In November 2022, OpenAI released ChatGPT and quietly ignited the most concentrated burst of corporate capital spending in modern history. What followed was not merely a technology trend but a financial phenomenon that has reshaped stock markets, geopolitical strategy, and boardroom priorities on every continent. Today, as the bills begin to arrive, investors and economists are wrestling with a question that carries trillions of dollars in consequence: is artificial intelligence a genuine economic revolution, or the most expensive collective hallucination Wall Street has ever produced?

The honest answer, supported by the data, is that it may be both — simultaneously one of the most important technological transitions of our era and one of the most financially distorted investment cycles in history. Understanding where the line falls between those two realities is the central challenge for anyone with money, or ambition, in the AI economy.

The Scale of the Bet

Start with the numbers, because they are genuinely staggering. Total corporate AI investment reached $252.3 billion in 2024 alone, a 44.5% jump from the prior year. The five largest hyperscalers — Google, Microsoft, Amazon, Meta, and Oracle — collectively logged $241 billion in capital expenditures in 2024. That figure is expected to exceed half a trillion dollars in 2026. Add in smaller cloud providers, sovereign AI programmes, and national infrastructure projects, and total AI infrastructure spending for 2025 is conservatively estimated at $600 billion by TCW Group, potentially cresting $1 trillion by 2027 or 2028.

For context: Big Tech firms invested a cumulative $750 billion in data centres between 2024 and 2025. According to financial investor Paul Kedrosky, AI data-centre capital expenditure in 2025 already surpassed the peak of telecom spending during the entire dot-com boom. The Apollo Global Management chief economist Torsten Sløk has noted that data-centre investment’s contribution to U.S. GDP growth in the first half of 2025 equalled that of consumer spending — an extraordinary displacement of economic weight.

At the firm level, the numbers become almost surreal. Microsoft’s AI infrastructure guidance points to $80–120 billion in capital expenditure, equal to 33–50% of its 2024 revenues. Meta is expected to spend roughly $70 billion — nearly 40–45% of revenue. Amazon, approximately $100 billion; Alphabet, $85–90 billion. These are mature, highly profitable businesses voluntarily reinvesting an unprecedented fraction of their earnings into a single technology bet.

Markets in the Mirror

The stock market reaction has been breathtaking. JP Morgan Asset Management estimates that AI-related stocks have accounted for roughly 75% of S&P 500 returns, 80% of earnings growth, and 90% of capital spending growth since ChatGPT launched in November 2022. The S&P 500 itself rose approximately 58% between 2023 and 2024, driven overwhelmingly by the so-called Magnificent Seven — Alphabet, Apple, Amazon, Meta, Microsoft, Nvidia, and Tesla. By October 2025, Nvidia’s market capitalisation alone exceeded $5 trillion, making it more valuable than the entire GDP of every country except the United States and China.

Nvidia’s financial performance is, by any measure, extraordinary. The chipmaker posted record profits of $32 billion in one year, a 65% jump. Its share price has increased approximately 300% over two years. For those invested early, these are generational returns. Yet even Jensen Huang, Nvidia’s chief executive, opened a recent earnings call by proactively attempting to deflate bubble concerns — an unusual posture for a CEO whose company sits at the apex of the boom.

Valuation metrics offer mixed signals. As of late 2025, the S&P 500 traded at roughly 22.3 times forward earnings, above its 10-year average of 18.7 but still approximately 10% below the July 1999 dot-com peak of 24.4 times. Technology, at roughly 27 times forward earnings, is the most expensive sector. Fidelity’s analysis suggests that earnings quality remains healthy: S&P 500 aggregate net profit margins of 13% in Q3 2025 topped their five-year average, and earnings restatements have declined from their post-pandemic peak.

“AI-related stocks have accounted for 75% of S&P 500 returns and 90% of capital spending growth since ChatGPT launched.” — JP Morgan Asset Management

The Circular Economy of AI

The most troubling structural feature of the current AI investment cycle is not its scale but its circularity. A detailed examination of major deals reveals a web of cross-ownership and cross-financing that raises fundamental questions about whether demand signals are real or manufactured.

Consider the network of relationships surrounding OpenAI. Nvidia invested $100 billion into OpenAI, which used the capital to purchase Nvidia chips for its data centres. OpenAI simultaneously holds a 10% stake in AMD — a rival chipmaker — while Microsoft, which owns roughly 27% of OpenAI, also spent approximately $35 billion on AI infrastructure in a single quarter. Microsoft accounted for nearly 20% of Nvidia’s annualised revenue as of fiscal year 2025. Oracle entered a $300 billion computing deal with OpenAI, and JPMorgan is reportedly arranging a $38 billion data-centre financing package for Oracle — despite OpenAI expecting its data-centre partner to ‘lose considerable sums of money’ on the arrangement.

Yale School of Management’s analysis summarises this structure pointedly: one major vendor grants a chip supplier equity to finance data-centre build-outs, then takes an ownership stake in a rival manufacturer for future chip development. The circular nature is not incidental — it is the architecture of the industry. Paul Kedrosky, cited in multiple analyses, notes that this kind of circular financing was last seen at comparable scale during the dot-com bubble.

Goldman Sachs has found that hyperscaler companies took on $121 billion in debt over a 12-month period — a more than 300% increase from their typical debt load. Some of this leverage is being kept off balance sheets through special purpose vehicles, a practice flagged by D.A. Davidson analyst Gil Luria. The effective economic life of current GPU and ASIC chips is approximately one year, yet creditors are underwriting these assets as if they were 7–15-year infrastructure with stable, predictable cash flows.

The Return on Investment Question

At the heart of the sceptical case is a simple arithmetic problem. CEIBS researchers Mattia Landoni and Renxuan Wang have set out the mathematics starkly. Current AI-related revenues are estimated at $15–20 billion globally. Even assuming a generous 25% gross margin and zero additional capex, breaking even on 2025’s investment alone would require scaling AI revenues to $160 billion per year. To generate a solid return on invested capital, revenues would need to reach $400–500 billion. And 2025 is just one year’s spending — if 2026 capex matches or exceeds 2025 levels, as many expect, the implied future revenue requirement runs into the trillions.

The enterprise adoption data makes this gap more concerning. An August 2025 report from MIT’s NANDA initiative — examining $30–40 billion in enterprise generative AI investment — found that 95% of organisations are generating zero measurable return. Fortune magazine reported that generic AI tools stall in enterprise deployment because they do not learn from or adapt to specific workflows. Swedish fintech Klarna became an emblematic cautionary tale: the company publicly boasted in early 2024 that its AI assistant was performing the work of 700 laid-off employees, only to quietly rehire those workers as gig staff by the summer of 2025.

Nobel Prize-winning MIT economist Daron Acemoglu has been explicit: ‘These models are being hyped up, and we’re investing more than we should. I have no doubt that there will be AI technologies that will come out in the next ten years that will add real value and add to productivity, but much of what we hear from the industry now is exaggeration.’

OpenAI’s own economics are instructive. The company is projected to generate approximately $20 billion in annualised revenue by 2025, yet it committed to spending $300 billion with Oracle over five years — averaging $60 billion per year — while still running significant operating losses. Former Fidelity manager George Noble has noted that OpenAI burns $15 million per day on its Sora video model alone, and one projection has the company running out of cash by mid-2027 absent further capital raises. OpenAI’s valuation, meanwhile, surged from $157 billion in October 2024 to $500 billion less than a year later.

“The AI boom is real, but the financial structure built around it appears to be expanding more quickly than any credible adoption curve can justify.” — Man Group

Key Data Points at a Glance

Sources: MIT NANDA, Goldman Sachs, JP Morgan Asset Management, Fidelity, TCW Group, Yale Insights, CEIBS, Man Group, NPR, Wikipedia (AI bubble), Institute for New Economic Thinking.

The Bull Case: Why This Time May Be Different

Dismissing the AI investment cycle as pure speculation would be its own error. The bull case rests on several substantive foundations that distinguish the current moment from the dot-com era.

First and most fundamentally, the companies doing the spending are overwhelmingly profitable. Unlike dot-com-era firms burning venture capital at unsustainable rates, today’s hyperscalers — Google, Microsoft, Amazon, Meta — generate enormous free cash flow. Fidelity’s analysis points out that companies are largely spending what they earn, not what they borrow. TCW Group notes that 76% of data-centre construction in 2024 and 74% in the first half of 2025 was pre-leased, not speculative — real contracts from real customers, with vacancy rates of just 2%.

Second, supply constraints are genuine and persistent. All major AI cloud providers are reporting supply shortages expected to persist into 2026. This is not a demand-side illusion; the bottlenecks run through energy infrastructure, semiconductor fabrication, and data-centre construction timelines. JPMorgan recently projected $5 trillion in global AI infrastructure spending over five years, arguing that demand for computing remains ‘astronomical.’

Third, the productivity gains, where they are occurring, are real. Ninety percent of software development professionals now use AI tools. Companies using AI in marketing report average cost reductions of 37% and revenue increases of 39%. Financial services firms deploying AI in compliance and settlement processes report 40% cost reductions. ChatGPT’s user base doubled from 400 million weekly active users in February 2025 to 800 million by September 2025.

The deeper argument is that generalised-purpose technologies — electricity, railroads, the internet — always overshoot in their early investment phase before their benefits diffuse through the broader economy. The current AI build-out, while financially excessive, may be laying infrastructure that will underpin decades of productivity growth. Those who built railroad track in the 1870s frequently went bankrupt; the railroads themselves changed the world.

Systemic Risk and Who Bears It

Man Group’s year-end 2025 analysis identifies what it considers the bubble’s ‘distinctive feature’: the quiet migration of risk away from tech company balance sheets and into institutions that do not necessarily see themselves as making technology bets. Utilities building power infrastructure for AI data centres, insurers underwriting those facilities, pension funds invested in private credit vehicles financing GPU clusters, and retail investment vehicles — all are absorbing exposure to what is, effectively, a bet on GPU cycles and AI adoption curves.

The Bank of England has warned of growing risks of a global market correction due to the possible overvaluation of leading AI firms, adding that investors have not been adequately cautioned about the consequences if AI falls short of market expectations. The concern is not merely financial but structural: when mismatched asset durations meet overstated demand signals, corrections can be swift and contagious.

DeepSeek’s surprise launch in January 2025 offered a preview. The Chinese chatbot’s unexpectedly strong performance sent Nvidia’s shares down 17% in a single trading session, briefly wiping hundreds of billions from market capitalisation before an 8.8% recovery the following day. The episode illustrated how fragile sentiment can be — and how quickly capital can move when the narrative shifts.

Verdict: Revolution With a Price Tag

The historical record of transformative technologies suggests a synthesis that is uncomfortable for both bulls and bears. AI is almost certainly as important as its most credible advocates claim. The technology is real, the applications are emerging, and the infrastructure being built will serve genuine demand for decades. The companies solving fundamental cost and distribution problems — making AI cheaper, faster, and more reliably useful in enterprise workflows — will likely generate extraordinary value.

What is also almost certainly true is that the financial architecture constructed around AI has become oversized relative to any plausible near-term adoption curve. Circular financing, off-balance-sheet leverage, chip obsolescence cycles mismatched with debt durations, and 95% enterprise pilot failure rates are not the characteristics of a market pricing risk accurately. Some combination of write-downs, consolidation, valuation corrections, and financial institution losses appears probable.

For investors and business leaders, the practical implication is a familiar one from every prior technology cycle: the technology wins even as many of the companies and financial structures built around it lose. The internet transformed the global economy; most of the companies that built it in the 1990s did not survive. Electricity reshaped industrial production; the utilities that over-invested in generation capacity during the boom faced decades of poor returns.

AI will likely follow the same pattern at different speed. The question is not whether the revolution is real — it is — but whether you are holding the infrastructure bonds when the debt comes due, or the applications that make the infrastructure indispensable.