AI Bubble Dashboard

In the eighteen months to June 2026, the share of hyperscaler capital expenditure funded by new debt went from 9% to 32%, and the aggregate free cash flow of the eight largest hyperscalers turned from plus US$180bn to an estimated minus US$64bn. Almost every published comparison between AI and the dot-com bubble argues from valuation — and on valuation the comparison fails, because the Nasdaq-100 trades near 25 times forward earnings against roughly 60 times at the March 2000 peak.

This dashboard argues that valuation is the wrong instrument. Capital cycles do not end when multiples reach a number; they end when the marginal source of funding stops. Measured on financing structure rather than price, the AI cycle sits in late 1998 or early 1999 — and the two indicators that have already tripped are both financing, neither is valuation.

What the Radar shows
A phase verdict, and the disagreement underneath it
Five dimensions are dated independently against the 1994–2001 cycle, then composited with a weighting that favours financing structure. The spread between them is the finding: the dimensions reading late are physical, and the ones reading early are behavioural.

The five clocks
Retail speculation and index valuation read 1996–97: 53% of 2025 IPOs were unprofitable against 76% in 1999 and 81% in 2000, on Jay Ritter’s series. Financing structure reads 1998–99: incremental debt tripled as a share of capex in eighteen months, and vendor financing has reappeared at scale. Credit stress in the levered periphery reads 1999–2000. And capex intensity with index concentration reads past 2000 outright — AI capex heads to 3.1% of GDP against a 1.2% peak for the entire 1990s telecom buildout, while the S&P 500’s top ten hold 38–40% of the index against roughly 27% at the bubble’s high.

Seven falsifiable tripwires
Each carries a stated threshold, set at the level where the 1990s cycle turned, plus a written status and an icon — so colour never carries the meaning on its own. Two have already tripped; both are financing, neither is valuation. The verdict moves toward 2000 as more of them trip, which makes the thesis testable rather than rhetorical.

Four rebuilt exhibits
Physical scale — AI and data-centre capex as a share of US GDP, against the 1.2% telecom peak.
Free cash flow crossing to negative in 2026 — aggregate FCF for the eight largest hyperscalers.
Then versus now — three independent readings; two say we are earlier than 2000, one says we are already past it.
Where credit is pricing stress — five-year CDS across the core and the levered periphery.
Every chart carries a table view, so the numbers are readable without interpreting a picture.

A signals log
Appended on each scan. Only entries that move a tripwire reading or contradict the standing verdict are recorded — routine commentary is discarded.

Two numbers this study corrects
Both circulate widely and both are wrong. The March-2000 trailing P/E figures in general circulation are folklore-grade and range 100–200x depending on how loss-makers are netted; only the ~60x forward figure is defensible, and that is what is used here. And the frequently cited “trillion dollars of telecom debt” refers to industry debt outstanding in 2002, not defaulted principal — a distinction routinely garbled in the comparison literature.

What you get
The live AI Bubble Radar — five clocks, seven tripwire gauges, four exhibits and the signals log, responsive with light and dark mode.
The full Excel data pack — every series, its source, and the estimates flagged as estimates. Where sources disagree it carries the range rather than a point estimate.
Updated on a daily scan, with the date of the last scan shown on the dashboard.
Free — we just need an email
No payment and no trial. Give us an address and the Radar opens immediately and stays on your account. One Atlas per address, so it is worth picking the map you actually work in — a PRO subscription opens the whole series along with the interactive models and the research studies.

Independent educational analysis by CoreValue Research; not investment advice. Figures marked E are estimates and are revised on each scan. Third-party figures are reproduced from their cited sources and dated at the point of use.