Maybe "is AI a bubble?" is the wrong question

It's no secret that "AI companies" have staggering valuations. At $5.5 trillion, Nvidia is now worth more than the GDP of every economy on earth except the US and China.

If these numbers give you even a tiny bit of an "I don't understand it, but something feels off" feeling, you're not alone.

One thing I want to get clear before going any further: I'm not disputing the usefulness of GenAI as a technology (I use it heavily myself and I can see how it's transforming the industry I'm working in). What I'm interested in understanding is how the companies behind this technology are valued. Are those valuations justified?

What can we learn about bubbles from history? In a nutshell: bubbles inflate and burst, and investors take giant losses even when the technology goes on to permeate the economy. Here are two examples:

• In the 1840s, the railway mania in the UK consumed ~7% of the country's GDP. Tracks were quintupled, but in the end builders saw only about a quarter of the returns they expected.
• The pattern repeated with fibre cables in the dotcom era in the US. Level 3, a US telecoms company, lost ~95% of its value laying down fibre that became the bedrock of the modern internet.

Today, the five biggest AI spenders are spending more than companies spent on oil in the shale boom or on telecoms in the dotcom boom. It certainly feels like the same bubble shape, doesn't it?

So is AI a bubble? In order to answer this for myself, I wanted to understand the two extremes of the opinion spectrum: bulls on one side, bears on the other.

The bull case, stripped down: the technology is real, the spend is rational, the payoff is coming. Sam Altman says both things in one breath: investors are overexcited about AI and AI is the most important thing to happen in a very long time. Jensen Huang points at real, booked demand for chips. Satya Nadella invokes Jevons (a famous English economist): make a thing cheaper and you get more use, not less. Jeff Bezos calls it a bubble in a good way: an "industrial" bubble that overbuilds but leaves society something productive, even if the early investors don't get their returns.

The bear case, stripped down: the technology is real, the prices and the financing are not, and the reckoning is near. Michael Burry is arguably the loudest bear. He says it isn't Enron, it's Cisco. Cisco was a real company selling real picks and shovels, but it also fell ~80% and never reclaimed its 2000 high. His actual claim is about demand quality. Nvidia's buyers are concentrated and their orders are distorted by a training-and-benchmarking phase that won't last. Paul Tudor Jones says it feels like 1999 and the eventual correction could be severe.

Both sides concede the usefulness of the technology. What they disagree about is whether the economic payoff arrives soon enough to justify today's valuations. The bulls and bears have different clocks.

Clock 1 — can unit costs fall faster than usage grows?

Total bill = price per token × tokens used. Total spend only falls if cost per token declines faster than token usage rises. Bulls are right that price per token is going down, and bears are right that tokens used is going up. The question is: which one wins?

Blended enterprise costs per million tokens fell 67% YoY, $18.40 → $6.07 (Q1'25 → Q1'26, from 2.4bn API calls). Goldman sees 60–70%/yr inference cost decline. Nvidia's Rubin platform targets ~10× inference cost cut versus its Blackwell architecture.

On the other hand, total enterprise AI bills tripled over the same period. Agentic workflows use 5–30× more tokens per task than a chatbot call (agents loop, call tools, self-check, etc.). Google now processes tens of billions of tokens a minute (~19bn by its own I/O 2026 figure), and AT&T went from 8B to 27B tokens/day after multi-agent deployment.

Today's valuations assume AI gets cheap enough to be everywhere and still turn a profit. If falling unit costs win, that holds. If ballooning usage wins, customers start rationing, and the demand those valuations are built on stalls.

Clock 2 — will the money hold out long enough?

AI companies have a runway. The question isn't whether they can fly, but whether they have enough runway to get liftoff.

Profits at the five biggest AI spenders (Amazon, Google, Meta, Microsoft, Oracle) keep climbing, but free cashflow after capex is now falling.

To keep building, they're borrowing heavily and committing to pay for data centres that don't exist yet. And the revenue meant to pay it all back isn't here yet. Bain estimates the industry will need around $2 trillion in annual revenue by 2030 to fund the build-out — and even on generous assumptions, it's on track to fall about $800bn short.

So who's right? Bulls say these are the richest companies in history. They can afford to lose money for years, funding the gap from cash and cheap loans until the AI revenue shows up. Bears say patience will run out. If revenue is slow, loans get pricey, or investors get nervous, the money can dry up before lift-off. You go bankrupt waiting, even if AI would eventually have paid off.

You can already see the nerves around the weakest of the five: after Oracle made a $300bn promise to OpenAI, its shares fell sharply (by one count ~57% from the peak, depending on the window) as investors worried it was betting too much on a single customer.

Clock 3 — is real external demand arriving fast enough to replace circular demand?

To know if AI is overvalued, you check the revenue: are real customers paying real money? If they are, the demand is real and the valuations might be fine.

But some of that "revenue" may be the same money going in a circle. Company A puts money into Company B, B spends it back with A, both report it as income. Nvidia invests in companies that then use the money to buy Nvidia chips. Microsoft "invests" in OpenAI partly as Azure credits. Amazon does the same with Anthropic.

For the valuations to hold, real external customers have to show up fast enough to take over from the circular stuff. The bull thinks the outside money is already arriving. The bear thinks the replacement rate is near zero.

So is it a bubble?

I still can't give a clean yes or no. But I've stopped thinking that's the right question. The better one: do the cash flows arrive before the financing patience runs out?

• Clock 1: can costs fall fast enough, soon enough?
• Clock 2: will the money last long enough?
• Clock 3: does real demand arrive fast enough to replace the circular kind?

Bull and bear mostly agree the technology is real. They disagree on the clocks.

Where do I land? I'm on the bearish side for now, but I'm watching these three clocks.

We may still get an AI revolution. Just not one that rewards today's investors.