The Correction Event
Part III of a series: Part I: The Global AI Bubble | Part II: The Bubble Is Accelerating
I. The Bubble That Became a Safety Net
The correction has already begun; it just looks like stimulus. What started as a speculative build-out now functions as economic ballast. Governments can’t deflate the AI bubble because the bubble is the growth. A trillion-dollar construction cycle props up GDP prints, employment metrics, and energy demand. To stop building would be to admit how much of “the recovery” was architecture, not productivity.
In 2023–2024 the frenzy still looked private—venture money chasing compute. By late 2025 it’s effectively public infrastructure: subsidized, securitized, and politically indispensable. Every hyperscaler expansion reads like fiscal policy in disguise. Legislators brand it innovation, central banks treat it as resilience, and balance sheets quietly merge with budgets. The market no longer clears; it sustains. The reflexive furnace that once burned inside corporate accounting has been nationalized.
The system can’t tighten without self-immolation. Monetary policy is functionally targeting confidence, not inflation. The capex itself is the stimulus. Keynesian multipliers are routed through GPUs.
II. The Reflexive Furnace as Policy Instrument
GDP and payroll prints now track AI capex cadence more than productivity. When Deutsche Bank warned in September 2025 that “AI machines are saving the U.S. economy,” it wasn’t metaphor—it was accounting. Every new data center means construction jobs, utility contracts, and taxable equipment purchases; it’s industrial policy masquerading as innovation.
Governments have learned to use the hype cycle as fiscal channel. Subsidies and accelerated-depreciation rules make private overbuild politically desirable. Energy-grid upgrades are justified as “national competitiveness.” Export controls turn semiconductor scarcity into patriotic theater. The result is reflexive Keynesianism—stimulus routed through corporate hype, where narrative replaces demand as the organizing principle.
Once capital becomes policy, the feedback loop hardens. The same reflexivity that inflated valuations now underwrites employment and tax revenue. Deflating the bubble isn’t correction; it’s contraction. And everyone in charge knows it.
The original mispricing was temporal—building infrastructure for a 2040s economy on 2025 balance sheets. The October acceleration showed the gap widening to 125:1, spending increasing while revenue stagnated. Now we see why stopping is impossible: the infrastructure spending has been absorbed into the fiscal apparatus itself. It’s no longer just market dynamics—it’s how governments maintain the appearance of growth.
Herbert Dreyfus documented the first version of this cycle in 1972. What Computers Can’t Do wasn’t just critique—it was prophecy wrapped in phenomenology. The GOFAI researchers promised human-level intelligence within a decade. DARPA funded it as strategic infrastructure. Universities built labs. Then the AI Winter hit in 1974, and again in 1987, because symbolic AI couldn’t deliver what it promised. Dreyfus’s diagnosis was simple: they’d mistaken simulation for understanding, formalization for intelligence.
The current bubble follows the same structure, just with better graphics cards and worse epistemology. At least the GOFAI researchers knew what they were trying to build—formal reasoning systems. Today’s AI builders can’t explain how their models work, why they work, or what ‘working’ even means beyond benchmark numbers. Dreyfus saw the first bubble because he understood what intelligence actually required: embodiment, context, skills that can’t be formalized. The symbolic AI people ignored him. The neural net people think they’ve solved the problem he raised. They haven’t—they’ve just made the outputs more impressive while the fundamental confusion remains.
Interlude: The Interest Rate Misdirection
The era of free money never ended—it just changed employers.
The standard narrative goes like this: Zero Interest Rate Policy flooded markets with cheap capital from 2008 through 2021. Speculative bubbles formed across asset classes as investors chased yield in a world of negative real returns. Then the Fed raised rates aggressively in 2022-2024, increasing the cost of capital from near-zero to 5%+. Bubbles should deflate. Unprofitable ventures should collapse. Market discipline should reassert itself.
By this logic, the AI infrastructure boom—with its 125:1 spending-to-revenue mismatch and $1.5 trillion annual capital deployment—should have already imploded. The fact that it hasn’t suggests something fundamental has changed.
What Actually Happened
The AI bubble didn’t deflate when rates rose because it stopped being market-financed. Three substitution mechanisms kicked in simultaneously:
Private credit replaced central bank liquidity. Apollo, Blue Owl, and KKR’s infrastructure funds became the shadow Fed for AI buildouts. Oracle’s $38 billion in structured data center debt didn’t come from traditional bank lending—it came from private credit desks packaging long-term commitments into securities sold to yield-hungry institutions. When bank lending tightened, private credit expanded to fill the gap.
Fiscal policy replaced monetary policy. Tax credits, accelerated depreciation schedules, and energy subsidies created synthetic ZIRP specifically for AI infrastructure. A hyperscaler building a data center in 2024 faced:
Investment Tax Credits reducing upfront costs 20-30%
100% bonus depreciation allowing immediate writeoffs
State-level energy subsidies and grid upgrade commitments
Federal semiconductor subsidies for chip purchases
The nominal cost of capital rose, but the effective cost for AI infrastructure stayed near zero through fiscal subsidy.
Strategic necessity replaced return requirements. The reclassification from “speculative venture” to “strategic infrastructure” meant projects no longer needed to justify themselves through traditional ROI. National security, technological leadership, and economic competitiveness became acceptable substitutes for profitability.
The result: AI infrastructure spending accelerated even as rates rose. US hyperscaler capex went from $400 billion (2024) to $490 billion (2025) despite the Fed funds rate sitting at 5.5%. The cost of capital increased in theory but was neutralized through fiscal substitution in practice.
The Mechanism: When Government Becomes the Bagholder
This is why the Capture Triangle matters. Traditional monetary tightening works by making unprofitable ventures unable to raise capital. But when governments need the spending to continue—for GDP, employment, tax revenue, political legitimacy—they simply reroute the subsidy through fiscal rather than monetary channels.
The mechanism is identical to wartime industrial mobilization, just marketed differently:
World War II: Government bonds finance direct production
AI Boom 2025: Tax incentives and subsidies finance private production that functions as fiscal stimulus
Deutsche Bank’s September 2025 assessment made this explicit: “AI machines—in quite a literal sense—appear to be saving the U.S. economy right now. In the absence of tech-related spending, the U.S. would be close to, or in, recession this year.”
Read carefully: AI spending isn’t contributing to growth—it is the growth. Remove the capex and the recovery evaporates. This is why central banks can’t tighten further. The traditional bubble-deflation mechanism requires the central bank to be willing to induce recession to restore discipline. But when the “bubble” has been absorbed into the growth statistics themselves, tightening becomes self-immolation.
The Evidence Across All Three Blocs
This pattern isn’t US-specific:
United States: Federal semiconductor subsidies, state-level data center tax abatements, DOE grid modernization programs. Private spending reclassified as infrastructure investment eligible for public support.
Europe: EU “Digital Sovereignty” programs, member-state subsidies for local data centers, energy transition funds repurposed for AI infrastructure. All accelerating in 2024-2025 despite ECB rate hikes.
China: Continued building despite 80% idle capacity and VC funding at decade lows. The state simply doesn’t require market validation—local governments build for political advancement, central planning absorbs the losses.
The cost of capital rose globally. AI infrastructure spending accelerated globally. The correlation broke because the funding source changed.
Why This Matters for Collapse Scenarios
The interest rate misdirection explains why traditional bubble-deflation timelines don’t apply. When people ask “how can this continue with rates at 5%?” they’re assuming market financing. But AI infrastructure hasn’t been market-financed since 2023. It’s been state-subsidized private construction functioning as fiscal stimulus.
This is what makes the correction so much harder to execute:
In a market-financed bubble, rising rates force discipline through cost of capital
In a state-subsidized bubble, “discipline” means voluntarily inducing recession
No government will choose that path until forced by crisis
The three scenarios I outlined—Soft Nationalization, Hard Landing, Managed Decay—all assume the correction happens despite government efforts to prevent it, not because of deliberate policy tightening. The interest rate channel is blocked. The correction will come from:
Revenue reality overwhelming narrative (consumer AI staying at $12B while infrastructure hits $2T+)
Physical constraints (energy grid limits, utilization staying below 30%)
Credit events in the private financing structures (Oracle debt, pension fund distress)
Political limits to continued subsidy (fiscal crises forcing reallocation)
Rate hikes were supposed to puncture the bubble. Instead, they proved it wasn’t a bubble at all—it was industrial policy wearing venture capital’s clothes. And industrial policy doesn’t deflate through interest rates. It deflates through fiscal exhaustion, political crisis, or the eventual recognition that the infrastructure was built for a future that isn’t arriving on schedule.
The wrong clock is still running. But now it’s synchronized with the fiscal calendar, not the market’s discount rate. And that makes the correction both more inevitable (governments can’t subsidize 125:1 mismatches indefinitely) and more dangerous (when it breaks, it breaks through the state apparatus itself).
III. The Capture Triangle
Every modern bubble ends the same way: policy becomes portfolio management. The AI build-out has reached that stage.
Fiscal Capture
Fiscal capture came first. Politicians discovered that AI capital expenditure looks like prosperity—tax receipts, ribbon cuttings, and “innovation corridors.” Every regional subsidy buys a headline and a quarter point of GDP. The political economy now depends on continuous construction; to pause would be to reveal the recovery as scaffolding.
Local governments compete for data center projects with tax abatements and utility guarantees. State legislatures fast-track permits. Federal programs subsidize energy infrastructure under the guise of “modernization.” The fiscal mechanism no longer distinguishes between productive investment and speculative overbuilding—it just counts the spending.
Monetary Capture
Monetary capture followed. Central banks can’t tighten meaningfully without detonating the debt that funds the build-out. Private credit desks—Apollo, Blue Owl, KKR—have become the shadow Fed, rolling loans into new data-center vehicles to avoid marking losses. Tightening becomes forbearance by another name.
The circular capital structures I documented in Part I (Nvidia investing in OpenAI, Oracle securitizing infrastructure debt) now operate with implicit central bank backing. Not formal guarantees, but the understanding that systemic risk means system support. Too big to fail has been rebranded as too strategic to question.
Narrative Capture
Then narrative capture sealed the triangle. “AI leadership” became the moral alibi for liquidity operations. Every bailout can be framed as competitiveness, every deferral as “strategic patience.” To question the logic is to sound anti-innovation, anti-worker, or worse—anti-future.
The same pattern appears across all three economic blocs:
United States: “Maintaining technological leadership”
Europe: “Strategic autonomy in digital infrastructure”
China: “Self-reliance in core technologies”
Different rhetoric, same mechanism: the bubble has become too embedded in national identity to deflate without political crisis.
Once all three captures lock, no lever works independently. The result isn’t policy; it’s momentum management.
IV. The Inevitable Intervention
No system this entangled collapses cleanly. It stages its own rescue.
Act I: Controlled Burn
The language shifts: “efficiency,” “responsible scaling,” “phase two.” Capital expenditure slows at the margins while governments announce retraining grants and “AI transition funds.” It’s a theatrical pause, not reform.
We’re already seeing the early signals. In September 2025, the FTC announced a review of AI infrastructure investment patterns. Industry groups formed the “Responsible AI Infrastructure Coalition.” The White House released a policy paper on “sustainable compute.” None of this addresses the core temporal mismatch. It’s signaling designed to buy time, to make it appear that someone is managing the risk. The building continues, just with better branding.
Act II: Asset Repricing
Valuations compress 30–40%; pensions publish “long-term positioning” letters. Legislatures hold hearings that generate more acronyms than solutions. Markets call it volatility; economists quietly call it triage.
This is where European pension funds face their reckoning. The 39% allocation to US tech equities that I documented in Part I becomes a political crisis. Retirement shortfalls appear. The ECB’s warnings prove prophetic, but by then the exposure is locked in.
The asymmetry becomes visible:
Executives who exercised stock options 2023-2025 are already out
Fund managers collected fees during the growth phase
Workers and retirees hold depreciated assets with no exit
Act III: Nationalization by Another Name
Idle infrastructure is folded into “sovereign compute” programs, energy-transition initiatives, or defense modernization projects. The stranded racks become public property via acronym. It’s the same mechanism China used for ghost cities—absorb the failure, rename it policy.
The US might call it “National Strategic Computing Reserve.” Europe might frame it as “Digital Sovereignty Infrastructure.”
China already has its National Integrated Computing Network.
Different names. Same mechanism: state absorption of stranded private capital.
Each act buys time, not resolution. The circular capital loop becomes a circular fiscal one, and the furnace that once burned on venture fuel is now wired into the treasury.
V. Monetary Policy Without Money
Every GPU rack is a Treasury bill in drag. The distinction between fiscal stimulus and corporate investment has evaporated. What once looked like capex now functions as liquidity provision.
Governments no longer print money; they issue incentives. Tax credits, depreciation schedules, and loan guarantees turn hardware into currency. Every billion in data-center construction props up aggregate demand the way a bond issue once did. The real economy is collateral for the simulated one.
Central banks play along because they must. If AI capex slows, headline GDP softens, employment dips, and the narrative of “technological resurgence” evaporates. Rate hikes are now constrained not by inflation but by fragility—too much tightening and the compute bubble implodes; too little and asset prices decouple entirely. Either way, the balance sheet wins.
This is the mechanism I identified in Parts I and II made explicit: the temporal mispricing (building for 2040s on 2025 timelines) meets the fiscal necessity (GDP requires the building to continue). The system can’t correct because correction means recession.
It’s wartime economics without the war. A managed mobilization cycle where capacity equals employment equals legitimacy. The fiscal deficit is no longer a problem to solve; it’s the mechanism keeping the furnace lit.
When money creation is routed through infrastructure instead of debt issuance, valuation stops being a number and becomes a story. And stories can’t be deflated gracefully.
Consider the mechanics:
Traditional stimulus: Government issues bonds → Central bank buys bonds → Money enters economy
AI stimulus: Government subsidizes infrastructure → Corporations build capacity → Employment rises → GDP increases → Tax revenue funds more subsidies
The loop is identical in function, just routed through private balance sheets. The distinction between public and private capital has collapsed.
VI. Scenarios for the Correction
The question isn’t whether the correction happens—the 125:1 mismatch and 70-80% collapse probability documented in Part II make that clear. The question is which form the correction takes.
Summary of paths (probabilities sum to 100%):
Soft Nationalization (40%): Gradual state absorption of stranded assets
Hard Landing (30%): Credit event triggers rapid cascade and emergency intervention
Managed Decay (30%): Decade-long stagnation with quiet consolidation
Scenario 1: Soft Nationalization (40% probability)
Governments quietly absorb the stranded assets. “Sovereign compute,” “public-AI grids,” “climate compute initiatives.” The names differ; the mechanism doesn’t. Losses are reclassified as investment, and the data centers become monuments to optimism. Inflation stays muted because the write-downs are hidden behind new acronyms.
How it unfolds:
United States: Federal “Strategic Computing Reserve” purchases idle capacity from failed startups and hyperscalers with SPV-shifted capex at above-market rates. Framed as national security infrastructure. Defense Department and energy grid “modernization” absorbs the hardware.
Europe: EU “Digital Sovereignty Program” consolidates stranded data centers under member-state control. Marketed as reducing dependence on US cloud providers. Pension fund losses partially offset through sovereign guarantees.
China: Already happening. The National Integrated Computing Network folds idle capacity into state planning. Local government debt vehicles restructured with central bank support. Officials who championed failed projects quietly reassigned.
Who pays: Taxpayers absorb losses gradually through reduced services or higher taxes. The infrastructure becomes “public goods” that generate minimal return. Workers avoid mass layoffs but face wage stagnation as fiscal capacity is redirected to asset absorption.
Timeline: 2026-2030, rolling implementation as different facilities hit distress.
Scenario 2: Hard Landing (30% probability)
A single credit event—Oracle’s structured debt, Nvidia customer default, or a European pension run—overwhelms the patchwork. Backstops lag, spreads blow out, and the political instinct to “do something” creates emergency nationalizations that look less like policy and more like triage. This is the 2008 replay, only global, and denominated in compute cycles rather than mortgages.
The trigger could be:
Nvidia’s two mystery customers (39% of revenue) defaulting simultaneously
Oracle’s $38 billion in securitized data center debt facing mass downgrades
European pension fund forced selling creating US tech equity cascade
Chinese local government financing vehicle unable to roll debt
Cascade mechanism:
Initial failure breaks circular capital loop
Counterparty exposure reveals hidden interconnections
Private credit markets freeze (Blue Owl, Apollo funds suspend redemptions)
Margin calls force asset sales
Governments intervene with emergency measures
Who pays: Workers face immediate mass layoffs (100,000+ in US across tech and adjacent sectors documented in Part II). European retirees see 20-40% pension value destruction. Credit markets seize, affecting borrowing across all sectors. Bailouts require austerity measures.
Timeline: 2026-2027, compressed into 6-18 month crisis period.
Post-crisis: Governments end up owning the infrastructure anyway, but at greater social cost. The “soft nationalization” endgame arrives via emergency rather than planning.
Scenario 3: Managed Decay (30% probability)
The most likely outcome: a decade of quiet stagnation. The AI sector ossifies into state-sponsored utilities, the hardware glut repurposed for climate modeling and defense workloads. GDP holds steady on paper, wages don’t, and public optimism erodes. Economists call it equilibrium; citizens call it exhaustion.
How it unfolds:
2026-2028: Capex growth decelerates. No dramatic collapse, just “pivoting to efficiency.” Companies quietly defer expansion projects. Some facilities repurposed for inference rather than training. Media narratives shift to “AI 2.0” and “sustainable compute.”
2028-2032: Consolidation phase. Smaller players acquired or shuttered. Big Tech absorbs competitors’ assets at steep discounts. Market concentration increases further. Data centers run at 40-60% utilization—enough to avoid complete write-offs but nowhere near justifying the buildout.
2032+: The infrastructure becomes legacy. “AI infrastructure” rebranded as “digital public infrastructure.” Used for government services, climate modeling, scientific computing—legitimate applications that generate far less revenue than required to justify the initial investment.
Who pays:
Mid-tier companies and their employees (slow bleeding via hiring freezes and stagnation rather than mass layoffs)
Retail investors who bought into AI hype
Pension funds via persistent underperformance over decades
Opportunity costs: resources that could have funded productive investment trapped in underutilized hardware
Timeline: 2026-2035, slow burn over a decade.
Distributional outcome: The losses are real but distributed over time and populations, making them politically manageable even as they compound economically.
VII. The Asymmetry Persists
In every version, the asymmetry holds:
Who profits:
Executives who exercised stock options during 2023-2025 growth phase
Fund managers who collected fees on asset appreciation
Private credit firms that structured the deals
Short-sellers who positioned for the correction
Consultancies and law firms that manage the restructuring
Who pays:
American workers facing layoffs with minimal severance and healthcare loss
European retirees whose pension funds were used as exit liquidity for US tech
Chinese citizens bearing opportunity costs of misallocated state resources
Taxpayers funding bailouts, subsidies, and state absorption of stranded assets
Governance morphs into administration—maintaining the illusion of motion long after momentum has died.
The three scenarios differ in timeline and mechanism, but the distributional outcome remains consistent: gains extracted early and privately, losses socialized late and publicly.
VIII. After the Fire
This is rail mania, not dot-com. The difference matters.
In the 1840s, Britain built 6,000 miles of railway in a decade—triple what the economy could support. Speculation was frenzied, capital was destroyed, the correction was brutal. But the rails stayed. They couldn’t be unmade. And over the next 40 years, as the economy grew into the infrastructure, the overbuilt capacity became the backbone of industrial Britain.
The AI buildout follows the same pattern. We’re constructing 2040s infrastructure on 2025 balance sheets. The data centers won’t vanish when the correction hits—they’ll just sit there, depreciating and underutilized, waiting for demand to catch up. China’s 80% idle capacity isn’t a warning; it’s the template. The West is building its own ghost data centers; we’re just better at the marketing.
Rail mania eventually justified itself—not on the timeline or terms the original investors imagined, but through brute economic evolution. The AI infrastructure might too. Climate modeling, scientific computing, defense workloads—legitimate applications exist. They just generate a fraction of the revenue required to justify the $2 trillion buildout.
The investors get wiped out. The infrastructure persists. And in 2045, when AI usage finally fills the capacity we built in 2025, economists will call it “visionary foresight” instead of what it actually was: a speculative frenzy that accidentally built something useful while destroying everything in its path.
The furnace still glows, but now it’s inside the treasury. The system learned to burn itself to stay warm. What follows isn’t recovery; it’s maintenance. The vocabulary of progress persists because no one can afford silence. We’ll call it “AI 2.0,” “sovereign innovation,” “the post-infrastructure era.” The branding will change; the balance sheets won’t.
It won’t vanish; it will ossify. The pattern’s likely to be closer to the railroad mania than to dot-com exuberance — speculative overbuild hardened into permanent substrate. The rails stayed long after the market that built them collapsed. Datacenters will too. The bubble ends, but the infrastructure becomes the new floor of the economy.
This is the completion of the arc I began documenting in Part I:
The Temporal Mispricing: Building 2040s infrastructure on 2025 balance sheets created a 50:1 mismatch between spending and revenue. The error was in the clock—assuming AI adoption would compress into financial return timelines.
The Acceleration: By October 2025, the mismatch had widened to 125:1. Spending increased 150% while revenue stagnated. The gap wasn’t closing—it was accelerating. The 70-80% collapse probability within 5 years reflected this worsening trajectory.
The Capture: Now we see why correction is impossible through market mechanisms alone. The speculative infrastructure has been absorbed into fiscal and monetary policy. Governments depend on the spending for GDP, employment, and tax revenue. Central banks can’t tighten without triggering recession. The narrative has been nationalized.
When capital becomes policy, collapse becomes governance.
The correction will happen—the math is unavoidable. But it won’t look like market discipline. It will look like emergency nationalization, or managed decay, or gradual absorption into sovereign balance sheets. The infrastructure will persist, operating at a fraction of projected capacity, justified through evolving narratives that have nothing to do with the original speculative logic.
China’s 80% idle data centers aren’t a warning anymore—they’re the template. The West is simply following the same path with different vocabulary. “Sovereign compute” is the new “strategic industries.” “Digital infrastructure” is the new “ghost cities.” The mechanisms converge because the underlying problem is identical: governments trying to govern away a temporal mismatch that can’t be legislated out of existence.
The clock is still ticking. But now it’s government property. And the wrong time keeps getting printed on official documents, validated by policy statements, and embedded in multi-year plans that assume the future we built for will eventually arrive to justify the present we’re financing.
The three blocs I documented—United States, Europe, China—each using different financing models (private capital, pension exposure, state direction) to build the same speculative infrastructure, have now converged on the same endpoint: nationalization by necessity, whether hard or soft, fast or slow.
The cross-national suicide pact I described in Part I has been executed. We’re living in the aftermath. It just hasn’t been officially declared yet because doing so would make the losses real instead of deferred, visible instead of distributed, political instead of technical.
The furnace still burns. It’s just that now when you ask who’s feeding it, the answer is: all of us, through taxes and inflation and pension shortfalls and the gradual recognition that a decade of “recovery” was built on infrastructure for a future that existed in pitch decks but not in revenue statements.
When capital becomes policy, collapse becomes governance. And the clock keeps ticking.