“Is AI a bubble?” is the finance equivalent of “Is this the year I finally get my life together?”

Why AI looks like a mania from the outside, a productivity engine from the inside, and a very normal general purpose technology in the long run.

Technically a fair question, functionally useless.

What people mean is:

“Is this tulips? Can I safely roll my eyes, buy an index fund, and never think about CUDA again?”

To answer that, you have to separate two things we keep mashing together:

1. AI as a technology

2. AI as an asset class / stock market bit of theatre

They are related. They are not the same thing.

Step one: stop letting “bubble” do lazy work

When civilians say “bubble”, they usually mean fraud or toy:

Beanie Babies.

ICOs named after dogs (initial coin offerings)

NFTs of pixelated men in hats, briefly worth more than actual houses.

In that sense, a “bubble” is something that leaves no real infrastructure behind. The chart goes vertical, collapses, and the physical legacy is… hoodies.

That’s not how most serious technologies behave.

Economists have a deeply unsexy phrase for the ones that actually matter:
general-purpose technologies. A general-purpose technology (GPT) is a technology that is:

• used everywhere,

• keeps getting better, and

• gets more valuable the more you rearrange the world around it.

Steam engines did that. Electricity did that. The internal combustion engine, semiconductors, the commercial internet. They all start as “cool gadgets” and end up re-architecting cities, firms, labour markets and whatever we call culture now.

And here’s the important bit:

Every canonical GPT comes wrapped in at least one ridiculous investment mania.

• Railways in the 1840s: thousands of miles of track authorised, investors torched, trains become the backbone of industrial life anyway.

• Early electrification: utility booms, municipal overreach, bankruptcies… followed by “of course the lights turn on when I flip this”.

• Dotcom: Pets.com dies; the fibre and data centres it helped justify become the substrate for Google, Amazon, Netflix and the rest of your screen addiction.

The pattern is almost boring:

The hallucination dies. The technology moves in and never leaves.

This is the level at which AI looks suspiciously normal.

Step two: two AIs for the price of one

Right now we are living with two overlapping AIs:

1. AI, the capital story

2. AI, the plumbing

The capital story is the one you see on CNBC:

• Data centres going up like mushrooms after rain.

• GPU clusters hoarded like Cold War tinned food.

• Corporate debt underwriting “the largest capex cycle in history”, while the investor deck insists this is obviously fine because TAM.

Those numbers only work if AI revenue turns up quickly, at scale, and with fat enough margins. If it doesn’t, somebody is left with beautifully cooled server barns.

That part absolutely has bubble energy:

pricing that assumes a flawless S-curve, a clear set of winners, and no awkward “oh, these unit economics are terrible” moment.

Now, beneath that, there is AI, the plumbing.

This is much less cinematic:

• People who write code ship more of it, because the autocomplete now does 60–80% of the boilerplate and refactors.

• People who write words draft in minutes and edit in hours instead of the other way round.

• Customer support stops being a slow-motion hostage situation.

• Small businesses behave like they have a half-competent research and ops team living inside a browser tab.

The early data lines up with the anecdotes:

• Within teams, AI used properly gives chunky productivity gains (double-digit type gains!) especially for people who already know how to frame problems and judge output.

• Across firms, the gains are uneven and slower, because you can’t just lob a model at a century of legacy process and call it transformation. You have to change workflows, incentives, and your tolerance for machines that occasionally go rogue and hallucinate HR policy.

At the macro level, the signal is still fuzzy but it is not zero.

Adoption curves for generative AI look steeper than early internet or PCs. Productivity numbers have nudged up at exactly the moment every knowledge worker quietly became a part-time prompt engineer.

None of that feels like tulips. It feels like a messy, early-stage general-purpose technology doing exactly what general-purpose technologies do:

• start in weird corners,

• get misused,

• slowly become infrastructure.

So… is AI a bubble?

If by “bubble” you mean:

“Does this contain speculative nonsense that will end in tears for some investors?”

Then yes. Obviously. Parts of this market are a costume party waiting for the lights to come on.

If by “bubble” you mean:

“Is this fundamentally hollow, like tulip futures or cartoon apes?”

Then probably not.

• The tools are already entangled with real workflows in ways that are hard to fully roll back.

• The infra, once built, will mostly be repriced and repurposed, not bulldozed.

• The skill mix the economy quietly rewards is already tilting toward “can you orchestrate and evaluate machine output” rather than “can you manually produce every token yourself”.

The cleaner diagnosis is:

AI is a genuine general-purpose technology currently wearing an over-priced bubble costume so capital will pay for the plumbing.

The costume will come off. Some companies will turn out to be the costume. Valuations will deflate. Debt will bite. Entire business models will be quietly memory-holed.

And in 20 years, we’ll pretend it was obvious and linear that thinking, drafting, searching, coding and coordinating all got cheaper and we rebuilt a lot of work on top of that fact.

The hallucination is temporary, but the structural shift probably isn’t.

Strange Loop Extra Credit: Go Deeper

1. General-Purpose Technologies & bubbles