The Contest Is the Bubble: the BIS Prices AI Overinvestment at 1.5 Times the Efficient Level

The fragility of the AI build-out is a property of competition itself. The Bank for International Settlements has now put a number on the race: capital committed at 1.5 times the socially efficient level, by the very logic of the contest that produces it.

1847: the race itself was the risk

In the 1840s, British railway promoters raced one another for exclusive corridors. Each route was winner-take-most: the first line linking two cities captured the traffic, and the laggard inherited scraps. Parliament sanctioned schemes in waves, capital chased charters rather than traffic forecasts, and the aggregate network absorbed far more investment than any sober projection of demand could justify. The reckoning arrived in October 1847, when the drain on the Bank of England grew severe enough that the government suspended the Bank Charter Act of 1844 to halt the panic. The mechanism deserves precision: individually rational firms, each competing for a dominant position, produced a collectively excessive commitment of capital, an overshoot embedded in the structure of the contest long before the first revenue disappointment. BIS Working Paper 1367published on 14 July 2026 by Phurichai Rungcharoenkitkul, places the AI build-out in exactly this lineage: the American canal mania of the 1830s, the English railway mania of the 1840s, the dotcom boom of the 1990s. All three ended in sharp corrections. Three precedents are sufficient to call it a pattern.

The current pattern

Rungcharoenkitkul's model treats the AI boom as a dynamic contest in a winner-take-most market, and the results are stark. In the baseline, firms invest around 50% above the socially efficient level. Calibrated to balance-sheet and deal data, overinvestment reaches around 1.5 times the efficient level, and climbs to around three times where demand proves less elastic. The scale is historic: hyperscaler capital expenditure is set to exceed $700 billion in 2026and the build-out is on track to outgrow every previous US technology-driven investment episode within three years. Five firms, Alphabet, Amazon, Meta, Microsoft and Oracle, have added approximately $350 billion in debt obligations over five years. The paper's warning is direct: "The larger the boom, the deeper the eventual bust. The race to commit early through debt and circular financing also makes a bust more likely."

According to AGORÀ Intelligence analysis of 4 primary sources, the significance lies in a shift of analytical frame within the official sector. The Bank of England and the Federal Reserve have described AI-related exposures in qualitative terms; the BIS has now produced a structural model with a number attached. Overinvestment ceases to be an anecdote about exuberance and becomes a measurable externality: the contest itself, rather than any error of judgment, generates the excess. That reframing converts AI capex from an equity story into a financial-stability variable, with fire sales of specialised hardware and contagion through chains of circular equity stakes as the identified transmission channels.

The mechanism: overshoot by construction

The model's engine is a contest externality. Each firm's additional dollar of capex raises its own probability of securing a dominant position and lowers every rival's. From inside any single boardroom the spending is rational; from the vantage of the system, part of it exists purely to cancel out rivals' spending. That cancelled portion, around 50% of the efficient level in the baseline, produces data centres, chips and power contracts whose social return sits close to nil. Debt finance and circular equity stakes then convert the overshoot into fragility: specialised hardware supports thin resale markets in a downturn, so a funding shock forces sales at fire-sale prices, and chains of cross-holdings let stress in one firm cascade to others. The paper's sustainability condition follows: the structure holds together under "a strong realisation of the technology's productivity", and it comes apart under revenue disappointment.

Three implications for capital allocation

1. Credit, 6–12 months. Hyperscaler investment-grade paper carries an embedded contest premium. The $350 billion in added debt across five issuers reprices on revenue guidance rather than on default probability; spread widening precedes any downgrade. Position for dispersion between hyperscaler credit and the broader investment-grade index.
2. Collateral, 12–24 months. The BIS identifies fire-sale vulnerability in specialised hardware. GPU-backed lending and data-centre asset finance deserve haircuts calibrated to thin secondary markets, well below the loan-to-value assumptions now common in private credit.
3. Equities, 12–36 months. Winner-take-most implies concentrated winners atop an aggregate overshoot. Index-level AI exposure holds the full 1.5×; single-name selection holds the prize. The trade is dispersion, long identified moats against the capex-heavy field, rather than beta to the theme.

What to watch

• 2027 capex guidance from Alphabet, Amazon, Meta, Microsoft and Oracle in the Q4 2026 earnings cycle (January–February 2027): the first material deceleration in guided growth marks the inflection of the contest (10-K filings and earnings transcripts).
• Debt issuance and circular-financing disclosures in hyperscaler SEC filings: growth beyond the $350 billion five-year base signals deepening fragility (SEC EDGAR; BIS WP 1367).
• Secondary-market pricing for AI accelerators and data-centre lease rates: the earliest observable proxy for the BIS fire-sale channel (industry trackers such as Digitimes).
• Follow-up publications from the BIS, FSB and IMF applying the WP 1367 framework: the paper trail of the supervisory regime change (bis.org, imf.org).