The rule behind everything on this site: every published number traces to a public primary source anyone can check. Where judgment is unavoidable, the judgment is written down in a versioned assumptions file, and changing it forces the entire history to be restated — so movement on the dashboard is never an artifact of changed methods. Specialist research (credited) calibrates assumptions; it never appears as data.
"Is the AI boom a bubble?" is unanswerable, so this dashboard decomposes it into three measurable questions: is the machinery paying for itself (the earnings dial); is the world getting value even where investors aren't (the benefit dial); and whose money burns if it goes wrong (the debt dial). The intellectual framework — including the adversarial reviews it survived — is documented in the project's research notes; the short version is in chapter 9 of The Price of Thinking.
One framing caution, stated plainly: each dial is a coverage level, not a probability. The earnings dial does not output "the chance of a correction" — it measures how much of the capital cost the earnings currently cover, which is an input to that chance, not the chance itself. Turning a level into a probability needs an expectations benchmark this project does not claim to have. So the honest reading is "how well covered, and which way is it moving," never "how likely a crash."
| Source | What we take | Why this source | Tier |
|---|---|---|---|
| SEC EDGAR (XBRL company facts) | Capital spending, operating cash flow, debt issuance, per firm, quarterly (derived from year-to-date figures) | Audited filings; the only undisputed record of what is actually spent and earned. Free, no key, machine-readable. | T1 |
| US Census, Construction Spending (C30) | Monthly data-centre construction put in place (private, SA) | The official measure of the durable, building-shaped part of the boom; published monthly since 2014 for this category. | T1 |
| US Treasury daily yield curve | 10-year yield | The primary source for the risk-free rate (chosen over FRED's mirror after FRED's keyless endpoint proved unreliable in production). | T1 |
| FRED | Baa corporate spread (and 10y cross-check) | The standard public series for credit spreads; retained with a retry-and-canary design. | T1 |
| vast.ai public marketplace | Weekly median rental price per GPU model | The only open, observable market price of AI compute — a direct market check on what the machinery is worth per hour. This archive cannot be backfilled, so we never delete it. | T1 |
| Damodaran (NYU) implied ERP | Equity risk premium, annual history | The de-facto public standard for the equity premium; used in the required-return construction. | T2 |
| FMP (Financial Modeling Prep) | Earnings-call transcripts; statements for foreign filers EDGAR misses (e.g. Nebius) | Licensed aggregator used as a conduit to filed information; full transcripts are never republished — only short, quoted sentences with provenance. | T2 |
| Company newsrooms, press-wire services, SEC filings | The deal ledger: every financing deal, one row, one primary link | Deals reported only in the press do not enter the ledger until re-sourced to a primary document — our validator enforces this mechanically. | T1/T2 |
| Census BTOS · BLS (wages, productivity) · EIA (electricity prices) · Anthropic Economic Index | Adoption, wages, energy prices, AI task-usage weights | The benefit dial's ingredients: official surveys where they exist, plus the only public task-level AI usage dataset — now active in the benefit construction (v5). | T1/T2 |
Tiers: T1 = filed or official statistics. T2 = company-stated or reputable aggregated data. T3 = press-reported (always flagged, never load-bearing alone).
capex (trailing 4 quarters) ÷ operating cash flow (trailing 4 quarters), per firm. Above 1.0, building consumes more cash than the whole company generates. Quarterly values are derived by differencing year-to-date cash-flow figures within each fiscal year. Caveat stated wherever shown: hyperscaler cash flow comes mostly from non-AI businesses, so this measures cross-subsidy capacity, not AI self-financing.
Each externally financed deal is one row: parties, instrument, amount, location bucket (F3 = credit-like external finance to compute owners; F4 = vendor or circular capital), primary source URL, and whether it was found contemporaneously or retrospectively — retrospective discovery has survivorship bias, and labelling it lets us measure that bias instead of hiding it.
v3 (2026-07-06): the ledger is now folded into F, not merely counted. The off-balance-sheet SPV, ABS, lease and vendor financing engineered to stay out of the filed capex/OCF ratio enters the weakest-link maximum as a leverage intensity (scaled by AI capex), and the reciprocal-capital (circularity) bonus specified in the assumptions is now applied. Each ledger deal is tagged reported (funded at close, amount disclosed) or estimated (revolver or build-out draw), so the share of the leverage resting on estimates is itself visible. Both are as-of aware, so the historical trajectory shows leverage accumulating rather than today’s ledger stamped on the past. The leverage term is now on a drawn basis (a per-deal drawn fraction in the ledger, defaulting by instrument type) rather than announced commitments \u2014 e.g. funded notes and ABS count in full, revolvers and the build-to-2028 SPV only partly; still coarse (drawn estimated, discovery bias), reported transparently, and it binds the headline only when it exceeds the filed edge; equity finance is excluded from the leverage term.
Σ over asset classes of (rᵢ + δᵢ) × Kᵢ — what the stock must earn each year to justify itself: required return plus wear. K accumulates from filed capex, split across chips/servers, buildings, and power by the published allocation assumptions. v2 attribution: each firm's capex is multiplied by an AI-share band (hyperscaler 45–85%, neocloud ~100%) before accumulation, so K is AI-attributed rather than total capex — and, crucially, the earnings ratio compares a firm's AI revenue with the AI capital cost of the same firms, never a three-firm numerator against a nine-firm denominator. (This corrected a downward bias flagged in adversarial review: the earnings coverage moved from ~0.10 under the v1 mismatch to ~0.50 once numerator and denominator share a population.) v5: the hyperscaler AI-share was raised from a more conservative 25–60% band to 45–85%, matching mid-2026 evidence that data-centre/AI is now ~70–90% of hyperscaler capital spending (Dell'Oro data-centre-capex tracker, company guidance; Amazon's logistics base pulls the blend down) — a larger, better-grounded AI capital base that lowers both coverage ratios (earnings mid ≈0.45→0.29). The intended SemiAnalysis vintage calibration remains pending, so this stays a documented band. δ (wear) is carried as a band: chips at 18–40% a year, deliberately spanning the gap between company accounting schedules and the sceptics' arithmetic. 2026-07-06: the chip mid was recentred 0.28→0.20 (~5-year life) toward the longer-life evidence that accumulated through mid-2026 — Exponential View’s State of the AI Economy (25 Jun 2026) defends a six-year life and shows GPU rental yields persisting past six years; Meta extended server life to 5.5 years; Nvidia argues 4–6 — while the 0.40 high is retained for Burry’s 2–3-year replacement thesis. The whole back-series was restated on the same run so the movement stays real (earnings mid ≈0.29→0.31). r combines the Treasury 10-year with Damodaran's equity premium (equity-funded share) and the Baa spread (debt-funded share), weighted by the observed financing mix. v5: this split is now active per firm — the debt-funded weight is trailing-4-quarter debt issued ÷ capex from filed cash-flow statements, so a firm that leans on debt carries a lower blended required return (its extra risk from that leverage is captured separately by the fragility dial, not double-counted here). At mid-2026 the equity return is ~8.75% and the debt return ~6.15%, blended by each firm's mix.
Revenues attributable to the capital, net of operating costs only — energy, labour, inference, sales. Depreciation is deliberately not netted here: it is already charged in the user-cost denominator, and netting it twice would double-count it (the margin band is therefore an EBITDA-type margin). A revenue figure is converted through that band (30–70%, the widest in the system, because inference margins are genuinely disputed). Revenue inputs are tier-labelled: filed segment data first, company statements from earnings calls second (each kept as a verbatim sentence with its transcript reference), press reports last and flagged. Two lines are computed: the compute owners' coverage and the wider ecosystem's. v5: the two mechanical revenue inputs now self-refresh from filings each harvest — neocloud revenue as the trailing four quarters of reported income-statement revenue (CoreWeave, IREN, Applied Digital) and the hyperscaler cloud-segment line from the revenue-segmentation filing (Google Cloud, Oracle Cloud & License) — rather than being hand-entered. The verbatim earnings-call AI-ARR figures (Microsoft, Amazon, Nebius) stay human-reviewed; if an auto series is ever missing, the dial falls back to the last curated value. v3: a second coverage line is now reported alongside the K-holder Cp \u2014 whole-stack ecosystem quasi-rents over the same capital cost, revenue from Exponential View’s deduplicated total (banded $110bn trailing to $175bn run-rate; tier T2, an external estimate flagged distinct from the filed line and kept off the trajectory). The wedge between the ecosystem and K-holder lines is the value captured above the capital-holders \u2014 the read-out that tells competition-eroding-appropriability apart from a bubble.
Task-level value coefficients from peer-reviewed experiments × measured usage by occupation × wages × adoption, expressed against the same user cost, at 10- and 30-year horizons. v5: the per-occupation employment weights are the exact BLS major-group counts (the same May-2023 OEWS table the Anthropic Economic Index uses), and the AI-usage intensities are now grounded in measured usage rather than judgement: the Anthropic Economic Index's task-level Claude.ai usage data, aggregated (task→O*NET-SOC) to the 22 SOC major occupational groups — re-derived reproducibly from the ~3,500 usage-weighted tasks on 2026-07-06 (see derive_occupations_aei.py), matching the published chart to two significant figures (Handa et al. 2025 — Computer & Mathematical 37.2%, Arts & Media 10.3%, Education 9.3%, Office & Administrative 7.9%, …, Farming 0.1%), normalised so the most-exposed group (Computer & Mathematical) = 1.0 and held distinct from the separate economy-wide adoption factor so the two are not double-counted. Because that index reflects Claude.ai consumer traffic — it excludes Enterprise and API use, and so under-represents enterprise-heavy occupations such as office support, sales and healthcare — the weight we publish is a 50/50 blend of this measured pattern with the prior structural estimate rather than the raw figure: a deliberate hedge against the consumer-traffic lens. Grounding the weights this way concentrates exposure toward software and quantitative work and moves the benefit dial's 10-year mid from ≈1.2 to ≈0.8. A full task-level re-derivation (aggregating the index's ~20,000 O*NET task-usage rows up to occupation) remains the eventual refinement. The per-task productivity coefficient (ρ) is likewise now explicit rather than asserted: a task-share-weighted average of four randomised trials — Cui et al. (software, +26% output per week), Noy & Zhang (professional writing, −40% time), Brynjolfsson, Li & Raymond (customer support, +14% per hour) and Dell'Acqua et al. (consulting, +12% tasks) — each weighted by its occupation's share of Claude conversations; that weighted mean is 0.26–0.30, so the mid is held at 0.27 and the 0.15–0.40 band spans the lab-to-field generalisation discount and the throughput-optimistic read (full table in data/curated/rct_effects.csv). Adoption — the share of firms actually using AI — is read from the Census Business Trends and Outlook Survey across its November-2025 question break: from ~10% on the old "AI in producing goods or services" wording to ~20% on the broadened "AI in any business function" wording (mid-2026), with the mid set at 0.15; counting firms rather than workers, it if anything understates employment-weighted use. Published explicitly as an estimate with bias in both directions, gross of harms; harm indicators (electricity prices in data-centre states, layoff notices) are tracked separately rather than netted.
The earnings dial (x) against the benefit dial (y). The current quarter is a single dot coloured by the financing-fragility dial — green cash-funded, amber leverage-rising, red debt-carried — with AI's own quarter-by-quarter path trailing behind it since 2023, and a short vertical reach showing where the benefit sits at the 10-year horizon and at the 30-year. Past booms (railways, electrification, the telecoms-and-fibre bust) sit behind as labelled regions, each with an arrow for the ordinal direction it travelled, placed from economic-history figures rather than the live ratios — so that comparison is qualitative, not cardinal. A companion view plots the same quarter in raw dollars — private-earnings flow against social-value flow, the dot sized by the annual capital cost — which de-couples the reading from the interest-rate environment. Both charts render the published JSON directly in the browser, so the same files reproduce the same picture for anyone.
Because the project's whole claim is "levels are unreliable; the movement is the signal," a single quarter's dot cannot test it. So each number is recomputed at past quarter-ends from data we already hold, producing a real trajectory back to early 2023. The capital stock (the denominator of both coverage dials) and the fragility index are rebuilt from filed capex and operating cash flow truncated at each quarter-end — the natural "as of then" stock. The earnings numerator uses the most recent AI-revenue figure a firm had stated on an earnings call as of that quarter, so it sits at zero until firms first disclosed AI revenue in 2025 and then switches on — an honest depiction, not a backward-projected guess. The benefit numerator is a forward productivity-value projection with no quarterly history, so it is held fixed and only its denominator moves through time; this is stated on the chart. Every other input (depreciation, margin, adoption bands) is a fixed assumption, so any movement on the fixed-mid line is real rather than a methods artifact.
On top of the fixed-mid line sits a ±20% revision band — not the wide assumption range, but the plausible restatement of the mid itself from data revisions. The call rule is deliberately strict: a move counts only when the mid leaves the prior quarter's revision band and the next quarter confirms it (a two-quarter rule), so a single noisy quarter is never called. This is the operational answer to the fair criticism that "watch the movement" is empty if the bands are wider than any plausible quarterly move.
Arrow (1962); Nordhaus (2004) "Schumpeterian Profits"; Jorgenson (1963) on user cost; Minsky (1986); Kindleberger (1978); Bernanke, Gertler & Gilchrist (1999); Jordà, Schularick & Taylor (2015) "Leveraged Bubbles"; Perez (2002); Janeway (2012); Pástor & Veronesi (2009); Greenwood, Shleifer & You (2019) "Bubbles for Fama"; Goetzmann et al. (2026); Kogan, Papanikolaou, Seru & Stoffman (2020) "Left Behind"; Noy & Zhang (2023, Science); Brynjolfsson, Li & Raymond (2025, QJE); Dell'Acqua et al. (2023; 2026, Organization Science); Cui et al. (2026, Management Science); Quinn (2018) on the bicycle mania; Garber (1989) and Goldgar (2007) on tulip historiography. Exponential View, The State of the AI Economy (2026) — the independent bottom-up reconstruction used to calibrate the mid-2026 depreciation band. Full citations in the repository's METHODS.md.