FINANCEMonths to result82% confidence

Schumpeter Boom-Bust Cycle Applied to AI

Map transformative tech waves to time your exit before the 90% failure shakeout

Problem it solves

Cuts through AI hype by locating current conditions inside a documented historical pattern

Best for

Investors and founders timing exposure to AI-adjacent speculative assets and identifying which stack layer survives shakeout

Not ideal for

Picking individual winners within the surviving layer or distinguishing infrastructure vs inference consumption plays

Overview

Why this framework exists

Joseph Schumpeter's creative destruction model maps every transformative technology — railways, telecoms, internet — through a four-stage arc: bank-financed overinvestment, a boom phase where valuations detach from revenue, the technology coming online and undercutting existing businesses, and a slump where roughly 90% of overinvested companies go bust. The infrastructure survives and society benefits long-term; the investors in the top 10% of operators become the rails everyone uses.

Applied to AI in 2026, Keen argues the overinvestment phase is now confirmed: big tech alone (Meta, Amazon, Microsoft, Alphabet, Oracle) is on track for $720 billion in AI infrastructure spend in 2026 against a 5:1 ratio of money spent versus revenue coming in. The 90% AI startup failure rate in 2026 — versus a 70% baseline for general tech — signals the bust has already begun at the startup layer but has not yet reached large-cap infrastructure valuations.

Keen's call: within 24 months of a severe contraction. The strategic question for portfolio positioning is which layer of the AI stack is most likely in the surviving 10%, and whether infrastructure or inference-consumption is more capital-efficient heading into the slump. Inference-as-a-service sits closer to the 'rails' layer than capex-heavy training clusters.

Core principles

5 total
  1. Bank-financed overinvestment is the structural driver of every technology boom — follow the credit, not the narrative.
  2. A 5:1 spend-to-revenue ratio across a sector is a reliable leading indicator of imminent contraction.
  3. 90% of companies in the boom layer will fail; position for the 10% that become durable infrastructure.
  4. The bust is priced into startups before it reaches large-cap infrastructure — lag matters for exit timing.
  5. Inference consumption is structurally closer to 'the rails everyone uses' than capex-heavy training infrastructure.

Steps

5 steps
  1. Identify the technology wave and its financing source
    Establish which transformative technology is attracting bank or institutional credit at scale. In the AI case, the financing source is big-tech balance sheets and venture capital, not retail bank credit — note the structural difference from the dot-com era. Map total sector capex and compare to sector revenue.
    Pro tipUse reported capex-to-revenue ratios from earnings calls; a ratio above 3:1 sustained for two or more years is historically the warning zone.
  2. Locate the current position in the four-stage arc
    Stage 1 = credit expansion; Stage 2 = boom and valuation detachment; Stage 3 = technology deployment undercutting incumbents; Stage 4 = slump and shakeout. Use failure rates, job market data (entry-level hiring trends), and enterprise adoption metrics (% of pilots reaching production) as stage markers.
    Pro tipA 90% startup failure rate alongside 95% enterprise pilot failure to reach production is a strong Stage 3-4 transition signal.
    WarningStages 2 and 3 can overlap for years — the boom can persist at the large-cap layer long after the startup shakeout has begun.
  3. Identify the stack layer most likely in the surviving 10%
    In every historical technology wave, the survivors become infrastructure: AT&T from telecoms, Amazon AWS from the dot-com boom. Ask whether your exposure is to the infrastructure layer or the application layer. Infrastructure survivors typically have lower capex-to-revenue ratios, network effects, and pricing power.
    Pro tipInference-as-a-service platforms (consumption-layer, not training-cluster) are structurally analogous to web hosting companies surviving the dot-com bust.
  4. Set a contraction timeline and adjust position sizing
    Use the spend-to-revenue ratio and failure-rate data to set a probability-weighted contraction window. Keen's 24-month call from early 2026 implies the bust window opens by Q1 2028. Reduce speculative exposure in the boom layer proportionally as each milestone confirms the timeline, and increase dry-powder positions.
    Pro tipThe bust hits sentiment and paper gains before it hits fundamentals — move before the narrative turns, not after.
    WarningDo not conflate 'bust is coming' with 'all AI is worthless' — the infrastructure survivors compound strongly after the shakeout clears.
  5. Monitor leading indicators for bust confirmation
    Track: big-tech capex guidance revisions (downward = bust signal), enterprise AI pilot-to-production conversion rates, AI startup funding rounds (volume and valuation), and entry-level job market data. Keen cites a 13% decline in entry-level hiring as an early confirmation signal.
    WarningGeopolitical shocks (supply chain disruption, energy price spikes) can accelerate the bust timeline by compressing margins before the organic cycle completes.

Checklist

Saved in your browser

Examples

3 cases
The 1990s Dot-Com Bubble

Bank and institutional credit financed a wave of internet infrastructure investment, with spend-to-revenue ratios across the sector reaching multiples far above 1:1. The Nasdaq peaked in March 2000 and fell 78% over 30 months. Roughly 90% of dot-com companies failed. The survivors — Amazon, Google, Akamai — became the infrastructure rails everyone used for the next two decades.

OutcomeThe infrastructure persisted and compounded; the speculative application layer was almost entirely wiped out. Amazon's AWS, built on the surviving infrastructure, became more valuable than the entire dot-com boom's peak market cap.
AI in 2026 — The Current Overinvestment Phase

Big tech is on track for $720B in AI infrastructure spend in 2026 at a 5:1 spend-to-revenue ratio. The AI startup failure rate has hit 90%, and 95% of enterprise pilots fail to reach production. Entry-level hiring is already down 13%. Keen records his call in early 2026: within 24 months of a severe contraction.

OutcomeThe bust has begun at the startup layer but has not yet reached large-cap infrastructure valuations — the framework predicts large-cap compression follows within the 24-month window.
The 1840s Railway Mania

British banks financed a wave of railway construction that massively overshot near-term demand. Hundreds of railway companies were formed; the majority collapsed in the bust of 1847. The track infrastructure survived and formed the backbone of British industrial logistics for 150 years.

OutcomeThe surviving rail operators became highly durable businesses; investors who bought rail infrastructure at bust prices compounded significantly over the following decades.

Common mistakes

5 traps
Treating the startup shakeout as the full bust
The 90% startup failure rate is Stage 3, not Stage 4. The large-cap infrastructure write-down comes later and is typically larger in absolute terms. Reducing exposure only to startups while holding large-cap AI infrastructure misses the second, larger wave of the correction.
Assuming safe-haven assets insulate against a technology bust
Keen explicitly notes that gold has been driven up then down, and Bitcoin is collapsing as well during the macro stress period. In a broad technology bust, correlations across speculative assets increase — diversification into other risk assets does not provide the insulation investors expect.
Confusing high revenue at the platform layer with surviving the bust
Big tech's $720B capex is less than 20% of their revenue — they can sustain the spend longer than pure-play AI companies. But when the slump hits sentiment, even profitable platforms see valuation compression. Revenue survival is not the same as price survival during a bust cycle.
Ignoring the exogenous acceleration risk
Schumpeter's cycle assumes a relatively stable macro environment. Supply chain disruptions (helium shortage, energy price spikes) can pull the bust forward by compressing margins and capital availability simultaneously. The cycle timing is endogenous but the catalyst can be exogenous.
Waiting for narrative confirmation before repositioning
By the time the 'AI bust' narrative dominates financial media, the large-cap correction has already begun. The framework's value is leading, not lagging — act on the capex-to-revenue ratio and failure-rate signals, not on media consensus.

Origin story

How this framework came to be

Steve Keen developed his heterodox macro framework from Hyman Minsky's financial instability hypothesis and applied it to predict the 2008 Global Financial Crisis before it happened. His intellectual lineage runs through Marx, Keynes, and Schumpeter — economists who treated debt dynamics and technology cycles as endogenous rather than externalities to equilibrium models.

Schumpeter's original creative destruction concept (from Capitalism, Socialism and Democracy, 1942) identified bank credit as the fuel for technology booms, with the bust as a necessary cleansing mechanism. Keen extends this by quantifying the signal: when spend-to-revenue ratios reach multiples of 5:1 across an entire sector, historical precedent from the 1990s telecom bubble suggests the slump is within a two-year window. He considers the current AI wave 'much bigger' than the dot-com bubble.

Source

Traced to primary
Source · PODCAST
Financial Crash Expert: In 3 months We'll Enter A Famine! If Iran Doesn't Surrender It's The End!
Steve Keen · 2026
Open source →

Related frameworks

Browse all Finance →