Index Methodology

The composite AICDI measures overall AI compute supply chain strain on a 0 to 100 scale. Higher values indicate greater strain across the infrastructure stack.

AICDI = Hardware(25%) + Foundry(20%) + Memory(20%) + Energy(20%) + Cost(15%)

Component Weights

Severity Levels

Each sub-index is normalized to 0 to 100 before weighting. The hardware demand index (unbounded 0 to infinity) is mapped through a compression function: min(100, raw * 0.2). Cost pressure is derived from average TCO: min(100, avgTCO / 300).

Measures demand pressure on AI accelerators. The index is intentionally unbounded (0 to infinity) because supply conditions can worsen without a theoretical ceiling. A value of 100 represents balanced supply; 200 means lead times have doubled; 400+ signals full parabolic demand with no relief in sight.

AICDI = weightedTightness * leadTimeMultiplier * capexAccelerator + bubblePremium

Components

• ◆Weighted Tightness · Tier-weighted average of per-chip supply tightness scores. Frontier chips count 3x, mainstream 2x, others 1x.
• ◆Lead Time Multiplier · Average lead time in weeks divided by a 12-week baseline. Currently frontier chips average 32+ weeks.
• ◆Capex Accelerator · Hand-curated scalar from hyperscaler earnings calls reflecting aggregate capex acceleration. Currently set at 1.85x.
• ◆Bubble Premium · Each frontier chip sold out past 24 weeks adds 15 points. Captures extreme supply deficit conditions.

Interpretation Scale

Measures supply-side concentration risk in semiconductor manufacturing. When 90% of AI chips are fabricated at one company in one country, the entire AI supply chain carries systemic risk. The FCI surfaces this.

Scale: 0 (diversified) to 100 (monopoly risk).

FCI = companyHHI(35%) + geoConcentration(30%) + packagingStrain(20%) + (100 - altReadiness)(15%)

Components

• ◆Company HHI · Herfindahl-Hirschman Index across foundries, normalized to 0 to 100. Measures market share concentration.
• ◆Geographic Concentration · Percentage of all tracked AI chips fabricated in the top single country. Currently Taiwan dominates.
• ◆Packaging Strain · Maximum packaging utilization across all foundries. CoWoS-L capacity is often the binding constraint.
• ◆Alternative Readiness · How viable alternative fabs are (0 to 100, higher = more alternatives). Inverted in the formula so low readiness increases the FCI.

Measures demand pressure on the global memory supply, particularly High Bandwidth Memory (HBM). Scale: 0 (balanced) to 100 (critical shortage).

MDI = aiAbsorption(30%) + supplyGap(25%) + supplierConcentration(20%) + pricingPressure(15%) + crossIndustryStrain(10%)

Components

• ◆AI Absorption · What percentage of total HBM output is consumed by AI workloads. Higher absorption means less slack.
• ◆Supply-Demand Gap · Demand exceeds supply by this percentage. Normalized: 0% gap = 0, 50%+ gap = 100.
• ◆Supplier Concentration · HHI across the three HBM suppliers (SK hynix, Samsung, Micron), normalized to 0 to 100.
• ◆Pricing Pressure · Weighted average year-over-year price change across HBM generations. Normalized: -30% = 0, +60% = 100.
• ◆Cross-Industry Strain · Non-AI demand growth competing for DRAM fabrication capacity (automotive, mobile, networking). Normalized: 0% growth = 0, 40%+ = 100.

Normalizes all AI compute systems to a comparable cost metric: $/PFLOPS/year. This is the metric datacenter operators actually optimize for when choosing between NVIDIA, AMD, Google TPU, or custom silicon. Lower is better.

TCO = (listPrice / 3 years + powerKW * 8760h * $0.05/kWh * PUE) / FP8_PFLOPS

Assumptions

• ◆Hardware Amortization · 3-year straight-line depreciation of list price.
• ◆Power Cost · $0.05/kWh, the average datacenter electricity cost. Actual rates vary by region ($0.02 to $0.12/kWh).
• ◆PUE Factor · Power Usage Effectiveness, per system. Liquid-cooled systems typically achieve 1.1; air-cooled systems 1.3 to 1.5.
• ◆Compute Output · FP8 PFLOPS as the denominator. FP8 is the dominant precision for AI training and inference on current-generation hardware.

When list price is unavailable but a monthly lease price exists, the annual cost uses leaseMonthly * 12 + annualPower instead.

Measures energy pressure on AI infrastructure. Scale: 0 (comfortable) to 100 (infrastructure cannot keep pace with demand).

EPI = powerDemandGrowth(25%) + gridStrain(20%) + ppaPriceEscalation(20%) + nuclearPremium(15%) + waterStress(10%) + permitDelays(10%)

Components

• ◆Power Demand Growth · AI datacenter power year-over-year growth rate, scaled: YoY% * 2, capped at 100.
• ◆Grid Strain · Average datacenter share of regional grid capacity across tracked regions, scaled by 5x.
• ◆PPA Price Escalation · Year-over-year change in average Power Purchase Agreement prices, scaled by 3x.
• ◆Nuclear Premium · Gap between average grid price and PPA price, reflecting the premium hyperscalers pay for dedicated clean power.
• ◆Water Stress · Share of tracked datacenter regions with high or extreme water stress, scaled by 1.5x.
• ◆Permit Delays · Average permit wait time across regions, normalized against a 48-month ceiling.

Each of 10 global datacenter regions receives a 0 to 100 readiness score. Higher scores mean the region is better prepared for large-scale AI infrastructure deployment.

Readiness = computeDensity(25%) + energySecurity(25%) + constructionVelocity(20%) + waterSustainability(15%) + regulatoryClimate(15%)

Components

• ◆Compute Density · Current datacenter MW capacity (normalized against the largest region) plus growth rate bonus. Captures existing scale.
• ◆Energy Security · Carbon-free power percentage, PPA affordability relative to peer regions, and grid headroom. A region with cheap, clean, abundant power scores higher.
• ◆Construction Velocity · MW under construction (normalized) and inverse of permit wait time. Captures how fast new capacity is being added.
• ◆Water Sustainability · Inverse of water stress level, bonus for air or liquid cooling, and free-cooling month count. Captures long-term operational viability.
• ◆Regulatory Climate · Derived from region status (growing, emerging, constrained, restricted, saturating) and operator diversity. More operators and a favorable status score higher.

Applied to 10 global datacenter regions including Northern Virginia, Iowa, the Nordics, Singapore, Tokyo, and more. Rankings are available on the Compute Hub.