Evidence before alerts.

What is live now

SEC EDGAR companyfacts supplies annual cash-flow and capital-expenditure facts for Microsoft, Alphabet, Amazon and Meta. Free cash flow is operating cash flow minus purchases of productive assets or property and equipment. Company values are accepted only when cash flow and capex share the same fiscal periods.

FRED supplies the ICE BofA US Corporate option-adjusted spread. Its 90-day change is scored as a broad funding-stress proxy. It is not an AI-issuer-specific spread or debt-issuance measure, so its confidence is low.

NVIDIA’s latest SEC-filed earnings release supplies quarterly Data Center revenue and year-over-year growth. Strong growth scores as lower cycle risk. This is a demand proxy, not a complete measure of backlog, cancellations, lead times or competitors.

OpenRouter supplies a current catalogue containing Artificial Analysis Intelligence Index scores, token prices and published Hugging Face identifiers. The highest-scoring proprietary model is the reference. Among open-weight models scoring at least 90% of it, the lowest combined list price for one million input plus one million output tokens is selected.

The monetisation proxy compares aggregate AWS and Google Cloud annual revenue growth with aggregate Amazon and Alphabet capex growth over exactly matched fiscal periods. It is deliberately labelled low confidence because cloud revenue includes non-AI workloads and Microsoft and Meta do not disclose directly comparable AI/cloud revenue.

Census private data-centre construction divided by BEA nominal GDP produces the macro-exposure reading. Construction excludes servers, chips, power generation and other equipment, so it understates total AI infrastructure investment.

State Street’s official SPY fund page supplies the top-10 holdings weight and the S&P 500 forward one-year P/E. These are the concentration and valuation components of market confirmation. Earnings-revision breadth remains unavailable and is not guessed.

Scoring guardrails

• Every scored signal includes a source, observation period, update time and documented transformation.
• Missing data is never converted to a neutral score.
• The composite is unavailable until at least six signals are verified and six are scored.
• Monetisation scores 1 when cloud revenue growth matches or exceeds capex growth; 2 when capex leads by no more than 15 percentage points; 3 at no more than 30 points; 4 at no more than 50 points; otherwise 5.
• Open-model pressure scores 5 when the benchmark gap is no more than 5% and the price discount is at least 50%; 4 at a gap of no more than 10% with the same discount; 3 at a gap of no more than 15% and discount of at least 30%; 2 at a gap of no more than 20%; otherwise 1.
• Market confirmation is scored only when at least two of three components are verified. Top-10 weight scores from 1 at 25% or less to 5 above 40%; forward P/E scores from 1 at 16× or less to 5 above 25×. The two verified component scores are averaged and rounded. Earnings breadth contributes nothing until connected.
• Amber requires a verified composite of 2.6–3.9 or three independent deteriorating signals.
• Red requires a composite of at least 4.0, five deteriorating signals, or a separately governed exceptional funding event.
• Daily, weekly and monthly controls select the last available observation in each calendar period. They never interpolate or manufacture missing readings.

Historical reconstruction

The chart includes twelve point-in-time month-end reconstructions from August 2025 through July 2026. Each point uses only evidence that had been published by that date. These readings provide historical context; they are not evidence that Boom Risk operated or issued real-time alerts during that period.

The reconstructed composite uses six defensibly recoverable scored markers: hyperscaler free cash flow, the FRED corporate-credit spread proxy, NVIDIA Data Center revenue growth, the AWS and Google Cloud monetisation proxy, market concentration and valuation, and US data-centre construction relative to GDP. Capex intensity remains context only to avoid double-counting monetisation. Open-model pressure is excluded because a reliable versioned archive combining benchmark results, model classification and contemporaneous prices was not available.

Update timing follows the underlying evidence rather than inventing a uniform frequency. Credit is sampled at month-end from daily observations; data-centre construction is monthly; semiconductor demand changes when quarterly results are published; free cash flow and monetisation change when comparable company filings become available; and market confirmation uses the contemporaneous concentration and valuation evidence available for the period.

The dashed line and hollow markers denote reconstructed six-signal readings. The solid line denotes monitored observations beginning 22 August 2026 and uses the coverage verified at each run. Tooltips disclose the score, date, origin and scored-signal count. Because historical and live coverage differ, small changes at the handover should not be treated as pure market movement.

Reconstruction sources include SEC EDGAR, FRED corporate spreads, NVIDIA results, US Census construction data, and contemporaneous S&P Dow Jones Indices and FactSet market publications.

Known limitations

Company fiscal years differ, XBRL tags can change, and annual filings are lagging indicators. Cloud revenue is not AI-only revenue. OpenRouter prices are hosted list prices rather than self-hosted economics, and Hugging Face identifiers are a practical open-weight classification rather than a licence audit. The market-confirmation score currently lacks earnings-revision breadth and its two components come from one official fund source. Issuer-level credit spreads, consensus earnings revisions and semiconductor lead times require further normalization or licensed data. The monitor is research tooling, not investment advice.

Planned feeds

The remaining market-confirmation priority is licensed earnings-revision breadth. Capex intensity remains verified context rather than a separate score because its capex-versus-revenue risk is already captured by monetisation; scoring both would double-count the same economic pressure. AI-specific credit and consensus datasets may require commercial licences.