.png){kind=small}
The Truth About the AI Bubble: A Perez–Azhar Playbook for Business Leaders
- Mark Evans MBA, CMgr FCMi
- Oct 6
- 8 min read
Updated: Oct 7
By Mark Evans MBA CMgr FCMI Aka Rogue Entrepreneur
Executive Summary
Are we in an AI bubble? Evidence in 2025 points to a large investment boom with visible froth, but not yet a confirmed bubble. The pattern mirrors past infrastructure overbuilds such as the railways and fibre optics, where capacity arrived before demand, valuations overshot reality, and corrections followed. Those corrections were painful but productive, leaving behind assets that enabled long-term growth.
This report applies Carlota Perez’s long-wave theory of technological revolutions and Azeem Azhar’s Exponential Age framework to examine the AI investment cycle. It finds that AI is in a late installation phase, exuberant, capital intensive, and productive in parts. If organisations develop the complements that translate capacity into results, this period can mature into a deployment era of broad economic value. If not, the market faces a short, sharp correction that will prune excess and reset expectations.
1. Defining the Bubble
A bubble occurs when capital, attention, and valuation run ahead of evidence for sustainable cash flow. Schumpeter’s theory of creative destruction explains how innovation reshapes markets by destroying incumbents and reallocating capital to new frontiers. In this sense, bubbles are not errors but the turbulence that comes with transformation.
Perez adds a deeper insight. Each technological revolution, she argues, begins with an installation period driven by speculative finance. The excitement of a new paradigm attracts capital faster than productive use cases can emerge. When institutions, regulation, and complementary capabilities catch up, the market enters deployment, where durable value is realised.
The critical question for 2025 is whether AI remains in speculative installation or is transitioning into productive deployment.
2. Historical Parallels
Railways: The Long Game of Infrastructure
The British Railway Mania of the 1840s was the defining investment boom of its age. Capital poured into duplicate routes, optimistic projections, and feverish speculation. The crash that followed bankrupted investors and shattered confidence. Yet the physical network built during that mania became the backbone of British industry for a century (Odlyzko, 2010).
Fibre Optics: The Overbuild that Powered the Internet
A century later, the late-1990s telecom bubble repeated the story. Companies laid thousands of kilometres of fibre optic cable, far exceeding immediate demand. Prices collapsed, firms failed, and investors fled. But the so-called “dark fibre” left behind would later power the modern internet (Richmond Fed, 2003; Wired, 2004).
The lesson from both eras is that speculative overbuilds can finance the infrastructure of future productivity. The issue is not waste but timing, paying for progress before it pays back.
3. The New Economics of Capex Cycles
AI’s capital intensity rivals that of historic infrastructure waves, but its economic profile is completely different. Railways and fibre were amortised over decades. They generated physical throughput (goods, passengers, data) with stable utility and long payback periods. Their durability allowed patient capital to compound.
AI infrastructure, by contrast, ages in dog years. GPUs and servers are refreshed every two to three years as architectures evolve. The value of each generation falls rapidly once new chips outperform them on cost per computation. That compresses the payback window and raises the threshold for justified spending.
Goldman Sachs (2024) estimates over a trillion dollars in AI-related capital expenditure, while Citi projects roughly 2.8 trillion dollars by 2029 (Reuters, 2025). Unlike railways or power grids, much of this spending creates computational capacity with a short half-life. The danger is that the depreciation curve outpaces adoption and revenue growth.
Hyperscalers have already extended the accounting lives of servers and network gear to about six years (Alphabet, 2025; The Register, 2024). Yet GPU refresh cycles remain fast, with NVIDIA’s Blackwell architecture ramping in 2025 (EE Times Asia, 2025). The energy and real estate supporting these systems (grid upgrades, substations, and cooling infrastructure) may last decades. This mismatch ties short-cycle technology to long-lived assets.
The result is a new kind of risk: the duration mismatch bubble. Rail investors had decades to earn back capital. AI investors might have just twenty-four months before obsolescence arrives. If utilisation lags, sunk assets quickly turn into stranded ones.
Figure 1. Duration Mismatch (authors interpretation)
4. Present-Day Signals
The data centre buildout underway is unprecedented. Citi forecasts 2.8 trillion dollars in AI capital expenditure by 2029, while the International Energy Agency (2025) expects electricity demand from data centres to more than double by 2030.
Power is now the limiting factor. Interconnect approvals and grid capacity expansion are multi-year processes. As one European developer observed, “the grid is the bottleneck, not demand.”
The United States remains the epicentre of this expansion, with high equity concentration in AI leaders (Apollo, 2025). GDP growth continues at 3.8 per cent annualised in Q2 2025 (BEA, 2025) however, much of that reflects infrastructure formation rather than productivity gains.
5. Perez and Azhar: Frameworks for Understanding
Perez’s long-wave theory explains why bubbles are often misdiagnosed. Each technological revolution starts with euphoria and overinvestment. The installation phase builds capacity ahead of its economic application. The crash that follows clears speculation and sets the stage for deployment, when institutional frameworks and social norms stabilise the technology.
Azhar (2021) provides the contemporary lens. He argues that technology now improves exponentially while institutions, laws, and human organisations adapt linearly. This widening exponential gap creates the turbulence we see in AI today, extraordinary capability colliding with lagging governance, ethics, and workforce skills.
In this context, AI’s exuberance is not irrational. It is the predictable behaviour of capital chasing exponential tools in a linear world. The danger is not excitement but imbalance: when capital outpaces adaptation, bubbles form.
Figure 2. Bubble Triangle Radar
6. Evidence from Firms and Markets
Empirical evidence supports cautious optimism. Brynjolfsson, Li and Raymond (2025) show that generative AI tools can lift productivity in customer support roles by double digits, especially among less-experienced workers. The OECD (2025) reports similar gains but warns that results depend heavily on complements such as process design, data quality, and governance.
McKinsey (2025) finds that while adoption has broadened, only a small share of firms have achieved meaningful EBIT impact. Many deployments remain experimental. PwC’s 2025 mid-year update notes that governance, not model performance, is now the primary determinant of ROI.
In short and IMO: AI is delivering task-level productivity, not yet enterprise-level transformation. The same was true of electrification before factory layouts changed.
7. The J Curve of AI Value
The productivity path of AI resembles a J curve. Early investment brings higher costs as organisations experiment, train models, and rewire workflows. Returns dip before rising sharply once data, skills, and governance align.
Perez would describe this as the transition from installation to deployment. The firms that cross this inflection point are those that invest heavily in complements such as human capital, data infrastructure, and process redesign.
AI only adds value when embedded into decision-making. As Agrawal, Gans and Goldfarb (2018) note, AI makes prediction cheap, but decisions still depend on judgment and context. The firms that treat AI as a prediction engine without redesigning how they decide will stay trapped in the left side of the J.
Figure 3. The J Curve of AI Value
8. Where the Risk Lives
The greatest danger lies in the short lifespan of AI hardware relative to the long payback of supporting infrastructure. Power facilities, substations, and data-centre shells are twenty-year assets. The GPUs they host may need replacing every two.
If utilisation stalls, the result is a modern echo of the fibre glut, that of vast capacity with falling unit costs and weak revenue coverage. Financing quality is another pressure point. Overreliance on vendor debt, take-or-pay contracts, or customer prepayments can amplify downside risk when markets tighten.
Yet these same investments can become productive foundations. Overbuilt rails became logistics arteries; dark fibre became cloud highways. The outcome depends on management discipline and the speed at which utilisation catches up.
Figure 4. The Four Factor Bubble Screen
9. A Scoreboard for Business Leaders
Business leaders should focus less on predicting bubbles and more on measurable indicators of resilience.
Capacity vs utilisation: monitor how quickly installed capacity translates into real workloads. Rising capacity with flat use is a red flag.
Complements share: at least one-third of AI spend should go to data, skills, process, and governance.
Financing quality: track debt ratios, prepayment reliance, and vendor exposure.
Duration alignment: avoid locking long-life assets to short hardware cycles.
Price signals: sharp falls in compute prices often precede corrections.
This practical scoreboard converts theory into management vigilance.
10. Author’s Verdict
AI in 2025 is not a bubble burst, but a boom under stress. It shares the excesses of past revolutions (overcapacity, narrative heat, and speculative finance) but, also their potential. The technology is very real, the productivity gains are emerging, and the infrastructure being built will form the backbone of the next economy.
As Perez reminds us, bubbles finance progress in advance. The question is whether today’s builders have the patience and governance to turn early excess into lasting value.
References:
Alphabet Inc. (2025) Form 10-K for fiscal year ended Dec 31, 2024. Available at: https://abc.xyz/assets/77/51/9841ad5c4fbe85b4440c47a4df8d/goog-10-k-2024.pdf (Accessed: 4 October 2025).
Apollo Academy (2025) Extreme concentration in the S&P 500 continues. Available at: https://www.apolloacademy.com/extreme-concentration-in-the-sp-500-continues/ (Accessed: 4 October 2025).
BEA (2025) Gross Domestic Product, 2nd quarter 2025 (third estimate). Available at: https://www.bea.gov/news/2025/gross-domestic-product-2nd-quarter-2025-third-estimate-gdp-industry-corporate-profits (Accessed: 4 October 2025).
Brynjolfsson, E., Li, D. and Raymond, L. (2025) ‘Generative AI at Work’, The Quarterly Journal of Economics, 140(2), pp. 889–942. Available at: https://academic.oup.com/qje/article/140/2/889/7990658 (Accessed: 4 October 2025).
Citi / Reuters (2025) ‘Citigroup forecasts Big Tech’s AI spending to cross $2.8 trillion by 2029’, Reuters, 30 September. Available at: https://www.reuters.com/world/china/citigroup-forecasts-big-techs-ai-spending-cross-28-trillion-by-2029-2025-09-30/ (Accessed: 4 October 2025).
EE Times Asia (2025) ‘Blackwell to dominate NVIDIA’s high-end GPU shipments in 2025’, EE Times Asia, 14 August. Available at: https://www.eetasia.com/blackwell-to-dominate-nvidias-high-end-gpu-shipments-in-2025/ (Accessed: 4 October 2025).
Goldman Sachs (2024) ‘Gen AI: too much spend, too little benefit?’ Top of Mind, 27 June. Available at: https://www.goldmansachs.com/insights/top-of-mind/gen-ai-too-much-spend-too-little-benefit (Accessed: 4 October 2025).
International Energy Agency (IEA) (2025) ‘AI is set to drive surging electricity demand from data centres’, IEA Newsroom, 10 April. Available at: https://www.iea.org/news/ai-is-set-to-drive-surging-electricity-demand-from-data-centres-while-offering-the-potential-to-transform-how-the-energy-sector-works (Accessed: 4 October 2025).
McKinsey & Company (2025) The state of AI: How organisations are rewiring to capture value. 12 March. Available at: https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai (Accessed: 4 October 2025).
Odlyzko, A. (2010) ‘Collective Hallucinations and Inefficient Markets: The British Railway Mania of the 1840s’, SSRN Electronic Journal. Available at: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=1537338 (Accessed: 4 October 2025).
Perez, C. (2002) Technological Revolutions and Financial Capital: The Dynamics of Bubbles and Golden Ages. Cheltenham: Edward Elgar. Available at: https://www.e-elgar.com/shop/usd/technological-revolutions-and-financial-capital-9781840649222.html (Accessed: 4 October 2025).
Richmond Fed (2003) ‘Boom and Bust in Telecommunications’, Economic Quarterly, 89(4), pp. 1–24. Available at: https://www.richmondfed.org/~/media/richmondfedorg/publications/research/economic_quarterly/2003/fall/pdf/wolman.pdf (Accessed: 4 October 2025).
The Register (2024) ‘Alphabet banked $3bn by stretching servers’ lifespan’, The Register, 31 January. Available at: https://www.theregister.com/2024/01/31/alphabet_q4_2023/ (Accessed: 4 October 2025).
Wired (2004) ‘Bandwidth glut lives on’, Wired, 30 September. Available at: https://www.wired.com/2004/09/bandwidth-glut-lives-on/ (Accessed: 4 October 2025).
FAQs
Are we in an AI bubble in 2025? Not yet. Evidence suggests a capital-heavy boom with speculative elements but underlying real demand.
Why compare AI to railways and fibre? Both were speculative overbuilds that collapsed before becoming essential infrastructure for the next economic wave.
Why are GPUs risky compared to railways? Their short lifecycles compress payback periods. Capital tied to fast-depreciating assets is exposed if utilisation and revenue growth lag.
How can boards track AI maturity? Measure utilisation against installed capacity, complements share of spend, and price signals for compute.
What is likely to happen next? A correction followed by consolidation. The survivors will own the rails of the next digital economy.