Is AI a Bubble? AI Boom vs Dot-Com Bubble, Measured in 2026

The Angle. This is a boom with a bubble inside it. The demand is real and the leader is priced sanely, but the spending is running years ahead of the revenue, the financing is getting more creative, and the stock market is more concentrated than it was in 2000. Expect a shakeout in the frothiest corners, not the end of AI.

The Trick

In August 2025, the chief executive of the most famous AI company in the world was asked whether investors were overexcited about AI. Sam Altman‘s answer was yes, and he warned that some would get very burnt. It is not every day the magician tells the audience the trick might fail.

Fourteen months later, the spending has only grown. Microsoft, Alphabet, Amazon and Meta spent about $416 billion on capital projects in 2025 and their latest guidance adds up to roughly $730 billion for 2026. Alphabet, Meta and Amazon have gone from throwing off cash to reporting negative free cash flow, according to analyses of their latest results, and are borrowing to fund the build.

The trick everyone falls for is the single word “bubble”. It makes you pick a side: bubble or not. The reveal is that bubbles are not one thing. They have a price part (are assets overvalued?), a money part (is the spending funded by real revenue or by hope and debt?) and an ending (do failures cascade?). On some of those gauges, today looks like 2000. On others, it looks nothing like it. So I measured eight of them.

Mojo's Law: the payback equation

Every data centre is a promise that future revenue will pay for today's concrete and chips. You can check the size of that promise with one line of arithmetic.

R_{\text{needed}} \approx \frac{C \times \left(\frac{1}{L} + r\right)}{m}

C is the capital spent, L is how many years the assets stay useful, r is the return investors expect, and m is the operating margin on the AI revenue R that the assets produce.

Now plug in round, clearly labelled assumptions. An FT tally puts the Big Four's cumulative capex at about $1.1 trillion from 2023 to mid-2026. Say the blend of chips and buildings lasts five years on average, investors want 10% a year, and AI services earn a 50% operating margin.

Then R is about $1.1 trillion times 0.3, divided by 0.5: roughly $660 billion of AI revenue a year, every year, just to pay back what has already been built. Not all of that capex is AI, which lowers the bar; GPUs may age faster than five years, which raises it.

This is the equation behind every bubble and every successful build-out in history. Railways, electricity grids and fibre networks all failed it for their first investors and passed it for society. The economist Carlota Perez calls this the installation phase: a frenzy of capital that builds the infrastructure, a crash, and then a long golden age when everyone uses what was built.

How much is actually being spent?

In nominal dollars, more than the whole dot-com telecom build-out, and it is still accelerating. The Big Four's capital expenditure rose from about $155 billion in 2023 to about $413 billion in 2025, and their latest guidance points to roughly $730 billion in 2026.

For 2026, Amazon guides to about $220 billion, Alphabet to $195 to $205 billion, Microsoft to about $170 billion for the calendar year and Meta to $135 to $145 billion, according to reports of their third-quarter results. Oracle, a late entrant, expects about $95 billion in its 2027 fiscal year, including customer-funded prepayments. McKinsey's base case, from April 2025, is $6.7 trillion of global data-centre investment by 2030, $5.2 trillion of it for AI.

The deals around the spending are where it gets dot-com familiar. OpenAI alone signed compute commitments worth roughly $1.15 trillion over 2025 to 2035 across Oracle, Microsoft, Broadcom, Nvidia, AMD, AWS and CoreWeave, by one analyst's tally.

Several are circular: suppliers investing in customers who then buy the suppliers' chips. Nvidia's investment commitments reached about two-thirds of its trailing revenue in late 2025, against roughly a quarter at Lucent in 1999 to 2000, according to an analysis by the investor Tomasz Tunguz.

But here the 2026 story is one of retreat, which is a healthy sign. Nvidia's headline $100 billion letter of intent with OpenAI was never finalised and became a roughly $30 billion equity stake. And announcements are not buildings: Epoch AI found in April 2026 that OpenAI's Stargate project had about 0.3 gigawatts operating, all in Abilene, Texas, against more than 9 gigawatts planned. Big numbers are easy to announce. Power stations are not.

Is the money coming back?

Faster than the sceptics of 2024 expected, and still far slower than the spending. That is the whole debate in one sentence.

The sellers of shovels are rich. Nvidia's data-centre revenue went from $47.5 billion in fiscal 2024 to $193.7 billion in fiscal 2026, and its second quarter of fiscal 2027 brought in $89.0 billion from data centres alone, up 117% on a year earlier. Cloud backlogs have swollen: Microsoft reports a commercial order book of about $677 billion, Oracle about $664 billion and Google Cloud about $561 billion, according to reports of their latest results.

The model makers are growing, and losing money. OpenAI's revenue rose from $3.7 billion in 2024 to about $13.1 billion in 2025, with an operating loss of about $20.9 billion, according to financial statements reported by the Financial Times. Anthropic's annualised revenue run-rate reportedly passed $65 billion by the end of July 2026, up from about $9 billion at the end of 2025. ChatGPT is close to a billion weekly users.

Businesses are adopting, slowly. The US Census Bureau's survey found 19.8% of US firms using AI in May 2026. McKinsey's 2025 survey found 88% of surveyed organisations using AI in at least one function, but only 39% reporting any effect on profit.

A February 2026 NBER survey of about 9,000 executives estimated AI added about 0.2 percentage points a year to their firms' productivity growth. And the much-quoted MIT finding that 95% of corporate generative-AI pilots fail referred to task-specific pilots reaching production with measurable profit impact within six months, from a small sample the authors called directional.

In June 2024, David Cahn of Sequoia Capital posed AI's “$600 billion question“: take Nvidia's data-centre run-rate, double it for the full cost of a data centre, double again for the margin the end user needs, and you get the annual AI revenue required to justify the spend. I re-ran his sum with 2026 numbers.

On Nvidia's full fiscal 2026, the answer is about $775 billion a year. On its latest quarter annualised, it is about $1.4 trillion. Against that, OpenAI's own 2026 revenue forecast of about $30 billion plus Anthropic's $65 billion run-rate come to roughly $95 billion. Cloud providers and software firms earn AI revenue too, so this is not the whole market, but the gap is measured in hundreds of billions. The $600 billion question did not get answered. It got bigger.

Does today actually look like 2000?

In scale and concentration, yes and then some. In the price of the leader and in failures, not even close. The Mojo Bubble Gauge puts eight indicators side by side, each expressed as a multiple of its dot-com level.

Five red. The ten biggest companies make up about 40% of the S&P 500's value, against roughly 26% in 2000. Nvidia, near $5 trillion in late 2025, was worth about 16% of US GDP; Cisco at its March 2000 peak was about 5.5%.

Cumulative Big Four capex of about $1.1 trillion since 2023 is more than double the $500 billion-plus that telecom companies poured into networks in 1996 to 2001 (in nominal dollars). Supplier financing is higher than Lucent's ever was. And Morgan Stanley's projection that data-centre debt could pass $1 trillion by 2028 would match the debt the telecom industry owed after its bust.

One amber. The Shiller cyclically adjusted price-to-earnings ratio for the whole US market was about 41.6 in October 2026, close to its December 1999 peak of 44.2 and more than double its long-run average of about 17.

Two green, and they matter most. Cisco traded at more than 150 times expected earnings at its peak. Nvidia trades at about 23 times, close to the market average of about 21 times, because its profits have grown nearly as fast as its share price. And the dot-com bust was a cascade of failures: dozens of upstart telecom carriers went bankrupt, followed by Global Crossing and WorldCom. As of November 2026, I found no default by an AI lab or a hyperscaler.

The first tremors have arrived all the same. South Korea's KOSPI, heavy with AI chip makers, fell about 40% in roughly five weeks in mid-2026 as a leveraged-fund unwind met fading AI enthusiasm, though it remained up for the year. OpenAI shut down its Sora video app in April, and its expected stock market listing slipped to 2027. Alphabet's shares fell up to 11% when it raised its capex plans again in late October.

What would a burst look like, and what survives?

If it comes, it will start where the debt is, not where the hype is.

Follow the borrowing. The 2026 shift that worries me most is not a valuation; it is cash flow. Analyses of third-quarter results show Alphabet, Meta and Amazon spending more on capex than they generate, and Alphabet's long-term debt reportedly rose from about $46.5 billion in December 2025 to about $125 billion by September 2026.

Oracle raised about $43 billion of debt in its 2026 fiscal year. Meta financed its Hyperion campus through a joint venture 80% owned by funds managed by Blue Owl, keeping roughly $27 billion of construction off its own balance sheet. Specialist GPU clouds such as CoreWeave borrowed against chips and customer contracts. These are not frauds. They are leverage, and leverage is what turns a correction into a crash.

Watch the deflation trap. The cost of a given level of AI capability is falling astonishingly fast: about tenfold a year by Andreessen Horowitz's estimate, and between 9 and 900 times a year depending on the task, by Epoch AI's. That is wonderful for users and dangerous for owners of expensive, fast-ageing chips. Bandwidth prices collapsed in the early 2000s in much the same way, and the owners of the fibre went bust while their customers thrived.

Remember what the last bubble left behind. In 2002, a Brookings analysis found no more than 2% of North American long-distance telecom capacity in use. Investors lost trillions. But that glut of cheap capacity is what made YouTube, cloud computing and the smartphone internet affordable a few years later. Internet users went from about 390 million around 2000 to about 6 billion by 2025. A burst would hurt the people who paid for the build. It would not un-build it.

AI in everything is the real hedge. The demand side is broadening. Gartner expected AI PCs to make up 31% of PC shipments in 2025 and more than half in 2026.

Humanoid robots are moving from demos to small fleets: Figure reported more than 350 of its latest robots delivered by April 2026, and Unitree's shares jumped about 500% on their first day of trading in Shanghai in August 2026.

Morgan Stanley projects a $5 trillion humanoid market by 2050. Those are forecasts and early numbers, not proof, but they show AI spreading from chatbots into washing machines, cars, phones and factories. That breadth is what separates a platform shift from a fad.

Claim Check

The Counter-Angle

The bull case deserves a fair hearing, and it is strong.

• First, the buyers are not dot-com start-ups; they are among the most profitable companies in the world, funding much of the build from their own operating cash, and in 2023 the ten largest US companies earned about 69% of the market's economic profit while holding 27% of its value, according to Morgan Stanley's Michael Mauboussin.
• Second, demand is visible in signed contracts, not page views: cloud order books in the hundreds of billions each.
• Third, the leader is priced on profits. Nvidia at about 23 times forward earnings is not Cisco at 150.
• Fourth, falling prices are expanding use faster than they shrink revenue, which is Jevons' paradox working in the industry's favour.

My answer:

Mojo's Forecasts

1. The Big Four's combined 2027 capex guidance, as stated by 28 February 2027, exceeds $900 billion.

2. Nvidia's forward price-to-earnings ratio stays below 40 on 31 December 2027. No Cisco-style blow-off.

3. The Nasdaq-100 suffers a peak-to-trough fall of 30% or more at some point before 31 December 2027.

All three go into the Mojo Forecast Ledger, and the Bubble Gauge will be refreshed after each earnings season.

Where this is debated next

FAQ

Every bubble builds something. The only question is who is holding the bill when the music stops.