The Bubble Is Accelerating
When Will It Pop?—A Probabilistic Assessment
I published my analysis of the global AI bubble as a temporal mispricing on October 14, 2025. Within hours, readers asked two questions: “How likely is this to actually collapse?” and “When?”
I went back to the data to quantify it. The answer is worse than I documented: spending has accelerated to $1.5 trillion, the mismatch is now 125:1, and the timeline is compressing.
Part I: The Numbers Have Deteriorated
My original analysis documented a 50:1 mismatch—roughly $600 billion in annual global AI infrastructure spending against approximately $12 billion in consumer AI revenue. That ratio was alarming enough to suggest systemic temporal mispricing: building for a decades-away future on quarterly-earnings timelines.
The October 2025 data shows the situation accelerating, not stabilizing:
Updated Key Metrics:
Global AI Spending: $600B → $1.5T (+150%) (Gartner 2025 total AI spend; mismatch ratio uses infrastructure capex vs consumer revenue to avoid double-counting cloud pass-throughs.)
Consumer AI Revenue: $10-15B → ~$12B (Stagnant)
Mismatch Ratio: 50:1 → 125:1 (+150%)
Collapse Probability: 60-70% (5-10yr) → 70-80% (5yr) (Higher & Faster)
Near-Term Trigger Risk: Not quantified → 40-50% (2-3yr) (New threshold)
The escalation is system-wide, not confined to any single region or player:
United States: Hyperscaler Acceleration
US hyperscalers alone are now projected to spend $490 billion on AI infrastructure in 2025, up from my original $400 billion estimate. Amazon’s individual commitment has reached $75 billion. Citi projects cumulative spending will hit $2.8 trillion by 2029 to support 55 gigawatts of new power capacity.
This isn’t just more of the same—it’s an acceleration into deeper dependency. The circular capital structures I documented (Nvidia investing in OpenAI, which buys Nvidia chips) are now operating at higher velocity with more total capital at risk.
China: Idle Capacity Worsening
China’s infrastructure bust isn’t stabilizing—it’s expanding. Recent reports indicate that 80% of newly built AI data center capacity sits unused. This isn’t legacy overcapacity from 2023-2024; this is fresh construction in 2025 immediately going idle.
The government’s response—building a National Integrated Computing Network to “repurpose” idle capacity—is a band-aid on a demand vacuum. You can’t policy-engineer demand that doesn’t exist. The soft budget constraint I described allows this to continue, but the opportunity costs compound: every yuan spent on empty racks is a yuan not spent on healthcare, education, or infrastructure people actually use.
Europe: Pension Exposure Unchanged
European pension funds maintain their structural vulnerability. The 39% allocation to US equities—concentrated heavily in the AI infrastructure buildout—hasn’t decreased. If anything, the automatic capital flows through index funds and benchmark-chasing have increased absolute exposure even as the underlying risk has grown.
The European Central Bank continues issuing warnings about concentration risk and potential “AI-related asset price bubbles,” but pension fund managers face the same coordination problem: reducing exposure means underperforming benchmarks in the short term, which is career-ending. So the exposure remains.
The Bain Reality Check
Perhaps most damning: Bain & Company’s analysis indicates the AI industry needs to generate $2 trillion in new annual revenue to justify current infrastructure spending trajectories.
Current consumer AI revenue: $12 billion.
That’s not a gap. That’s a chasm requiring a 167x increase in consumer monetization—or a revelation that enterprise revenue (currently mostly cloud pass-throughs) will somehow materialize at unprecedented scale. Neither seems imminent.
Enterprise AI ‘revenue’ largely manifests as higher cloud bills; to avoid double-counting with infra spend, this analysis benchmarks direct consumer monetization as the falsification gate.
Part II: Probabilistic Risk Assessment
The original question was: how likely is this to collapse, and when? Using historical bubble patterns, current trajectory data, and structural vulnerability analysis, here’s the updated risk framework:
Overall Collapse Likelihood: 70-80% within 5 years (2025-2030)
This represents an increase from my original 60-70% estimate, driven by three factors:
Accelerated spending without corresponding revenue growth widens the temporal mismatch
Increased concentration in fewer players amplifies cascade risk when any link breaks
Compressed timeline as sunk costs mount, making orderly de-escalation harder
Timeline Breakdown
Near-Term Trigger (2026-2028): 40-50% probability
A collapse within the next 2-3 years would likely result from:
Capital market shock: Interest rate increases, credit tightening, or major player default
Revenue disappointment: Quarterly earnings consistently missing AI monetization projections
Single-point failure: Nvidia customer default, OpenAI liquidity crisis, or Oracle debt restructuring
The circular capital structures I documented create tightly coupled failure modes. When Nvidia’s two mystery customers represent 39% of revenue, and those customers are financed by Nvidia’s own investment vehicles, a single crack cascades through the entire loop.
Medium-Term Structural Failure (2027-2029): 30-40% probability
A collapse in this window would stem from:
Cumulative overbuilding: $2.8 trillion in spending meets utilization reality
Energy constraints: 55 gigawatts of new power capacity requirements hit grid limits or regulatory opposition
Pension fund rebalancing: European institutions begin forced de-risking after sustained underperformance
This is the “dark fiber” scenario—not a sudden crash but a grinding recognition that the infrastructure was built for demand that won’t materialize on the assumed timeline. Companies don’t collapse overnight; they gradually admit stranded assets, write down investments, and restructure. But the pain compounds over years.
By Region: Specific Vulnerabilities
United States: The Circular Capital Breaking Point
Primary risk: 60-70% probability within 5-10 years
The US model depends on maintaining the circular flow: investment enabling purchases enabling revenue enabling more investment. This works until it doesn’t. Historical precedent (dot-com, 2008) shows these structures break suddenly, not gradually.
Key fragility points:
Nvidia’s customer concentration: 39% of revenue from two customers creates catastrophic dependency
Oracle’s securitization: $38 billion in data center debt packages echo 2008’s mortgage-backed securities
OpenAI’s equity desperation: Receiving AMD warrants to secure chip supply indicates capture, not strength
Worker impact probability: 80-90%
When this breaks, the impact will be broader than the dot-com bust. That crash primarily affected software engineers and venture-backed startups. The AI infrastructure boom has pulled in construction workers, logistics operators, energy sector employees, and manufacturing. The ripple will be wider, and American workers will face it with minimal safety nets: no mandatory severance, healthcare tied to employment, limited unemployment benefits.
The “flexibility” that The Economist praised becomes fragility when the music stops.
Europe: The Pension Crisis
Primary risk: 70-80% probability if US tech valuations crash
Europe’s vulnerability is structural and unavoidable. Pension funds allocating 39% of equities to US stocks—concentrated in the AI infrastructure players—have created direct exposure to American financial engineering they don’t control.
The double exposure compounds the risk:
Direct US holdings: When Nvidia, Oracle, and hyperscaler valuations correct, European pension funds take immediate losses
Domestic AI infrastructure: Europe’s own data center boom (tracking at ~$100 billion) faces the same utilization challenges
Historical precedent: 2008 was a preview
The Eurozone crisis following 2008 saw:
Youth unemployment exceeding 50% in some countries
Austerity measures decimating public services
Political fragmentation and rise of extremist parties
Brexit partially traceable to economic wounds that never healed
This time the exposure is larger (4x the subprime crisis scale), more concentrated (fewer companies, not thousands of mortgage originators), and European labor protections are weaker after a decade of “competitiveness” reforms.
If The Economist’s advice has been followed—weakening worker protections to enable “flexibility”—European workers will face this crisis with less cushioning than they had in 2008.
China: Chronic Waste, Selective Resilience
Ongoing misallocation: 80-90% probability (already happening) Acute system collapse: 20-30% probability
China presents a different risk profile because the state’s soft budget constraint allows continued building despite obvious overcapacity. The 80% idle rate on new data center capacity isn’t forcing a market correction because market mechanisms don’t apply—local officials continue building for political advancement, state-owned enterprises continue investing to meet government directives.
This creates chronic waste rather than acute collapse. The opportunity costs are real but distributed across the population: resources flowing to empty data centers aren’t available for healthcare expansion, education improvement, or infrastructure that serves actual demand. But Chinese workers won’t face mass layoffs American-style because employment isn’t purely market-driven.
The wild card: Can China monetize stranded assets? (50-60% probability they’ll attempt consolidation)
China’s response to ghost cities was eventual state absorption and gradual utilization over decades. They may attempt the same with AI infrastructure—consolidating idle centers into a national platform, repurposing capacity for non-AI compute tasks, or converting them into strategic reserves for geopolitical leverage.
This won’t erase the misallocation, but it might prevent complete write-offs. The question is whether AI data centers prove more or less salvageable than empty apartment buildings.
Part III: Falsification Criteria—Updated Probabilities
My original analysis included specific, measurable criteria for what would prove the bubble thesis wrong. Here’s how likely each is to occur based on current trajectories:
Five-Year Outlook: Scenarios That Would Prevent Collapse
Consumer AI revenue exceeds $100B annually
Likelihood: 10-20%
Currently stagnant at $12B with no killer consumer application visible. Enterprise revenue remains mostly cloud infrastructure pass-throughs rather than true AI product adoption.
Data center utilization sustains above 70%
Likelihood: 20-30%
China already shows 80% idle capacity. New facilities are coming online faster than workloads globally. Capacity expansion continues to outpace actual demand.
Capex shifts from training infrastructure to inference and applications
Likelihood: 40-50%
Some movement visible but training infrastructure still dominates capital budgets. Sunk cost fallacy and competitive dynamics favor continued scale buildout over operational efficiency.
EU pension funds reduce tech exposure by 10+ percentage points
Likelihood: 10-20%
No policy signals emerging. Fund managers remain benchmark-chasing. Career incentives prevent voluntary de-risking even as warnings multiply.
Combined probability all criteria are met: <5%
The falsification criteria were designed to be generous—any ONE of them being met would validate some portion of the buildout and reduce collapse risk. The fact that even the most favorable scenario (China’s state-directed consolidation) carries only 50-60% probability indicates how structurally entrenched the overbuilding has become.
Part IV: Collapse Scenarios
The question isn’t just “if” but “how.” Based on historical patterns and current structural vulnerabilities, here are the three most likely paths:
Scenario 1: Rapid Cascade (50% probability, 2026-2027)
Trigger: Major revenue disappointment coincides with rising capital costs (rate hikes, credit tightening)
Mechanism: The circular capital structures break when any major link fails. Nvidia reports customer payment delays. OpenAI misses revenue projections. Oracle’s RPO-backed securities get downgraded. Any of these creates cascading effects because the dependencies are tightly coupled.
Regional Impact:
United States:
Nvidia-OpenAI investment loop exposed as synthetic demand
Hyperscaler spending cuts trigger immediate layoffs: 100,000+ across tech and adjacent sectors (construction, energy equipment, logistics)
SPVs holding off-balance-sheet AI infrastructure face forced liquidation
GPU secondary markets flood with hardware from failed startups
Europe:
Pension funds experience 10%+ equity losses concentrated in US tech holdings
Triggers pension shortfalls requiring government intervention
Political pressure for austerity measures to shore up retirement systems
EU data center investments become stranded as funding from US sources evaporates
China:
State absorbs domestic overcapacity through policy bank restructuring
Exports AI infrastructure globally at steep discounts (dumping accusations)
Local government officials who championed projects face quiet reassignments
National platform consolidation accelerates but can’t create organic demand
Asymmetry:
Who profits: Executives who exercised stock options 2023-2025; fund managers who collected fees on asset growth; officials who secured promotions via visible projects
Who pays: Tech workers facing layoffs with minimal severance; European retirees seeing pension values drop; construction and logistics workers in suddenly-cancelled projects; taxpayers funding any bailouts or pension gap-filling
Scenario 2: Slow Burn (30% probability, 2027-2029)
Trigger: Energy grid constraints and geopolitical trade disruptions gradually choke growth
Mechanism: No single dramatic failure, but cumulative reality: data centers can’t get power allocations, trade tariffs spike component costs, utilization rates stay below 50% for years. Companies don’t collapse—they quietly write down assets, defer expansions, and gradually exit the space.
Regional Impact:
United States + China:
Trade war tariffs increase infrastructure costs 20-30%
Power utilities deny or delay data center interconnection requests
Hyperscaler capex declines 40-50% year-over-year
Gradual shift to efficiency (like DeepSeek’s model) over scale
Europe:
Dollar hedging strategies fail amid persistent volatility
Forced rebalancing into lower-yield European assets
Pension funds gradually acknowledge “structural underperformance”
No dramatic crash but years of below-benchmark returns
Global:
$2.8 trillion in cumulative spending (2025-2029) hits physical limits
Dark fiber 2.0: vast infrastructure sitting idle, waiting for demand that arrives slowly or never
Market consolidation: Big Tech absorbs failed competitors’ assets at cents on dollar
Asymmetry:
Who profits: Large incumbents with balance sheets to weather multi-year writedowns; short-sellers who positioned early; efficiency-focused startups (DeepSeek model)
Who pays: Retirees first (gradual pension erosion vs sudden crash); workers via hiring freezes rather than mass layoffs; mid-tier companies unable to raise capital; taxpayers via indirect costs (unemployment, social services)
Scenario 3: Mitigated Pivot (20% probability, 2028+)
Trigger: Genuine killer app emerges; consumer AI revenue jumps to $100B+ range
Mechanism: Something like autonomous vehicles achieving mass deployment, AI-powered personal assistants reaching genuine utility, or breakthrough applications in healthcare/education creating sustainable consumer willingness to pay. Revenue growth, while insufficient to justify ALL the infrastructure, validates enough of it to prevent complete collapse.
Outcome: Think dot-com resolution: the internet DID transform the economy, but 85-95% of fiber optic cable still went dark. Most builders went bankrupt. The technology mattered; the timing killed the pioneers.
In this scenario:
Some data centers reach 70%+ utilization serving the killer apps
But the vast majority (60-80% of total capacity) remains stranded
Infrastructure gets gradually absorbed: repurposed for non-AI workloads, converted to other uses, or simply abandoned
Regional Impact:
United States:
Big Tech consolidates: absorbs competitors’ assets at massive discounts
Market concentration increases further (antitrust concerns intensify)
Some workers rehired, but at consolidated companies with fewer total positions
Europe:
Pension funds stabilize after 20-30% drawdowns
“Rebalancing” narrative emerges: “We’re pivoting to the winners”
Political debate over market concentration vs innovation
China:
State declares “successful transition to AI applications phase”
Some capacity repurposed for actual demand; rest quietly written off
Officials credit national planning; critics note massive waste
Asymmetry:
Who profits: Dominant platforms who survive consolidation; early-stage investors who exited during the hype; efficiency-focused competitors who avoided overbuilding
Who pays: Workers at failed/absorbed companies; late-stage investors; bondholders of failed infrastructure ventures; taxpayers if consolidation involves government support
Part V: Why the Acceleration Matters
The deterioration from 50:1 to 125:1 isn’t just a bigger number—it represents a qualitative shift in the nature of the risk.
The Temporal Mismatch Is Widening, Not Closing
Original thesis: Building infrastructure ahead of demand creates a timing problem where the technology will eventually matter, but the financial timeline doesn’t align with the adoption timeline.
October reality: Not only is the timeline misaligned, but the two curves are diverging. Infrastructure spending is accelerating (up 150% year-over-year) while consumer revenue stagnates. The gap isn’t narrowing naturally; market forces aren’t correcting the allocation.
This means:
Less time to course-correct: My original analysis suggested 5-10 years before structural failure. The acceleration compresses this. When spending grows 150% annually while revenue is flat, the breaking point approaches faster. Updated estimate: 40-50% probability of a trigger event within 2-3 years.
Larger stranded assets: Every additional trillion dollars of infrastructure built ahead of demand is another trillion that needs to be written off, restructured, or absorbed when reality hits. The dot-com bubble left perhaps $100 billion in stranded fiber. This will leave trillions in stranded compute.
Faster cascade when it breaks: Higher velocity of capital flows means faster collapse when any link breaks. The circular structures (Nvidia-OpenAI, Oracle securitizations, pension fund allocations) are operating at higher pressure. When pressure increases in a closed system, failures are more violent.
The Prisoner’s Dilemma Intensifies
Each player now has MORE sunk cost, making stopping even harder:
US hyperscalers: $490 billion committed for 2025 alone
European pension funds: Larger absolute positions in US tech
Chinese officials: More megawatts of empty capacity to somehow justify
The coordination failure I documented isn’t resolving—it’s calcifying. Every additional quarter of investment makes the next quarter’s stopping even more costly. Everyone knows the math doesn’t work, but reversing course means immediate pain (stock price collapse, benchmark underperformance, political embarrassment) while continuing means distributed future pain.
So they continue.
Historical Precedent: Accelerating Into the Wall
The 2008 crisis didn’t pop when the first mortgage defaults appeared. It popped when the acceleration of bad loans hit a critical mass and the entire securitization structure revealed itself as circular. The dot-com bubble didn’t pop at the first sign of overvaluation; it popped when the acceleration of capital into unprofitable ventures hit escape velocity.
Both cases showed the same pattern: acceleration before collapse. The final phase of a bubble isn’t slow deflation—it’s manic inflation followed by sudden rupture.
We’re in the manic phase. Spending up 150% year-over-year. New capacity going immediately idle. Bain calculating $2 trillion in new revenue is needed. The ECB warning about concentration risk. Yet the building accelerates.
Closing: The Clock Is Ticking Faster
The core thesis hasn’t changed: we’re building infrastructure for a 2040s economy on 2025 balance sheets. What’s changed is the pace. The wrong clock isn’t just running—it’s accelerating.
The original mispricing:
50:1 infrastructure spending to consumer revenue
$600 billion building for $12 billion in demand
60-70% collapse probability within 5-10 years
The October reality:
125:1 infrastructure spending to consumer revenue
$1.5 trillion building for $12 billion in demand
70-80% collapse probability within 5 years
40-50% probability of near-term trigger within 2-3 years
China’s 80% idle capacity isn’t a warning anymore—it’s a preview. The People’s Republic already lives in the future where the infrastructure is built and the demand never came. They can absorb it through state mechanisms. The West cannot.
Europe’s pension funds aren’t just exposed—they’re increasingly concentrated in exactly the assets most vulnerable to this correction. When the US tech sector corrects, European retirees will pay for American financial engineering they never participated in. The “exit liquidity” role I documented is now operating at higher volume.
The US circular capital structures aren’t just fragile—they’re operating under more load. When Nvidia invests billions in OpenAI to buy billions in Nvidia chips, and this loop is running at higher velocity every quarter, the snap when it breaks will be proportionally more violent.
Three economic blocs. Three financing models. One accelerating coordination failure.
The falsification criteria I laid out remain valid: consumer revenue >$100 billion, sustained utilization >70%, capital shifting to applications over infrastructure. But the probability of any of these occurring before structural failure is now lower than when I first analyzed this, because the acceleration widens the gap faster than organic adoption can close it.
Bain’s calculation is the epitaph: $2 trillion in new annual revenue needed to justify the trajectory. We’re at $12 billion. That’s not a gap that closes gradually—it’s a gap that forces a reckoning.
The music hasn’t stopped. But it’s playing faster, louder, more frantically. Everyone’s still dancing. The floor is just more crowded, the exits more distant, and the building more structurally unsound than when we started.
The wrong clock is running the world. And now it’s ticking faster.
References and Data Sources
This analysis draws on October 2025 data from multiple sources documenting the acceleration of AI infrastructure spending:
Global AI Spending Projections: Gartner’s September 2025 forecast projects $1.478 trillion in total AI spending for 2025, representing a 150% increase year-over-year. This includes infrastructure, software, and related services across all sectors.
US Hyperscaler Capital Expenditure: Major technology companies have announced significant increases in AI-related capital expenditure for 2025. Amazon’s capex guidance indicates $75+ billion with continued growth expected. Aggregate hyperscaler spending on AI infrastructure is projected at approximately $490 billion for 2025, with Citi Research forecasting cumulative spending of $2.8 trillion through 2029 to support 55 gigawatts of new power capacity.
China’s Infrastructure Overcapacity: MIT Technology Review documented widespread underutilization of newly constructed AI data centers in China, with industry reports suggesting 80% of recent capacity additions remain idle. The Chinese government’s response includes the National Integrated Computing Network initiative to consolidate and repurpose stranded infrastructure.
European Pension Fund Exposure: The European Central Bank’s May 2025 Financial Stability Review continues to warn about concentration risks in US tech equities and the potential for “AI-related asset price bubbles” with significant spillover effects given the integration of global equity markets.
Revenue Gap Analysis: Bain & Company’s 2025 technology report indicates that the AI industry would need to generate approximately $2 trillion in new annual revenue by 2030 to justify the infrastructure expansion trajectory, compared to current consumer-facing AI revenue of approximately $12 billion annually. (Bain 2025 tech outlook; $2T incremental revenue by 2030).
Methodology Note: The 125:1 mismatch ratio is calculated using global AI infrastructure capital expenditure ($1.5T) versus direct consumer AI revenue ($12B), excluding enterprise software pass-throughs to avoid double-counting cloud infrastructure spending. This methodology mirrors the approach used in the original analysis but reflects the accelerated spending and stagnant consumer monetization observed through October 2025.
Additional analysis draws on corporate earnings guidance, financial institution research reports, and government policy documents as cited in the original essay. Data on utilization rates, pension fund allocations, and regional spending patterns are compiled from multiple analyst projections and regulatory disclosures.
Note: DeepSeek’s widely-cited $6M training cost represented only the final training run, not total development costs. Later reporting confirmed the actual R&D investment approached $500M—still more efficient than Western competitors but far from the ‘democratized AI’ narrative. This correction strengthens rather than undermines the efficiency-vs-scale thesis: even the most cost-effective approach still requires nine-figure investments for incremental improvements.
References
China built hundreds of AI data centers to catch the AI boom. Now many stand unused. - MIT Technology Review, March 26, 2025
Financial Stability Review, May 2025 - European Central Bank, May 21, 2025
Nvidia Q2 FY26 10-Q Filing - SEC, July 27, 2025