AI Boom vs. Dot-Com: A Larger, Darker Scale

The dot-com analogy has become a comfort. Every market correction since 2000 has been measured against the original, and the artificial intelligence boom, with its $725 billion in annual hyperscaler capex, its Shiller cyclically adjusted price-to-earnings ratio above 40, and its $4.3 trillion semiconductor company, invites the comparison more readily than most. The IMF has reached for it. Howard Marks has reached for it. Richard Bernstein has produced calculations showing that AI investment as a share of the U.S. economy is roughly a third larger than internet investment was at the 2000 peak. The frame has become so familiar that it has stopped doing useful work. The interesting question is not whether AI is a bubble. It is what kind of bubble it is, and whether anyone is positioned correctly for the specific shape it has taken.

The shape is this. In the dot-com cycle, capital died first and labor died second. Companies failed, and the people who worked at them lost their jobs as a consequence. The graveyards of Pets.com and Webvan and dozens of less-remembered casualties were full before the unemployment statistics caught up. The 2026 sequence has reversed. The companies most aggressively shedding workers are not the failing ones. They are Amazon, Microsoft, Meta, Salesforce, JPMorgan Chase, Ford. Firms whose balance sheets have rarely been stronger. The AI-driven layoffs are arriving during the boom, justified by productivity gains that the financial statements have not yet confirmed. Whatever else this episode turns out to be, it is not a repeat of the dot-com era. It is something less rehearsed and, in at least one specific way, considerably more difficult to think about.

The Spending

What the hyperscalers told their shareholders this spring is, even by the standards of an industry given to large numbers, remarkable. Microsoft set 2026 capital expenditure at $190 billion, with CFO Amy Hood attributing $25 billion of that figure to rising memory and component costs alone. Amazon will spend approximately $200 billion. Alphabet guided $185 billion. Meta, which has no public cloud business through which to monetize the spend, raised its 2026 range to $115 to $145 billion. Combined: roughly $725 billion, a 77 percent jump on the prior year. The equivalent annual output of Switzerland, deployed in twelve months, by four companies, on infrastructure for a technology whose enterprise return remains, by the most generous reading, unsettled.

The figure dwarfs anything assembled in the previous cycle. Total U.S. internet and telecommunications capex in 1999 and 2000 combined, adjusted for inflation, comes to roughly two-thirds of what these four firms will spend this year alone. The AI cycle has not just rhymed with the dot-com era. It has been transposed several octaves higher.

The Resemblances That Hold

Three things about the present moment rhyme with 2000 honestly enough to deserve attention, and a fourth that is sometimes invoked but is, on closer inspection, weaker than it appears.

Market concentration. AI-related companies drove approximately 80 percent of the S&P 500’s gains in 2025. The Shiller cyclically adjusted price-to-earnings ratio crossed 40 in late 2025, a threshold previously reached only on the immediate eve of the dot-com unwind. The reading is uncomfortable not because such valuations cannot persist, they have persisted longer than skeptics expected in late-cycle environments, but because the historical record of what follows them is monotonous in one direction.

The financing architecture. Nvidia invests in OpenAI; OpenAI commits to purchasing Nvidia GPUs; Microsoft funds OpenAI; OpenAI runs on Azure. The OpenAI-Nvidia commitment alone is estimated to account for as much as 13 percent of Nvidia’s projected $272 billion in 2026 revenue. This is precisely the structure by which Lucent and Nortel financed telecommunications carriers in 1999, equipment makers lending customers the money to buy their equipment, and it ended badly, in waves of bankruptcy from the carriers and revenue collapse at the suppliers. UBS calls the current AI arrangement “not overwhelming.” Janus Henderson calls it a “virtuous circle.” The skeptical name is round-tripping. The dispute is about the size of the dependency, not about whether one exists.

Retail enthusiasm at unprecedented velocity. Morgan Stanley estimates that U.S. retail investors have moved more than $700 billion into equities since January 2026, roughly five times the pace of the dot-com run-up. Nasdaq-linked exchange-traded funds have continued attracting over $120 billion in inflows even as S&P 500 earnings revisions have turned negative. The late-cycle pattern is reliable: the conviction of the marginal buyer increases as the underlying fundamentals soften.

Where the analogy weakens. Each of the resemblances above is real. None of them, however, addresses the question that mattered most in 2000, which was solvency. The dot-com bust was a credit event. Companies that could not service their obligations went bankrupt, equipment makers wrote down receivables, and the damage spread by default. The 2026 AI spenders are different animals. Nvidia recorded $215.9 billion in revenue at a 53 percent operating margin in its most recent year. Microsoft’s AI revenue surpassed a $37 billion annual run rate. Google Cloud’s backlog has reached $460 billion. The four hyperscalers can fund their current capex out of operating cash flow, for now. The risk in front of them is not insolvency. It is something subtler and slower: overbuilt capacity at firms too profitable to fail, depreciating quietly, while the demand thesis takes longer than the cycle assumed to land.

The Inversion

The single most important difference between the two cycles, and the one most underweighted in current commentary, concerns labor.

The 2001 calendar year produced 168,395 tech-sector layoffs, the worst figure Challenger, Gray & Christmas had recorded at that point in its history. Across the full 2000 to 2002 unwind, conservative estimates put cumulative tech-sector job losses above one million. In the San Francisco Bay Area, four out of every five dot-com companies went out of business in 2000 and 2001 alone, eliminating roughly 30,000 direct internet positions in that region. The displaced workers came from companies that had failed. The cause and effect was unambiguous: the company was gone, and therefore the job was gone.

The 2026 AI layoff profile inverts that sequence. Through the first five months of this year, U.S. employers have announced 397,755 job cuts, according to Challenger, Gray & Christmas. AI has been cited as the principal reason in 87,714 of them, already exceeding the 54,836 AI-attributed cuts recorded in all of 2025, and now the leading reason given for layoffs three months running. The technology sector has led all industries, with 123,653 cuts year-to-date, up 66 percent from the same period in 2025 and the steepest concentration since early 2023. Amazon eliminated 14,000 corporate roles in October 2025, framing the decision as an AI-enabled flattening of management layers. Workday cut 8.5 percent of its global workforce, roughly 1,750 people, explicitly to reallocate spending toward AI investment. Meta described its May 2026 reductions as a means of offsetting the cost of its AI infrastructure commitments. These are not the cuts of bankruptcy. They are the cuts of strategic choice, made by firms whose financial position is uncontested and whose stock prices are at or near records.

There is a complicating factor worth naming, which is that some of these layoffs are not, in any rigorous sense, caused by AI. Resume.org found that nearly 60 percent of hiring managers planning cuts in 2026 cited AI as the principal reason, while only 9 percent reported that AI had fully replaced any specific role. The gap between citation and reality has acquired its own label: AI washing. Hiring managers reportedly prefer the AI framing because investors receive it more favorably than the candid alternatives, post-pandemic over-hiring, deferred cost discipline, the conclusion of an unsustainable headcount cycle. The displaced worker, however, does not particularly care which framing is technically correct. The rent is still due.

A widely cited MIT NANDA study, surveying 300 enterprise AI initiatives along with 500 executives and employees, found that the substantial majority of deployments had yet to produce a material profit-and-loss return. The headline figure derived from the study, that 95 percent of enterprises see zero measurable ROI, has been contested for its methodology and for the narrowness of what it counts as a return. The more defensible reading is the more useful one: at this stage of the AI cycle, the productivity dividend is conspicuously absent from the financial statements of the firms that have spent most aggressively on the technology. OpenAI is on track to lose approximately $14 billion in 2026, nearly triple its 2025 losses, even as it projects $100 billion in revenue by 2029. The projection is not implausible. It is also not yet evidence.

Combine the three observations, the front-loaded labor adjustment, the unresolved productivity case, and the absence of the cushion that a profitable underlying enterprise would normally provide its workforce, and the picture is one no part of the 2000 playbook prepared the present generation of analysts for. A correction in capital, if it arrives, will land on a labor market that has already been hollowed out by the AI boom itself. The dot-com cycle gave back, in compensation, the boom-time wages it had paid out. The current cycle is collecting on the labor adjustment in advance.

Where the Comparison Misleads, in Detail

The cleanest illustration of the structural difference between the two cycles is the side-by-side of the principal actors at the top of each market.

Four Things Worth Watching

A consultant’s job, properly understood, is not to predict the timing of a correction but to identify which signals would change the recommendation. Four are worth disciplined attention over the next eighteen months.

Free cash flow inflection. Amazon is projected to turn cash-flow negative in 2026 for the first time in years. When AI capex begins to exceed operating cash flow at the other hyperscalers, the bet shifts from self-funded to debt-funded, and the character of the risk changes materially. Self-funded overbuilding is a depreciation problem. Debt-funded overbuilding is a balance-sheet problem.

Disclosed AI-specific margins. The first quarter in which a hyperscaler discloses AI-specific operating margins separately from broader cloud revenue will be material. Markets have so far accepted backlog growth as a proxy for AI returns. The moment that proxy is replaced by direct measurement, the analytical landscape will move sharply in whichever direction the disclosure permits.

Any visible unwinding of a circular financing deal. A single instance of an AI lab failing to meet a chip-purchase commitment, or a chip maker writing down an equity stake in a customer, would alter the perceived risk geometry of the entire ecosystem within a trading session.

The power grid constraint. Memory chips, GPUs, and skilled engineering labor are the visible bottlenecks of the current AI cycle. Electricity is the quieter one. When a major data-center project is delayed or canceled for grid-interconnection reasons rather than capital reasons, and the first such instances are already appearing in Virginia, Ireland, and parts of Texas, the question shifts from “is demand uncertain” to “is supply impossible.” That constrains the upside of the bull case as much as it complicates the bear case.