AI Stress Index Methodology

The AI Infrastructure Stress Index is a single number between 0 and 100. It answers one question: is AI a bubble, or can the current pace of AI infrastructure spending pay for itself?

Twelve public data signals feed into a weighted composite, updated daily. Weights are derived using the same multi-criteria framework the OECD, UN, and World Bank use for their composite indexes. Higher means more stress. Lower means the buildout looks financially sustainable.

We don't predict crashes or recommend trades. The score tracks how far apart AI capex spending and revenue are, and whether that gap is growing.

Organized into three tiers by information type.

The most common criticism of composite indexes: "Why 15% and not 12%?" Without a framework, that question has no answer.

We score each signal 1–5 on five dimensions, then derive weights proportionally from the totals:

Where $s_i$ is the total score for signal $i$ across all five dimensions. The denominator is the sum of all twelve signal scores (237 in our case). This is what the OECD Handbook on Constructing Composite Indicators calls the "Budget Allocation Process." The UN Human Development Index and World Bank Doing Business Index use the same method.

The five dimensions:

• ·Lead time measures how far in advance of a crisis the signal moves. Leading indicators get more weight.
• ·Update frequency. Daily data is worth more to a daily-updating index than quarterly data.
• ·Data reliability is about the source. Audited SEC filings and Federal Reserve data rank higher than web-scraped estimates.
• ·Historical precedent. How well has this type of signal called past bubbles? Track records matter.
• ·Uniqueness asks whether the signal adds information the others don't. Correlated signals shouldn't double-count the same risk.

The resulting range is 6%–11%, compared to a pure equal weight of 8.3% per signal. Credit spreads score highest on every dimension. Enterprise AI sentiment scores lowest because the methodology is newer and partially correlated with other fundamentals.

That convergence toward near-equal weighting is expected. An IMF paper (Brave & Butters, 2022) found that equal-weight financial conditions indexes perform comparably to more complex statistical approaches. The structure gives justified deviation from equal weight without arbitrary extremes.

Weights are reviewed annually. When we have 12+ months of continuous data, we'll publish a PCA validation study comparing statistically-derived weights against these multi-criteria weights.

Each signal is normalized to a 0–100 scale. 0 means no stress. 100 means maximum stress. At launch, normalization uses min-max scaling against theoretical floor and ceiling values derived from historical extremes:

Clamped to [0, 100]. All signals are sign-aligned so that higher always means more stress.

Once we have 12+ months of data, we'll move to percentile ranking against observed history, which is the method SentimenTrader and most professional composite models use. Percentile ranking is more resistant to outliers and self-calibrates as more history accumulates.

The composite score is a weighted average of all twelve normalized signals:

Where $w_i$ is the weight from the multi-criteria framework and $\hat{x}_i$ is the normalized score. All weights sum to 1.0. Linear aggregation means a low-stress reading on one signal can offset a high-stress reading on another. One signal spiking while everything else is calm matters less than several rising together.

If a signal can't be updated (data source outage, API failure), the most recent valid value is carried forward for up to 30 days. After that, the signal is excluded and remaining weights rescale to sum to 100%. Any exclusion is flagged on the site.

The composite updates daily at market close. Eight signals update daily or better (credit spreads, GPU pricing, data center REITs, insider trading, options sentiment, AI adopter performance, data center infrastructure, AI power demand). One updates monthly (semiconductor supply chain). Three update quarterly (NVIDIA growth, hyperscaler capex, enterprise sentiment). When a quarterly signal updates, say after NVIDIA earnings, the composite may shift more in a single day. That's expected. The site shows the "as of" date for each signal so you can tell which data is fresh and which is carried forward.

The 25-point zones are chosen for simplicity. They aren't statistically derived. With limited history, there's no empirical basis for optimizing thresholds yet. As data accumulates, zone boundaries may be adjusted, with changes announced in advance.

The index draws on three frameworks from financial economics. Michael Burry's AI prediction draws on similar foundations, though our index tracks the data directly rather than making timing calls.

Minsky's Financial Instability Hypothesis. Long periods of stability encourage risk-taking until the system becomes fragile. Borrowers evolve through three stages: hedge finance (cash flow covers all obligations), speculative finance (can cover interest but must refinance principal), and Ponzi finance (can't cover interest or principal). Our credit spread and capex/cash flow signals are built to catch this transition in AI infrastructure.

Kindleberger's Five Bubble Stages. Displacement (ChatGPT launch), boom (hyperscaler capex ramp), euphoria (AI = 92% of GDP growth), profit-taking/distress, panic. The index targets the transition from stage three to stage four. That's the highest-value moment for anyone watching.

Credit-to-GDP cycle analysis. Fed, IMF, and BIS research all point to the same finding: the rate of change of credit, not its absolute level, predicts financial crises. AI-related debt issuance exceeds $200B in 2025. JPMorgan projects data center paper will reach 7-10% of structured credit issuance by 2026-2027. AI infrastructure accounts for 4% of GDP but drove 92% of GDP growth in H1 2025. The question of AI spending sustainability comes down to whether this credit growth can be serviced by actual revenue.

All data comes from public sources. Nothing is proprietary or paywalled at the source level.

• ·Credit spreads: FRED (Federal Reserve Economic Data), FINRA TRACE
• ·GPU spot pricing: AWS, Azure, GCP, Vast.ai, Lambda Labs public APIs
• ·NVIDIA financials: quarterly earnings via SEC EDGAR
• ·Hyperscaler financials: quarterly earnings from Microsoft, Alphabet, Amazon, Meta (SEC EDGAR)
• ·Earnings call transcripts: SEC EDGAR filings
• ·Semiconductor data: SIA (Semiconductor Industry Association), TSMC investor relations, SEMI
• ·Data center REITs: market data for EQIX, DLR, VNQ; quarterly SEC filings
• ·Insider trading: SEC Form 4 filings via EDGAR
• ·Options data: CBOE, OCC (Options Clearing Corporation)
• ·AI adopter performance: market data for SNOW, CRM, NOW, PLTR, SPY via Polygon.io
• ·Data center infrastructure: market data for VRT, ETN, XLI via Polygon.io
• ·AI power demand: market data for CEG, VST, NRG, XLU via Polygon.io

The AI Infrastructure Stress Index is focused on one thing: the financial sustainability of AI infrastructure spending. That narrow focus is deliberate. It means the index does not cover:

• ·Whether AI technology is valuable (it is)
• ·Individual stock prices or short-term market moves
• ·Geopolitical risk, regulation, or policy changes
• ·Spending by governments or defense contractors
• ·The pace of AI research breakthroughs

A high score doesn't mean AI is failing. It means the money being spent on AI infrastructure may not generate enough return to justify itself.

Some signals have decades of history (credit spreads, semiconductor data). Others have 1–3 years (GPU spot pricing, AI-specific credit). Normalization and weight calibration will get better as history accumulates.

Three of twelve signals update quarterly. Between updates, those signals carry forward stale values, which means the composite can underreact to fast-moving conditions in those dimensions.

The multi-criteria framework is more structured than pure expert judgment, but the criteria scores still involve judgment. We tested this: perturbing weights by ±3 percentage points across 10,000 simulations produces deviations of less than ±2.5 points on the composite. The score is not fragile.

In a real downturn, signals tend to move together. The individual signal breakdown, available to Pro subscribers, helps isolate which risk dimensions are driving the composite.

And one more thing: the score measures current stress, not timing. The 2008 housing market was unsustainable for years before the crash. A high score means stress is elevated. It doesn't tell you when.