Artificial Intelligence April 3, 2026

The AI Bubble Is Deflating – And the Fallout Could Dwarf the Dot-Com Crash

The numbers tell a story of profound contradiction. Nearly 88% of companies report using artificial intelligence in some capacity, yet only about one-third have scaled it enterprise-wide, and just 39% see measurable financial impact. Meanwhile, the infrastructure buildout powering generative AI has ballooned into a multi-trillion-dollar bet – one increasingly financed by debt rather than earnings, and one that now accounts for an outsized share of both stock market gains and GDP growth.

Three years after ChatGPT ignited a frenzy of corporate AI spending, the gap between adoption metrics and actual returns is widening into a chasm. The S&P 500’s three-year, $30 trillion bull run has been driven overwhelmingly by a handful of tech giants and their AI infrastructure satellites. If those companies stop rising, the indexes – and the retirement portfolios and pension funds tethered to them – follow them down. What’s unfolding isn’t a simple market correction. It’s the slow-motion deflation of what one University of Michigan professor has called an “order-of-magnitude overvaluation bubble,” and the economic consequences could be severe.

The Scale of the Bet

To understand why the current moment feels so precarious, consider the sheer magnitude of capital being deployed. OpenAI alone plans to spend $1.4 trillion in the coming years but projects $115 billion in losses through 2029 before generating positive cash flow in 2030. Globally, an estimated $6.7 trillion must be invested in data centers between 2025 and 2030 to meet AI demand. Hyperscaler companies – Alphabet, Microsoft, Amazon, and Meta – are projected to spend more than $400 billion on capital expenditures in the next 12 months, the vast majority directed at data center infrastructure.

This spending has become the engine of the American economy in ways that few appreciate. Harvard economist Jason Furman estimates that AI-driven infrastructure investment accounted for 92% of US GDP growth in the first half of 2025. Strip out AI-related capital spending, and the US economy would have been in recession. AI capex alone contributed an estimated 100 basis points – a full percentage point – to quarterly GDP, growing tenfold faster than underlying consumer spending.

The Depreciation Trap Eating Big Tech Margins

The financial strain is already visible in the earnings of the world’s largest technology companies. Alphabet, Microsoft, and Meta’s combined quarterly depreciation costs from data centers rose from roughly $10 billion in Q4 2023 to nearly $22 billion in the September 2025 quarter. That figure is projected to hit $30 billion by late 2026.

Metric Q4 2023 Sep 2025 Quarter Late 2026 (Projected)
Combined Depreciation (Alphabet, Microsoft, Meta) $10 billion $22 billion $30 billion
Mag 7 Earnings Growth Strong acceleration Decelerating 18% (4-year low)
Free Cash Flow After Shareholder Returns Positive Narrowing Negative (Meta, Microsoft)

In 2026, Meta and Microsoft face negative free cash flow after accounting for shareholder returns, while Alphabet is expected to roughly break even. Magnificent Seven earnings growth is projected at 18% – the slowest in four years and barely above the broader S&P 500. Big Tech’s historic value proposition was rapid revenue growth at low cost, producing immense free cash flows. The AI pivot has inverted that model entirely, transforming these companies from cash machines into capital-intensive infrastructure builders with uncertain returns.

Why This Isn’t the Dot-Com Bubble – It’s Potentially Worse

Comparisons to the dot-com crash of 2000-2002 are everywhere, but the scale differences are staggering. Consider CoreWeave, the Nvidia-backed AI infrastructure firm that went public in March 2025 at $40 per share. After its latest earnings revealed widening losses and infrastructure constraints, shares plunged 30% in two days, erasing roughly $23 billion in market capitalization. That single two-day loss was 56 times the peak $410 million valuation of Pets.com – the poster child of dot-com excess – which collapsed within 12 months of its February 2000 high.

“It takes a hype-driven tech stock to instantly destroy $20 billion in wealth,” observed Erik Gordon, an entrepreneurship professor at the University of Michigan’s Ross School of Business who researches financial markets and technology. Gordon has warned that “more investors will suffer than suffered in the dot-com crash, and their suffering will be more painful” because tech now comprises a far larger share of the market and far more people are invested through retirement accounts and index funds.

The dot-com collapse saw the S&P 500 fall approximately 9% in 2000, 12% in 2001, and 22% in 2002. Today, US equity market capitalization is nearly twice GDP – significantly higher than at the dot-com peak. An equivalent crash would erase approximately $33 trillion of value, more than total US GDP. The unemployment trajectory from that era offers a sobering precedent: it took 47 months for unemployment to return to previous levels and seven years for the S&P 500 to recover.

The Broken Economics of Generative AI

Beneath the infrastructure spending frenzy lies a fundamental problem with generative AI’s business model. Traditional software economics were built on a simple truth: build once, sell infinitely. The marginal cost of adding a new user approached zero, allowing aggressive scaling with high margins. Generative AI has shattered that model.

Every interaction with a GenAI feature generates real expenses – computing power, database searches, and third-party API fees. A legal-tech firm that integrated an LLM to summarize case files discovered that for every $1.00 in subscription revenue, it spent $1.20 on processing and infrastructure costs. The product worked beautifully. Users loved it. But the company was effectively subsidizing its customers’ work with venture capital.

Autonomous AI agents multiply this cost volatility exponentially. A single user prompt like “Plan my marketing campaign” can trigger 50 to 100 internal reasoning loops, each incurring inference costs. In traditional software, loops cost negligibly. In generative AI, an agent stuck in a reasoning loop is the cloud equivalent of leaving a tap running in a drought. Sequoia Capital has highlighted this structural mismatch as “AI’s $600 billion question” – the massive gap between billions spent on infrastructure and actual revenue realized by applications.

The Debt Financing Tipping Point

Perhaps the most alarming development is the shift in how AI infrastructure gets funded. What began as investments financed from Big Tech’s enormous free cash flows is increasingly being funded by debt.

Recent mega-deals illustrate the complexity and risk. Meta’s $27.2 billion data center financing with Blue Owl combined asset-backed securities, commercial mortgage-backed securities, and investment-grade debt in off-balance-sheet structures. Oracle has sold tens of billions in bonds to fund data centers, and a gauge of Oracle’s credit risk hit its highest level since 2009. Credit default swap spreads on these bonds are already widening – by as many as 40 basis points relative to investment-grade bonds since September – potentially an early sign of investor discomfort.

The parallels to 2008 are uncomfortable. In that crisis, banks discovered they owned far more US housing risk than internal reports suggested. Financial institutions may soon discover the same about data-center and digital infrastructure risk, with exposures spread across corporate, real estate, infrastructure, fund financing, and alternative credit books.

Market Concentration Creates Systemic Fragility

The concentration of gains in AI-adjacent companies has created a market structure of extraordinary fragility. Roughly 75% of S&P 500 returns since the rally began stem from AI-related spending. Those same companies account for 80% of earnings growth and a staggering 90% of capital spending growth in the index. Nvidia alone, with a market capitalization exceeding $4.5 trillion, sits at the center of most major infrastructure deals.

Morgan Stanley’s chief investment officer has described the current market as a “one-note narrative” almost entirely dependent on massive AI capital expenditures, warning of a potential “Cisco moment” – referencing the company that was briefly the world’s most valuable before an 80% stock plunge during the dot-com bust. The concern extends to circular financing patterns, where Nvidia invests in OpenAI, which has deals with Oracle and AMD, creating interwoven financial dependencies that increase systemic risk.

Why This Bubble Behaves Differently

One critical distinction separates the AI bubble from its dot-com predecessor. During the 1990s telecom boom, companies massively overbuilt fiber-optic infrastructure in anticipation of internet demand. Much of it remained “dark” and unused for a decade. The AI bubble faces the opposite problem: excess demand overwhelming supply.

Every GPU depreciates at Moore’s Law speed. Every data center requires constant, massive power draws. Every technological advance makes yesterday’s infrastructure obsolete within quarters, not decades. Companies cannot pre-build AI capacity the way they laid fiber-optic cable. This creates a bubble dynamic that could stall not from lack of demand but from the physical inability to supply what the market needs – and from the crushing economics of perpetual infrastructure replacement.

An MIT-linked study found that roughly 95% of generative AI pilots fail to generate meaningful business impact. BCG research shows approximately 74% of companies struggle to achieve and scale value from AI initiatives. The enterprise adoption target of 50% by 2027 remains uncertain, with scenarios showing adoption potentially stalling at current 20% meaningful-deployment levels rather than accelerating.

What Experts Say Investors and Enterprises Should Do Now

The expert consensus converges on several key actions. Erik Gordon advises bracing for significant wealth destruction, warning investors to avoid hype-driven stocks and expect sudden $20 billion-scale market cap wipes. Jim Morrow of Callodine Capital urges rigorous ROI assessment, noting that the industry is in the “anteing up” phase where actual returns must materialize. Michael O’Rourke of Jonestrading recommends monitoring growth plateaus closely, as any deceleration in growth projections will trigger market exits, particularly for debt-laden firms.

Eric Clark of Rational Dynamic Brands Fund warns that with trillions crowded into a small group of themes and names, “when there’s the first hint of that theme even having short-term issues, they’re all leaving at once.”

The Road Ahead

The AI bubble’s deflation will not look like a single dramatic crash. It is more likely to manifest as a grinding erosion – slowing earnings growth, rising depreciation costs, tightening credit conditions, and a gradual recognition that adoption metrics without profitability metrics are meaningless. The 18% earnings growth projected for the Magnificent Seven in 2026 may not sound alarming in isolation, but for stocks priced for perfection, deceleration is the trigger.

The real danger lies in the economy’s dependence on AI investment for growth. If AI capital spending contracts, GDP growth evaporates almost immediately – there is no diversified base to absorb the shock. Geopolitical instability, semiconductor supply chain risks concentrated in Taiwan, and record public debt issuance could compound any downturn. Financial institutions, enterprises, and individual investors face a stark choice: continue riding a narrowing bet on AI’s transformative potential, or begin the difficult work of diversification before the market forces it upon them.

Kevin O’Leary has argued that AI shows measurable productivity unlike dot-com hype, and that returns can be measured “on a dollar-by-dollar basis.” He may ultimately be right about AI’s long-term value. But as every bubble in history has demonstrated, being right about the technology does not protect you from being wrong about the timing – or the price.

Sources