The AI Bubble: A Comprehensive Analysis of Collapse Scenarios

Introduction

I spent last weekend at Kilkenomics in Kilkenny, Ireland. This is economist David McWilliams' festival combining economics and comedy. Across three days of talks, one question dominated every panel: are we in an AI bubble, and if so, what will make it burst?

The rapid rise of artificial intelligence, particularly generative AI models like ChatGPT, has sparked one of the most significant technology booms in history. Global AI spending is projected to reach $375 billion in 2025 and $500 billion in 2026, with AI-related capital expenditures surpassing U.S. consumer spending as the primary economic growth driver in the first half of 2025. The "Magnificent Seven" technology companies (Apple, Microsoft, Google, Amazon, Meta, Nvidia, and Tesla) have seen their valuations soar to unprecedented levels, accounting for the vast majority of stock market gains since late 2022.

However, this explosive growth has raised concerns among investors, economists, and industry analysts about whether the AI industry is experiencing a speculative bubble similar to the dot-com bubble of the late 1990s. Multiple high-profile reports from institutions such as Goldman Sachs [4], Yale School of Management [5], the Brookings Institution [6], and Harvard Business Review [7] have examined whether AI valuations have become detached from fundamental economic realities.

This report synthesises findings from 120 reputable sources to provide a comprehensive analysis of the potential collapse scenarios facing the AI industry. The research examines ten major risk factors, assesses their probability of occurrence, and analyses how these factors might interact to trigger a broader market correction.

Methodology

This analysis is based on a systematic review of 120 high-profile reputable studies, reports, and academic papers published between 2019 and 2025. Sources include:

• Academic research papers from peer-reviewed journals and preprint servers (arXiv, SSRN)
• Major financial institution reports (Goldman Sachs, Bank of America, Bernstein, Deutsche Bank)
• Think tank analyses (Brookings Institution, Yale Insights, CSET Georgetown, Open Markets Institute)
• Industry research (MIT, Stanford HAI, Gartner, Deloitte)
• Government and regulatory studies (U.S. Congress CRS, International Energy Agency, NIST)
• News and media analyses from reputable outlets (Reuters, Forbes, CNBC, The Atlantic, Wired)

Probability assessments for each collapse scenario were determined using a multi-factor framework:

1. Frequency of citation across multiple independent sources
2. Severity of potential impact on AI industry viability
3. Timeline proximity - how soon the risk could materialise
4. Evidence strength - empirical data versus theoretical concerns
5. Mitigation difficulty - how challenging it is to address the issue

Top 10 Reasons the AI Bubble Might Burst

This section details the top 10 reasons the AI bubble might burst, ranked by probability. Each reason is supported by evidence from the 120 sources compiled for this report.

1. Return on Investment (ROI) Failure

Probability: 70-75%

The most significant and immediate threat to the AI bubble is the growing disconnect between the massive capital investment in AI and the tangible returns generated for businesses. A landmark 2025 study by MIT found that a staggering 95% of businesses deploying generative AI have realised no value from their investments, which totalled between $30-40 billion [13, 14, 15]. This "productivity paradox" of AI adoption, where initial implementation hinders short-term productivity, has been noted by researchers at MIT Sloan [17]. With an estimated $3 trillion invested in AI, the lack of significant returns is becoming a major concern for investors and executives alike [10, 11].

The Berkeley Executive Education programme has questioned whether traditional ROI metrics are even appropriate for measuring AI success, suggesting that the industry may be using the wrong framework entirely [16]. Meanwhile, research from MIT Sloan demonstrates that AI adoption in manufacturing firms actually hinders productivity in the short term, contradicting the optimistic projections that have driven much of the investment boom [17].

2. Infrastructure Constraints (Electricity & Power)

Probability: 65-70%

The exponential growth of AI is placing unprecedented demands on global energy infrastructure. Data centres, the backbone of the AI industry, are incredibly power-intensive. According to a 2025 report from the Cato Institute, data centres could consume between 6.7% and 12% of all electricity in the United States by 2028 [25]. Research from MIT Technology Review projects that AI power demand could reach 165-326 TWh per year by 2028 [23]. This surge in demand is already causing bottlenecks in data centre construction and straining power grids [19, 21]. Goldman Sachs notes that data centres accounted for 4% of U.S. electricity consumption in 2024, a figure expected to more than double by 2030 [28].

The International Energy Agency has highlighted that AI-driven data centres are set to drive surging electricity demand, with the U.S. economy consuming more electricity for data processing than for manufacturing all energy-intensive goods by 2030 [20]. Utilities are already grappling with how to power these facilities, with some regions facing resource-adequacy constraints that could severely limit growth [27].

Deloitte's analysis questions whether U.S. infrastructure can keep pace with the AI economy, noting that the build-out of new power generation and transmission capacity takes years and faces significant regulatory and community opposition [18]. The Atlantic has gone so far as to suggest that infrastructure constraints represent the most likely path to an AI crash [26].

3. Market Valuation Correction

Probability: 60-65%

The current valuations of many AI companies are reminiscent of previous technology bubbles. A Bank of America Global Research survey from October 2025 found that 54% of investors believe AI stocks are in a bubble [9]. This sentiment is fuelled by the extreme concentration of market gains in a few large tech companies, often referred to as the "Magnificent Seven," which have accounted for 75% of the S&P 500's returns since the launch of ChatGPT. The gap between the tech sector's market capitalisation and its share of net income has widened significantly since late 2022, leading analysts at Bernstein to declare that a bubble is a "likely outcome."

Global stock markets have already shown signs of fragility, with sharp falls in November 2025 attributed to AI bubble fears [91]. Goldman Sachs' October 2025 report "AI: In a Bubble" notes that AI bubble concerns have returned amid the rise in AI-exposed companies' valuations, ongoing massive AI spending, and increasing circularity of investments [4].

Morningstar analysis reveals that AI investment drove approximately two-thirds of the S&P 500's profit growth in Q3 2025, making the market critically dependent on continued AI momentum. Reuters reports that despite $3 trillion in AI investment, early studies show companies are not yet reaping significant returns, creating a dangerous disconnect between valuations and fundamentals [9].

4. AI Commoditisation and Competition

Probability: 55-60%

The rapid pace of innovation in AI is also leading to its rapid commoditisation. The emergence of powerful open-source models from China, such as DeepSeek, which offer comparable performance to proprietary models at a fraction of the cost, is a significant driver of this trend [29, 31, 37]. As AI capabilities become more widespread and accessible, it becomes increasingly difficult for companies to maintain a sustainable competitive advantage. An article in the MIT Sloan Review from May 2025 argues that AI will not provide a sustainable competitive advantage, but will instead become "table stakes" for businesses [42].

The DeepSeek phenomenon has sent shockwaves through the AI industry and financial markets, demonstrating that cost-efficient training methods and open-source ecosystems can democratise access to advanced AI capabilities [34, 37]. The World Economic Forum reported that DeepSeek has fundamentally shaken up the AI tech sector, challenging the assumption that only well-funded Western companies can develop state-of-the-art models [31].

Academic research by Abonamah, Tariq, and Shilbayeh (2021), cited by 82 researchers, presents a compelling argument that AI is moving rapidly toward commoditisation, proposing new organisational architectures to address the challenges this creates [39]. Forbes analysis suggests that achieving competitive differentiation with AI requires accompanying business transformation, as the risk of commoditisation is happening faster than expected [43].

5. Performance Plateaus and Diminishing Returns

Probability: 50-55%

There is growing evidence that the rapid performance gains in AI models are beginning to slow down. Researchers and industry experts are observing that large language models may be nearing their limits, with new models not achieving the same scaling benefits as their predecessors [83, 84, 85]. The scaling laws that have driven AI progress are showing diminishing returns, and innovation is becoming more expensive [86, 87]. Practical constraints, such as the availability of high-quality data and the sheer cost of training larger models, are also contributing to this slowdown [88].

EDUCAUSE reported in September 2025 that large language models may be nearing their limits, challenging assumptions about the transformative potential of artificial intelligence [85]. Forbes noted in November 2024 that recent research suggests new models aren't getting the same amount of benefit from scaling, marking what some call "The Big AI Slowdown" [84].

Wired's analysis suggests that the AI industry's scaling obsession is headed for a cliff, with models soon offering diminishing returns compared to smaller, more efficient alternatives [87]. Research by Lu (2025) introduces the relative-loss equation, a time- and efficiency-aware framework that extends classical AI scaling laws and acknowledges the reality of diminishing returns [86].

6. Data Quality and Training Data Exhaustion

Probability: 45-50%

The quality and availability of training data are critical for the development of powerful AI models. However, there are growing concerns that the well of high-quality data is running dry. In January 2025, Elon Musk stated that AI companies have "exhausted" the sum of human knowledge for training models [57]. This is supported by research from the MIT Data Provenance Initiative, which found a dramatic drop in the content available for AI training due to restrictions from web sources [56]. Furthermore, the proliferation of AI-generated content is leading to data contamination, where AI models are trained on the output of other AI models, leading to a degradation in quality [58, 59].

Research by Xu, Guan, Greene, and Kechadi (2024), cited by 94 researchers, provides a comprehensive survey of benchmark data contamination in large language models, noting that with the rise of AI-generated content, the issue is becoming more serious [58]. Deng, Zhao, Tang, and Gerstein's 2024 study, cited by 177 researchers, investigates data contamination in modern benchmarks for large language models [59].

7. Geopolitical Risks and Export Controls

Probability: 40-45%

The increasing strategic importance of AI has made it a focal point of geopolitical competition. The United States has implemented export controls on advanced computing items to limit China's access to cutting-edge AI technology [54, 55]. While intended to protect national security, these measures also risk fragmenting the global AI market and harming the competitiveness of U.S. companies.

CSET Georgetown has provided recommendations on export controls for artificial intelligence, assessing options for controlling AI software and examining highly effective points of export control [54]. Duke University research examines the battle for chips and semiconductors' crucial role in AI development, analysing restrictions on China's access to advanced chips [55].

8. Regulatory Compliance Costs

Probability: 35-40%

The rapid development of AI has outpaced the development of regulations, but that is beginning to change. Governments around the world are now considering and implementing new regulations for AI, which will inevitably lead to increased compliance costs for businesses. The American Enterprise Institute estimates that just three AI regulations in the U.S. could cost between $18 billion and $52.1 billion [78]. These costs will be a significant burden for AI startups, which are already operating on tight R&D budgets [77, 80].

Wu and Liu's 2023 research, cited by 6 researchers, provides a field deployment perspective on compliance costs of AI technology commercialisation, noting that these costs have become a huge financial burden for AI startups already constrained on research budgets [77].

9. Market Concentration and Antitrust Action

Probability: 30-35%

The AI industry is already highly concentrated, with a few large tech companies dominating the market. According to the Stanford HAI 2025 AI Index Report, nearly 90% of notable AI models in 2024 came from industry, a significant increase from 60% in 2023 [72]. This concentration of power has raised concerns among regulators about potential antitrust violations. The U.S. Congress, the Federal Trade Commission, and the Department of Justice are all investigating competition in the AI market [70].

Korinek and Vipra's 2024 NBER Working Paper examines market concentration dynamics in foundation models, analysing the evolving structure and competition dynamics of the rapidly growing market for foundation models [62]. The Yale Law and Policy Review argues that an unregulated AI oligopoly has undesirable economic, national security, social, and political consequences [63].

10. Environmental and Sustainability Backlash

Probability: 25-30%

The environmental impact of AI is a growing concern. The massive data centres that power AI consume vast amounts of electricity and water. Google's greenhouse gas emissions have risen by 48% since 2019, largely due to its AI and data centre operations. Training a single large AI model, like GPT-3, can consume as much electricity as hundreds of homes for a year.

Research by Ding et al. (2025), cited by 11 researchers, tracks the carbon footprint of global generative artificial intelligence, compiling data on 369 GAI models released globally from 2018 to 2024 to examine their energy consumption and entire life cycle carbon emission levels. The Environmental and Energy Study Institute reports that a medium-sized data centre can consume up to roughly 110 million gallons of water per year for cooling purposes, equivalent to the annual water consumption of a small city.

Cascade Scenario Analysis

Most Likely Collapse Sequence

The ten risk factors identified in this report are not independent; they interact and reinforce each other in complex ways. The most likely path to a significant AI bubble correction involves a cascade of events:

Phase 1 (2025-2026): ROI Concerns Mount, Infrastructure Bottlenecks Emerge

The cascade begins with the growing realisation that AI investments are not delivering promised returns. As more companies publish results similar to the MIT study showing 95% failure rates, investor confidence begins to waver. Simultaneously, infrastructure constraints become increasingly apparent as data centre construction faces power supply limitations. Energy costs rise as utilities struggle to meet demand, squeezing margins for AI companies.

Phase 2 (2026-2027): Valuation Corrections Begin, Commoditisation Accelerates

As ROI concerns mount and infrastructure limitations become undeniable, the first valuation corrections begin. The "Magnificent Seven" tech stocks experience their first significant decline, triggering broader market nervousness. Chinese open-source models continue to improve and gain adoption, accelerating the commoditisation of AI capabilities. Price competition intensifies, and margins compress across the industry.

Phase 3 (2027-2028): Performance Plateaus Evident, Market Restructuring

Performance improvements in AI models slow noticeably, with diminishing returns from scaling becoming impossible to ignore. Data quality issues and training data exhaustion contribute to the plateau. Companies that have invested billions in ever-larger models struggle to justify continued spending. A wave of consolidation begins as weaker players exit the market or are acquired.

Phase 4 (2028-2030): New Equilibrium Established at Lower Valuations

The market reaches a new equilibrium with valuations reset to more sustainable levels based on actual demonstrated value rather than speculative potential. Companies that have found genuine product-market fit and can demonstrate clear ROI survive and thrive. The AI industry continues to grow, but at a more measured pace with more realistic expectations.

Probability Assessments

Key Differentiators from Dot-Com Bubble

1. Real Technology: AI demonstrably works and provides value in many applications, unlike many dot-com companies that had no viable business model.
2. Established Companies: The AI boom is largely driven by established, profitable tech giants rather than unprofitable startups.
3. Infrastructure Investment: Much of the AI investment is in real infrastructure (data centres, chips, power) rather than purely speculative ventures.
4. Diverse Applications: AI has applications across virtually every industry, providing multiple paths to value creation.

However, these differentiators do not eliminate bubble risk. They simply suggest that the eventual correction will be less catastrophic and will be followed by continued growth rather than a prolonged "AI winter."

What We're Left With: The Railway Tracks Debate

A recurring theme at Kilkenomics 2025 and on David McWilliams' podcast was the question of lasting value: when this bubble bursts, what will we actually be left with? The pessimistic view, articulated by several panellists, centres on the disposability of AI infrastructure. GPUs, the engines of the AI revolution, need to be replaced approximately every 18 months as newer, more powerful chips emerge. Unlike the railway bubble of the 1840s, which left behind permanent physical infrastructure (tracks, stations, bridges) that remained useful for decades, the AI bubble's physical assets rapidly depreciate into obsolescence.

This is a sobering parallel. When Britain's Railway Mania collapsed in 1847, investors lost fortunes, but the nation gained a comprehensive rail network that transformed commerce and society for generations. What comparable lasting infrastructure will remain after the AI bubble deflates? Vast data centres filled with outdated hardware? Power infrastructure built for computational loads that no longer justify the expense?

The AI bubble, regardless of when or how it corrects, is producing similar invisible infrastructure. Transformer architectures, reinforcement learning techniques, efficient training algorithms, open-source model ecosystems, fine-tuning methodologies, and practical deployment patterns are the "railway tracks" of the AI era. Unlike physical GPUs, these innovations don't depreciate. They become shared knowledge, freely available for the next generation of builders.

Perhaps more significantly, the AI boom is fundamentally democratising access to capabilities that were previously locked behind expensive gatekeepers. A solo developer can now access language models, computer vision systems, and code generation tools that would have required entire research teams and millions in funding just five years ago. Barriers to entry across industries—from software development to content creation to data analysis—have collapsed. This democratisation persists regardless of market valuations.

The disposability of GPUs is real, but focusing solely on hardware obsolescence obscures the more durable legacy: the techniques, the training data, the open-source ecosystems, the reduced barriers to innovation, and most importantly, the widespread knowledge of what's possible. These intangible assets, like the economic coordination enabled by Victorian railways, will outlast any market correction and provide the foundation for whatever comes next.

So while the pessimists at Kilkenomics are correct that we won't have physical infrastructure with century-long utility, they may be underestimating the value of the technological infrastructure we're building. This infrastructure exists in codebases, research papers, trained minds, and proven methodologies rather than in depreciating silicon.

Conclusion

The analysis of 120 high-profile reputable studies reveals a complex and multifaceted picture of the AI industry. While the transformative potential of AI is undeniable, the current market is characterised by many of the classic signs of a financial bubble: extreme valuations, concentration of gains in a few companies, speculative investment, and a growing disconnect between spending and returns.

The most pressing risks are ROI failure (70-75% probability), infrastructure constraints (65-70% probability), and market valuation correction (60-65% probability). These three factors are mutually reinforcing and could trigger a cascade effect leading to a significant market correction in the 2025-2027 timeframe.

However, several factors distinguish the AI bubble from the dot-com bubble of the late 1990s. AI technology demonstrably works and provides genuine value in many applications. The boom is driven largely by established, profitable companies rather than speculative startups. Much of the investment is in real infrastructure rather than purely speculative ventures. These factors suggest that while a significant correction is likely, a catastrophic collapse is less probable.

The most likely scenario is a significant correction (30-50% decline) that resets valuations to more sustainable levels, followed by continued growth at a more measured pace. Companies that can demonstrate clear ROI and have found genuine product-market fit will survive and thrive. Those that cannot will fail or be acquired.

The AI industry will emerge from this correction more mature and sustainable, with more realistic expectations and a clearer understanding of where AI can and cannot create value. The technology itself will continue to advance and find new applications, but the speculative frenzy of 2023-2025 will give way to a more grounded and sustainable growth trajectory.

The AI revolution is real, but the current bubble is also real. Navigating the coming correction will require clear-eyed assessment of risks, realistic expectations, and a focus on genuine value creation rather than speculative hype.

References