Methodology

Last updated: 2026-10-17

The question, made precise

"Is AI a bubble?" is decomposed into questions that can be wrong:

1. Overbuild: is capital being deployed faster than demand that can service it will appear? Measured per layer of the stack.
2. Financing: who funds the build, with what instrument (cash flow, equity, bonds, private credit, vendor financing, off-balance-sheet SPVs), and who bears the loss if returns disappoint?
3. Pricing: are securities (equity, debt) priced for outcomes that the evidence makes unlikely?
4. Resolution: concrete, dated events whose probability we estimate and later score.

Technology can be real and transformative while the capital deployed to build it does not all earn a return (railways in the 1840s, fiber in 1999). Both can be true at once; the report keeps them apart.

Lenses

• Mechanism: Minsky and Kindleberger stages (displacement, boom, euphoria, profit-taking, panic); Dalio's bubble indicators (prices high relative to fundamentals, prices discounting rapid future appreciation, broad bullish sentiment, leverage-financed purchases, forward purchases, new entrants, tightening policy); Perez's installation-vs-deployment cycle. Each becomes a scored indicator with a stated current reading.
• History: the closest precedents, what signalled the top, and how long the overbuild took to absorb.
• Plumbing: where the debt sits, depreciation assumptions, unit economics of inference and training, and whether end-user revenue grows fast enough to service the capex.

Evidence rules

• Every number carries a source, an "as of" date and a type. Nothing is stated from model memory.
• Source types: primary (SEC filings, earnings releases and transcripts, company statements, central-bank and statistical-agency data, court filings), secondary (reporting by named outlets), tertiary (aggregators, social media, analyst commentary). Secondary and tertiary claims about private companies are labelled "reported, unverified".
• Confidence grades for a claim:
  • A: primary source, directly stated.
  • B: reputable secondary source, consistent with primary data, or derived from primary data by simple arithmetic.
  • C: single secondary source, leak, or estimate; plausible but unverified.
  • D: tertiary, contested, or contradicted by other sources.
• Conflicting evidence is recorded, not resolved by omission.
• No paywalled sources were available at the time of writing. Coverage of public companies is therefore strong; coverage of private labs and neoclouds relies on reported figures and is graded accordingly.

Forecast rules

• Each forecast is a question that resolves yes or no by a stated date, from a stated public source.
• Horizons: 6, 12 and 24 months from the forecast date. Error bars at 24 months are large and are stated as such.
• Each forecast records the probability, the date, the reasoning, and every later revision with its cause. Forecasts are scored (Brier) once resolved.
• Per-layer summaries estimate the probability that capital deployed in that layer is materially impaired (equity drawdown or debt loss) within the horizon.

Agents

The research is produced by AI agents under a lead agent (Claude), steered by a human. Claude subagents write research notes; Codex (OpenAI) independently re-researches contentious notes and reviews conclusions; Grok (xAI) scans x.com for new information, which is then traced to primary sources before use. Nothing from a social-media scan is cited as evidence on its own. The lead agent is an Anthropic model and assesses Anthropic by the same rules as every other lab; the Codex cross-check is the control for that blind spot.

The agent sessions that built the report are not published.