~20 GW of net new capacity per year. The gap between announced AI data center capacity and actual grid delivery creates a structural tension that most AI investment theses ignore. AISI tracks the pressure building inside that structure.
Demand-Supply Gap — Data Center Power vs Grid Capacity
The core tension: projected US data center electricity demand by 2028 versus committed grid capacity additions. DOE/LBNL projects 325–580 TWh by 2028 (up from 176 TWh in 2023). We convert to GWavg (= TWh ÷ 8760 hours) for comparability with grid capacity figures. "GWavg" is average continuous power draw, not peak load.
DC Demand 2028 (range)
37–66 GWavg
325–580 TWh
Midpoint (index input)
52 GWavg
Committed Capacity
45 GW
EIA AEO 2026
Gap
7 GWavg (13.5%)
⚠️ Unit clarification
GWavg = TWh ÷ 8760 hours (average continuous power draw over a year). Grid constraints often bind at peak load, not average. This metric captures the structural demand-supply imbalance, not instantaneous grid stress.
Source: DOE/LBNL — Data Centers and AI (Jan 2025) + EIA Annual Energy Outlook 2026 · as of Jan 15, 2025 · GWavg = TWh/8760 (average continuous power, not peak load). Range reflects uncertainty in AI adoption pace and efficiency gains. "Data center" includes all workloads, not only AI.
Consensus Dispersion — Public Forecaster Disagreement
How much do public institutions disagree on data center power demand? Only freely accessible, clickable sources are used as index inputs. A high coefficient of variation signals deep structural uncertainty — the market has not converged on a consensus scenario.
| Forecaster | DC Demand 2028 (GWavg) |
|---|---|
| IEA (Conservative) (Low) | 35 |
| DOE/LBNL (Low) | 37 |
| NERC | 55 |
| DOE/LBNL (High) (High) | 66 |
Spread
31 GWavg
CV
26.7%
Mean (arithmetic)
48.25 GWavg
Source: Compiled from IEA, DOE/LBNL, NERC — all public, no auth required · as of Jan 15, 2025 · All sources are free, public, and clickable.
Infrastructure Lag — Interconnection Queue
The US grid interconnection queue is the physical bottleneck of the AI buildout. Projects must wait years for grid connection approval — a structural constraint that no amount of capital spending can bypass in the short term.
Queue Backlog
2,300 GW
gen + storage
Avg Queue Wait
5 years
projects completed 2024
Projects in Queue
10,300
Completion Rate
14%
of capacity, 2000–2023
📊 Queue trend
Queue declined from ~2,600 GW (end-2023) to ~2,300 GW (end-2024) as FERC Order 2023 reforms took effect. Generation: 1,400 GW. Storage: 890 GW.
Source: DOE/LBNL — Queued Up: 2025 Edition (data through end of 2024) · as of Dec 31, 2024 · Queue backlog includes all generation types (solar, wind, gas, nuclear, storage), not only data center loads. Completion rate is capacity-weighted (14% of capacity, 2000–2023). Queue declined from ~2,600 GW (end-2023) to ~2,300 GW (end-2024) as FERC Order 2023 reforms took effect.
Price & Risk Signal — Wholesale Electricity
Wholesale electricity prices in data-center-heavy regions (PJM, ERCOT) are an early indicator of grid stress. Rising prices signal that demand is outpacing supply in the regions where AI infrastructure is concentrated.
PJM Western Hub DA LMP
$42.5/MWh
30-day rolling avg
ERCOT Houston Hub DA SPP
$38.2/MWh
30-day rolling avg
YoY Change
+18.5%
PJM Western Hub
DC Premium (est.)
12%
qualitative only
⚠️ DC premium: qualitative only, not used in AISI scoring
The data center premium (12%) is estimated from public PPA disclosures and industry reports. It is shown for context but is not independently verifiable at scale and does not feed into the AISI composite score. Only the PJM/ERCOT wholesale price YoY change drives the Price & Risk component.
Source: PJM Data Miner (Western Hub, DA LMP) + ERCOT MIS (Houston Hub, DA SPP) · as of Feb 24, 2026 · Series: PJM Western Hub Day-Ahead LMP, ERCOT Houston Hub Day-Ahead SPP. Window: 30-day rolling average.
Wall Street AI Infrastructure Outlook 2026
⚠️ Reference only — not used in AISI scoring
The forecasts below are from institutional research reports, many of which are paywalled or proprietary. They are shown for context and comparison only. None of these numbers feed into the AISI composite score. Only publicly accessible, clickable sources (DOE/LBNL, IEA, NERC, EIA) are used as index inputs.
Institutional forecasts for AI capex, power demand, and grid bottleneck risk. The spread between bullish capex projections and cautious grid assessments is itself a stress signal — when capital commitments outpace infrastructure delivery, the system is more exposed to timeline corrections.
Reference range: AI power demand 35–90 GW by 2028 (avg 64 GW). Not used in AISI scoring.
| Institution | AI Capex 2026 ($B) | Power 2028 (GW) | Grid Risk |
|---|---|---|---|
| Goldman Sachs | $250B | 72 | moderate |
| Morgan Stanley | $220B | 68 | high |
| IEA | N/A | 35 | low |
| NERC | N/A | 55 | high |
| McKinsey | $200B | 65 | moderate |
| Bernstein | $280B | 90 | severe |
Data Freshness — Asynchronous Timeline
| Data Source | Frequency | Last Updated | Status |
|---|---|---|---|
| EIA Annual Energy Outlook | Annual | Jan 15, 2025 | LIVE |
| Institutional Forecasts | Quarterly | Jan 15, 2025 | LIVE |
| DOE/LBNL Queue Data | Quarterly | Dec 31, 2024 | LIVE |
| PJM / ERCOT Prices | Monthly | Feb 24, 2026 | LIVE |
Frequently Asked Questions
What is the AI Infrastructure Stress Index (AISI)? ▾
The AI Infrastructure Stress Index (AISI) is a composite measure of structural tension between AI compute demand growth and energy infrastructure delivery capacity. It synthesizes 4 signals — demand-supply gap, consensus dispersion, infrastructure lag, and price risk — into a single 0–100 score. Current score: 46/100 (moderate).
How much power does AI consume? ▾
Public forecasts for US data center power demand by 2028 range from 35 GWavg (IEA conservative) to 66 GWavg (DOE/LBNL high scenario), with a mean of 48.25 GWavg across 4 public forecasters. For context, the entire US grid capacity is approximately 1,200 GW. The wide spread (31 GWavg) reflects deep uncertainty about AI adoption pace, model efficiency, and inference scaling. Sources: DOE/LBNL, IEA, NERC — all publicly accessible.
What is the interconnection queue backlog? ▾
As of Dec 31, 2024, approximately 2,300 GW of generation and storage capacity is waiting in the US interconnection queue, with an average wait time of 5 years for projects completed in 2024. This backlog includes all generation types (solar, wind, gas, nuclear, storage), not only data center loads. Historical completion rates are approximately 14% of capacity (2000–2023). Source: DOE/LBNL Queued Up 2025 Edition.
Why do AI power demand forecasts vary so widely? ▾
The 31 GWavg spread between the lowest (35 GWavg, IEA) and highest (66 GWavg, DOE/LBNL high scenario) public forecasts reflects fundamental uncertainty about: (1) AI model efficiency gains vs scaling laws, (2) inference-to-training compute ratio evolution, (3) sovereign AI buildout pace outside the US, and (4) whether hyperscaler capex commitments translate to actual power draw. AISI tracks this dispersion as a stress signal — wider disagreement indicates higher structural uncertainty.
Is the AISI investment advice? ▾
No. The AI Infrastructure Stress Index (AISI) is a structural stress audit tool developed by AhaSignals for educational and analytical purposes only. It does not constitute investment advice, financial advice, or trading recommendations. Users are responsible for their own decisions.
Methodology — AISI v0.1-beta
The AI Infrastructure Stress Index (AISI) is a composite measure of structural tension in the AI-energy nexus. It does not predict AI power demand or energy prices — it measures the stress building between demand narratives and infrastructure delivery reality.
Component Formulas:
1. Demand-Supply Gap (40%): score = clamp(gapPct × 2.5, 0, 100). Source: DOE/LBNL + EIA AEO.
→ Current: gapPct=13.5% → score=34
2. Consensus Dispersion (30%): score = clamp(CV% × 1.5, 0, 100). Source: IEA, DOE/LBNL, NERC (public only).
→ Current: CV=26.7% → score=40
3. Infrastructure Lag (20%): score = clamp(avgQueueYears × 15, 0, 100). Source: DOE/LBNL Queued Up.
→ Current: avgYears=5 → score=75
4. Price & Risk Signal (10%): score = clamp(yoyChangePct × 3, 0, 100). DC premium is qualitative only, not scored. Source: PJM + ERCOT.
→ Current: yoy=+18.5% → score=56
Composite:
AISI = Σ(weight_i × score_i) / Σ(weight_i)
→ Current: (0.4×34 + 0.3×40 + 0.2×75 + 0.1×56) = 46
Signal: low <31 | moderate 31–55 | elevated 56–75 | high ≥76
Index Input Sources (all free, public, clickable):
- DOE/LBNL — Data Centers and AI (demand range, queue data)
- EIA Annual Energy Outlook (committed capacity additions)
- NERC Long-Term Reliability Assessment (load growth projections)
- IEA World Energy Outlook (conservative demand estimate)
- DOE/LBNL — Queued Up (interconnection queue)
- PJM Data Miner (Western Hub DA LMP)
- ERCOT Market Information (Houston Hub DA SPP)
Wall Street research (Goldman, Morgan Stanley, McKinsey, Bernstein) appears in the reference table above but does NOT feed into any scored component.
Known Limitations:
- AI power demand projections are inherently uncertain; efficiency gains could reduce actual demand significantly.
- Queue backlog includes all generation types, not only data center loads; historical completion rate is ~14% of capacity (2000–2023).
- Wholesale prices are hub-specific averages (PJM Western Hub, ERCOT Houston Hub); actual data center costs depend on bilateral PPAs.
- Data center premium is qualitative only — not used in AISI scoring. Actual premiums vary by region and contract.
- Infrastructure Lag uses queue wait time only (avgYears × 15). The previous cross-unit ratio (queueBacklog / announcedDCCapacity) was removed because announced DC capacity had no verifiable public source.
- This is an experimental research tool (v0.1-beta). Not investment advice.
📎 Cite This Data
APA:
AhaSignals. (2026). AI Infrastructure Stress Index (AISI). Retrieved from https://ahasignals.com/ai-infrastructure-stress-index/
MLA:
"AI Infrastructure Stress Index (AISI)." AhaSignals, 2026, ahasignals.com/ai-infrastructure-stress-index/.
Chicago:
AhaSignals. "AI Infrastructure Stress Index (AISI)." 2026. https://ahasignals.com/ai-infrastructure-stress-index/.
Deep Dive — AI Energy Research
Detailed analysis pages exploring specific dimensions of the AI-energy nexus. Each page uses only publicly verifiable data sources.
Power by Company
Hyperscaler electricity footprint audit: Microsoft, Google, Amazon, Meta.
Grid vs AI Demand
Regional grid capacity analysis against projected AI load growth.
PJM Electricity Prices
Wholesale price trends in the largest US data center corridor.
AI Capex Tracker
Hyperscaler capital expenditure vs deliverable power capacity.
NERC Reliability
Large load integration risk audit from NERC assessments.
Queue Tracker
Interconnection queue backlog (~2,300 GW), wait times & completion reality.
Related Trackers
📊 Get Updates
Get weekly updates when the dashboard state changes materially, plus new research on consensus fragility and market divergence. Research-only. Not trade signals.
🔒 No spam. Unsubscribe anytime. 2,000+ researchers and practitioners as of Apr 2026.