How likely is an AI bust, and how bad would it be?

Model output from the working paper “When the Buildout Breaks”, recalibrated on market, rate and filing data. loading…

The odds

Cumulative probability, median of the pooled estimate, with the 10th–90th percentile range.

Pooled odds across stored runs

Why the four methods differ

By end-2028. Each method is an independent estimate; the headline pools them. Credit markets are usually the outlier on the bust.

Economic bust

Market crash

Fundamentals = Model G revenue simulation · History = past booms · Market = options and credit · Indicators = credit growth, run-up, volatility.

If a bust happens

Two damage models, weighted by the odds above. Figures are averages over all shock draws, not conditional on a bust, so they understate losses in a deep one.

Neocloud debt maturity wall ($B)

Model D vs F, in short

Model D: 12-sector contagion network with legal friction. Fat tail, cliff above a ~22–25% shortfall.
Model F: weekly agent-based model of firms with nine switchable mechanisms. Tail is capped by contracts and rescues.
“S&P 500 down 30%” in Model F is a synthetic index, not the real one.

Inputs and data sources

Inputs fed to the models

Data sources

Credit spreads, option-implied volatility, lab revenue and neocloud backlog have no free feed. They come from manual_inputs.json and are marked manual. Model G stays at its paper calibration.

What to keep in mind

The odds rest on stated judgment calls (when the boom started, how a shortfall maps to Model F, how strongly the Fed and sovereign buyers respond).
Model F ablations are noisy and only loosely paired across seeds; small differences mean nothing.
Model F’s no-shock baseline is not quiet: some seeds lose money even at zero shock.
Live fetchers have had one real-server run. Check the source table for stale or failed feeds.
Credit spreads, implied volatility, lab revenue and backlog are manual inputs.