AI Bubble vs Dotcom Bubble: What the Data Actually Says

Every few months somebody declares the AI trade a bubble. A few weeks later somebody else declares that this time is genuinely different. Both camps reach for the year 2000 as their reference point, and both usually stop at the surface: a chart of the Nasdaq, a scary valuation number, a confident conclusion.

The comparison deserves better than that. The dotcom bubble is one of the most thoroughly documented manias in financial history. We know what the capital spending looked like, what the earnings looked like, what the balance sheets looked like, and what happened to all three when the music stopped. That gives us something rare in markets: a control group.

So let us put the 2026 AI complex next to that record, line by line, and see where the two actually overlap. No verdict up front. The numbers first, the argument second.

Part One: The Anatomy of the Last One

Between January 1995 and March 2000, the Nasdaq Composite went from roughly 751 to 5,048.62. That is a rise of about 572% in five years, and the final year did most of the damage: the index gained 86% in 1999 alone.

The peak arrived on 10 March 2000. From there the index fell 78%, bottoming at 1,114.11 on 9 October 2002. More than five trillion dollars of market value was erased. The March 2000 high was not reclaimed until April 2015, fifteen years later.

Those are the headline numbers. The mechanics underneath them matter more.

The earnings were not there

At the peak, roughly 14% of the technology companies in the Nasdaq-100 were profitable. The Nasdaq-100 traded at a trailing price-to-earnings ratio near 200x and a forward multiple around 60x. Shiller’s cyclically adjusted P/E for the broad market hit 44, still the highest reading in the series.

Cisco Systems became the most valuable company in the world at roughly $555 billion of market capitalisation on net income of about $2.67 billion. That is well over 100x trailing earnings for a hardware business selling into a customer base that was about to stop buying. Cisco’s revenue, it should be said, was real. Its multiple was not.

The IPO machine ran on narrative

1999 produced roughly 308 technology IPOs with an average first-day gain of about 73%. VA Linux rose 697% on its first day of trading. Somewhere between three-quarters and four-fifths of those 1999 tech listings were unprofitable at the time they went public. The listing standard was a story and a domain name.

The infrastructure build was funded with other people’s money

This is the part most retellings skip, and it is the part that matters most for the AI comparison.

Telecom carriers raised roughly $1.6 trillion in equity and a further $600 billion in bonds. They invested about $500 billion in physical network between 1996 and 2001, laying 80.2 million miles of fibre optic cable across the United States. Capital expenditure peaked around $120 billion in 2000, which is something like $213 billion in today’s money.

Years later, up to 85% to 90% of that fibre was still dark. Unlit. Built for demand that arrived a decade late. When the carriers failed, bondholders recovered around twenty cents on the dollar. WorldCom and Global Crossing became case studies rather than companies.

The monetary backdrop turned

The Federal Reserve raised rates from 4.75% in mid-1999 to 6.5% by May 2000. The bubble did not die of valuation alone. It died of valuation plus a tightening cycle plus a funding window that slammed shut on companies with no cash flow of their own.

The aftermath

More than half of all dotcom companies were gone by 2004. Venture funding fell 95% from its peak. Amazon, one of the genuine survivors and eventual winners, fell 94% from high to low. The UK 3G spectrum auctions raised £22.5 billion from carriers who could not afford it, in what is still one of the cleanest examples of auction-driven capital destruction on record.

Hold that shape in mind. Enormous spending, funded by debt and equity issuance, by companies with no profits, against demand that had not yet appeared, at valuations that assumed it already had.

Part Two: The AI Boom by the Numbers

Now the present. All figures below are as of mid-2026.

The capital expenditure

The big four hyperscalers have guided to roughly $725 billion in combined capital spending for 2026. That is up about 77% from approximately $410 billion in 2025. The split runs roughly: Amazon around $200 billion, Alphabet $175 to $185 billion, Meta $125 to $145 billion, Microsoft $110 to $120 billion and above.

Goldman Sachs projects $5.3 trillion of hyperscaler capital expenditure across 2025 to 2030. Total technology investment now sits near 4.4% of US GDP, which is close to the share it reached at the dotcom peak.

By raw scale, this build is larger than the telecom build. That is not a defence of it. It is the reason the question matters.

The revenue

Here the picture diverges sharply from 1999. Microsoft reports an AI revenue run rate above $37 billion, growing 123% year over year, with commercial remaining performance obligations of $627 billion. Google Cloud carries a backlog above $460 billion. AWS is growing 28%.

Nvidia’s fiscal 2026 revenue came in at $215.9 billion, up 65%. The most recent quarter delivered $81.6 billion, up 85% year over year, with data centre revenue of $75.25 billion. The company trades near 31x trailing earnings and roughly 23x to 26x forward earnings at a market capitalisation around $5.3 trillion.

Compare that to Cisco at the 2000 peak: well over 100x trailing earnings. Nvidia is expensive in absolute terms and cheap relative to the bellwether it is most often compared to. Roughly half its data centre revenue comes from the hyperscalers, which is a concentration issue we will return to.

The concentration

The top ten companies now make up 41% of the S&P 500, against 27% at the 2000 peak. The Magnificent Seven alone account for about 32.5% of the index and trade near 28x forward earnings versus roughly 23x for the S&P overall. They drove more than half of index returns across 2023 and 2024.

Read that again. Index concentration today is materially higher than it was in March 2000. Whatever else is different, this is worse.

The valuation

Shiller’s CAPE sits near 41, the second highest reading in its history behind only 2000. The equity risk premium is the thinnest it has been since 2000, with the ten-year Treasury yielding roughly 4.3% to 4.7%. The technology sector trades at a 1.34x premium to the broad market, against 2.0x at the dotcom peak.

So: broad market valuation close to 2000, sector premium roughly a third of 2000, concentration worse than 2000, earnings support enormously better than 2000. That mixture is why smart people land on opposite conclusions using the same data set.

Part Three: Where the Two Eras Genuinely Rhyme

1. Index concentration

Already covered, and it is the strongest single parallel. When 41% of an index sits in ten names correlated to one theme, diversification becomes an illusion at exactly the moment you need it. The 2026 data shows early signs of that correlation loosening, with pairwise correlation among the Magnificent Seven down to 0.27 and dispersion above thirty percentage points, but the index-level concentration risk remains.

2. Capital spending as a share of the economy

Technology investment near 4.4% of GDP is dotcom-peak territory. Capital spending at that scale creates its own gravity: it flatters GDP on the way up and subtracts from it on the way down, whether or not the assets end up productive.

3. The new-era narrative

“The internet changes everything” and “AI changes everything” are the same sentence. Both may even be true. The internet did change everything, and it still wiped out 78% of the Nasdaq on the way to doing so. A correct thesis and a survivable position are separate problems.

4. Circular capital flows

Nvidia has committed up to $100 billion into OpenAI. Microsoft holds roughly a 27% stake in OpenAI. OpenAI in turn has compute commitments exceeding $1 trillion: approximately $250 billion with Azure, $300 billion with Oracle, $138 billion with AWS, $22.4 billion with CoreWeave, $10 billion with Broadcom. Bloomberg has framed this structure as an echo of Lucent and Nortel.

The comparison is fair as a structural observation. The suppliers are funding the customers who are buying the suppliers’ product. Where 2026 differs is that the money is coming out of genuine operating cash flow rather than fresh debt issuance, and the end demand is currently real. That reduces the fragility. It does not eliminate the circularity.

5. The bellwether trade

Cisco was the pick-and-shovel proxy for the internet. Nvidia is the pick-and-shovel proxy for AI. Both became the largest company in the world. Both saw a single vendor absorb a vast share of an industry’s capital budget. The difference is the multiple, and the multiple is a very large difference.

6. The speculative fringe

Every boom grows one. In 1999 it was the pets.com layer. In 2026 it is the wrapper-startup layer: thin applications built on somebody else’s model, priced on revenue multiples that assume no competition. That layer will not survive, and its failure will be loud, and it will be used as evidence about the entire complex regardless of whether it is.

7. The monetary backdrop

1999 into 2000 saw 175 basis points of tightening land on companies with no cash flow. Today’s equivalent risk is not the policy rate itself but the long end. With the equity risk premium at its thinnest since 2000, a meaningful move up in the ten-year yield does arithmetic damage to long-duration equity, and almost everything in the AI complex is long-duration equity.

8. Deep-pocketed, price-insensitive buyers

The Saudi Public Investment Fund and the UAE’s MGX $100 billion vehicle are deploying capital into AI infrastructure on strategic rather than return-based criteria. Sovereign buyers with non-financial objectives inflate asset prices in ways that do not respond to valuation signals, which is a familiar late-cycle feature even though the specific actors are new.

Part Four: Where They Break Apart

Mark Minervini laid out the bull case for structural difference in a widely circulated post on X. His argument runs across five points, paraphrased here with the underlying data attached.

One: the spending is larger and still accelerating. This is factually correct. The 2026 hyperscaler capital budget of roughly $725 billion dwarfs the telecom peak in nominal and real terms, and it is growing rather than rolling over. Note that this point cuts both ways. Larger spending means larger consequences if the demand assumption is wrong.

Two: AI is already monetised. Hyperscaler cloud sales have exceeded $375 billion over the trailing four quarters. This is the strongest single difference from 1999. The 2000 infrastructure was built for revenue that did not exist. The 2026 infrastructure is being built by companies already collecting the revenue.

Three: earnings support the investment. Hyperscaler valuations have actually compressed relative to earnings even as the stocks have risen, because earnings have risen faster. Minervini flags semiconductors as the more vulnerable link in the chain, which is a distinction worth keeping.

Four: balance sheets are far stronger. The telecom build was funded with $600 billion of bonds against no operating cash flow. The AI build is largely internally funded from some of the strongest cash-generating businesses ever assembled. There is no equivalent of the 2001 telecom debt wall currently visible.

Five: demand is strong and capacity is tight. Data centre vacancy rates are exceptionally low. This is the direct opposite of dark fibre. The 2000 build produced capacity that sat unused for a decade. The 2026 build is currently constrained by power and land rather than by lack of takers.

Three further differences worth adding

IPO discipline. 2025 produced roughly fifty technology IPOs against 308 in 1999. Median software listings arrive with more than $200 million in annual recurring revenue. First-day pops run 18% to 31% against 73% in 1999. The public listing market is behaving with restraint that has no 1999 analogue.

Nvidia’s multiple. A bellwether at 31x trailing earnings is a fundamentally different animal from a bellwether at over 100x. Even a severe earnings disappointment leaves Nvidia in a valuation range that Cisco could only have reached after an 80% decline.

National security framing. Export controls, sovereign fund participation and explicit government interest in domestic AI capability create a demand floor that had no 1990s equivalent. Strategic buyers are less price sensitive on the way up, and also less likely to disappear entirely on the way down.

Part Five: The Risks That Fit Neither Template

This is the section that matters most, and it is the one both camps tend to skip. Some of the largest risks in the AI complex are not dotcom echoes at all. They are new.

The depreciation question

Meta extended assumed server life from three years to five and a half. Microsoft and Alphabet moved to six. Amazon shortened a subset back to five. These are accounting choices with very large earnings consequences, because a longer assumed life spreads the same capital cost across more years and raises reported profit today.

Michael Burry has argued that GPUs have a real useful life closer to two or three years, that depreciation across the sector is understated by roughly $176 billion over 2026 to 2028, and that by 2028 this could overstate Oracle’s earnings by around 27% and Meta’s by around 21%. These are his estimates, not settled fact, and he holds short positions in Nvidia among others, with a reported basis near $198.09. Weigh the argument on its merits and account for the position.

The underlying question is legitimate regardless of who is asking it. If the assets wear out faster than the books assume, reported earnings across the complex are flattered, and the earnings support that anchors the entire bull case is softer than it looks.

The scale of the circularity

Covered above, but worth restating as a risk rather than a rhyme. OpenAI’s compute commitments exceed $1 trillion against a revenue run rate near $25 billion. Reported projections put 2028 compute spending at $121 billion against a roughly $74 billion loss. The company filed a confidential S-1 on 8 June 2026 and raised $122 billion at an $852 billion valuation.

Anthropic reached a $30 billion annualised run rate by April 2026, up from roughly $1 billion fifteen months earlier, with about 85% of it enterprise. That is one of the fastest revenue ramps in the history of software. It is also a fraction of the infrastructure commitments being underwritten against expectations of continued growth at that pace.

Private-market opacity

In 1999 the excess was on public exchanges where anyone could see it. In 2026 a large share of the risk sits in private companies, private credit and structured financing. CoreWeave’s $8.5 billion delayed draw term loan was the first investment-grade financing collateralised by GPUs. That is a genuinely new instrument, and new instruments are not stress-tested until they are.

The revenue-to-spending gap

One estimate puts enterprise AI revenue near $100 billion against roughly $400 billion of infrastructure spending. Goldman’s own analysts disagree publicly on the productivity payoff: Joseph Briggs models around a 9% productivity gain, while Daron Acemoglu models closer to 0.5%. Jim Covello has been openly sceptical that a killer application justifying the spend has appeared.

That gap can close two ways. Revenue rises to meet spending, or spending falls to meet revenue. Only one of those is pleasant to own.

Power as the binding constraint

Data centre electricity consumption ran near 460 terawatt-hours in 2024 and could reach 1,300 by 2035. Microsoft signed a twenty-year 835MW power purchase agreement with Constellation covering Three Mile Island. Amazon contracted 1.9GW from Talen’s Susquehanna facility. Google has committed to small modular reactors with Kairos.

Fibre could be laid faster than demand appeared. Power cannot. This constrains the build in a way that limits overcapacity risk, and simultaneously introduces execution risk that the 2000 build never faced.

Sentiment fragility

Two data points show how thin the cushion is. DeepSeek’s January 2025 model release took $589 billion off Nvidia’s market capitalisation in a single day, the largest single-day loss in market history at the time. Meta fell 9.25% the day after raising capital expenditure guidance, which tells you the market has begun pricing capex as a cost rather than an investment. Citi has noted Amazon’s free cash flow is likely negative.

When capital spending starts being punished rather than rewarded, the reflexive loop that funds the boom begins running in reverse. Watch that signal more closely than any valuation metric.

Part Six: The Installation Phase Lens

Carlota Perez’s framework for technological revolutions splits each one into an Installation phase and a Deployment phase, separated by a crash.

Installation is when financial capital builds the infrastructure, speculatively and wastefully, at valuations disconnected from near-term returns. The crash clears the excess. Deployment is when the infrastructure gets used productively by the wider economy, and where most of the actual wealth is created.

Railways, electricity, mass production, the microprocessor and the internet all followed this shape. On this framework AI sits in the Installation phase of a possible sixth technological revolution.

The uncomfortable implication is that Installation phases historically end in a crash even when the technology is completely real. The fibre laid in 1999 powered the streaming economy a decade later. It also bankrupted the companies that laid it. Being right about the technology and right about the equity are separate bets, made at different times, with different odds.

Figure 3 · Side by Side: 2000 versus 2026

Bubble Scorecard: Eleven Tests

One row per structural test. The verdict column asks a single question: on this specific measure, is 2026 safer than 2000, roughly the same, or worse?

Five tests come out clearly safer, four look broadly similar or introduce a risk that did not exist in 2000, and two are outright worse. That distribution is the honest answer to the headline question, and it is why a single-word verdict fails.

Part Seven: The Trader’s Lens

Here is where most bubble analysis fails. It resolves into a prediction, and a prediction is not a plan. You cannot trade “it is a bubble” any more than you can trade “it is not.” Both are opinions about a future that will arrive on its own schedule.

Run it through Mind, Method, Money instead.

Mind

Notice which conclusion you want to be true, because that is the one you will find evidence for. If you are long the complex, you will find Minervini’s five points compelling and Burry’s depreciation argument motivated. If you are short or sitting in cash watching others make money, you will find the reverse. The data above genuinely supports both readings. That is the whole point of laying it out.

The 2000 lesson on this front is brutal and simple: the people who were right about the internet and the people who kept their capital were largely different groups.

Method

Bubbles do not end on valuation. They end on price action. The Nasdaq’s trailing multiple crossed 100x well before the top, and anyone who shorted on that signal was carried out long before March 2000 arrived.

What actually marked the turn was distribution: failed breakouts, leadership narrowing, heavy volume on down days, and the bellwether names losing their trend while the index still looked fine. Watch the semiconductor complex specifically, which is the segment Minervini himself flags as most vulnerable, and watch how the market reacts to capital expenditure announcements. Meta falling 9.25% on a capex raise is exactly the kind of behavioural shift that precedes a change in regime.

Money

This is the part that survives regardless of which camp turns out to be right. Position sizing that assumes you might be wrong. Concentration limits that account for the fact that owning an S&P index fund in 2026 is a 41% bet on ten correlated names. Stops placed on evidence rather than hope. Enough dry powder that a 40% drawdown in the complex is an opportunity rather than an obituary.

Amazon fell 94% and then became one of the great compounders of the modern era. Both facts are true. Only the traders who survived the first got to participate in the second.

The Verdict, Such As It Is

The AI boom is not the dotcom bubble repeating. The earnings are real, the balance sheets are strong, the capacity is being used, and the bellwether trades at a third of the multiple. On the specific mechanism that killed 2000, which was enormous debt-funded spending against revenue that did not exist, the 2026 complex is materially safer.

It is also not obviously safe. Concentration is worse than 2000. Valuation on the broad market is close to 2000. The circular financing structures are real, the depreciation assumptions are contested, a meaningful share of the risk has moved into private markets where it cannot be observed, and the gap between infrastructure spending and enterprise revenue is currently wide.

The honest summary is this. The most likely failure mode is not a 78% index collapse. It is a violent repricing concentrated in the most speculative layers, semiconductors and the wrapper-startup tier, while the profitable hyperscalers absorb the damage and continue. That would look less like 2000 and more like 2022, which is precisely where Minervini lands.

But the Perez framework offers a caution that no earnings quality argument fully answers. Installation phases have historically ended in a crash even when the technology was entirely real. Being right about AI and being right about AI equities are separate propositions, and history suggests they resolve at different times.

Plan for both. Size for the one you have not priced.