I’ve been speaking to a lot of smart people with contrasting views of what the future holds lately.

I guess that’s often true, but rarely does it feel so concentrated on a single topic. In this case - AI.

Of course none of us really know where this great experiment leads, and many (probably wisely) are choosing to play the cards they’re dealt today, rather than concerning themselves with hypothetical futures.

But I am not one of those people.

So today’s newsletter includes two things:

1. My summary of the arguments for and against an AI bubble.

2. Links to the latest podcast episode of Humans in the Loop, with Investor, Researcher and Fellow at MIT’s Institute for the Digital Economy - Paul Kedrosky. Paul is convinced that AI is a capex bubble and he’s one of the smartest and most well-reasoned voices on the topic.

Till next time,

Seb

The Great AI Bubble debate

I’ve come to believe that AI is both a transformative set of technologies that will reshape societies AND an overhyped market which is due some sort of correction.

But does that make it a bubble?

This is a topic of fierce debate. As if the questions around the true capabilities of these technologies weren’t enough, now comes the semantic nit-picking over what does and doesn’t constitute a bubble.

As far as I can tell bubbles seem to come in two forms:

1. Snake oil - this might be said of the original dotcom bubble, in the sense that many of the companies had no obvious utility and no real revenue. It might also be said of at least parts of crypto.

2. Infrastructure overbuilds - think railroads or telecoms infrastructure as examples. US railroads were the AI of its day - hyped as a revolution that would open up new frontiers. This was of course true, but that didn’t stop the hype leading to overinvestment and overbuilding. Once the hype died down and valuations on railroad companies adjusted, the US ended up with miles and miles of half-built, non-functional railroads (alongside the usable stuff).

Some people claim AI is a snake oil type bubble, and I can forgive them for doing so. The Sam Altman’s of this world certainly give off a distinctly snake-oily vibe; but dismissing AI as having no utility and/or no revenues seems naive or maybe a form of trying to manifest what you want to be true.

But I think we may be set for an AI overbuild.

Of course I might be wrong. And it’s not an opinion I’m heavily wedded to. It’s just the conclusion I come to when weighing up the arguments for and against.

You will of course draw your own conclusions but let me try to do a good faith writeup of the arguments on both sides.

AI is a bubble

• Spending levels are unprecedented - Alphabet, Meta, Amazon and Microsoft are expected to spend $700bn on AI buildout in 2026, exceeding the annual GDP of a country like Sweden. Not a standalone argument for a bubble but reason to pay attention.

• Spending is disproportionate to revenue - even without any accounting tricks the infrastructure spend is estimated to be at least 7x the revenue generated (this compares to 2x during the railroad bubble as a historical example)

• Spending was being funded out of cashflow but is now being funded by debt - even the richest companies in history are no longer funding their own spending through equity. As of recently the hyperscalers (Alphabet, Meta, Amazon etc.) are now borrowing more than the 6 largest banks in the US combined.

• Some of the AI companies are burning record amounts of money - OpenAI for example, lost an estimated $38 billion in 2025.

• The actual numbers may be far worse - big AI firms have been using accounting tricks like SPVs (special purpose vehicles) to move the spending off their books, meanwhile many circular deals amongst the big AI firms may artificially inflate the revenue numbers e.g. Nvidia invests $100bn in OpenAI, to fund the build out of data centres which OpenAI needs to fill with Nvidia chips.

• Tokens are becoming a commodity - the primary models are taking similar technical approaches and trained on the same underlying data so all perform very similarly. Customers then start to move around based on where’s cheaper/more convenient. You can perhaps see this as OpenAI loses market share to Google (whose other services are already well embedded).

• Demand has been subsidised so far but AI is about to get more expensive - as AI companies try to stem the losses, prices are being significantly increased, and with it question marks over how much demand persists. Many tech companies previously tokenmaxxing are now reevaluating their spending on AI.

• Software engineering is an anomaly - tools like Claude Code drastically increased token usage and seemed to allay some concerns of a bubble. Paul Kedrosky argues that software engineering is an anomaly and most work won’t follow that expansive pathway.

• This isn’t infrastructure that lasts - ~60% of the cost of a data centre is chips. These chips are being treated like a long term asset but may actually only last a few years before they fail or need upgrading.

• LLM improvement is slowing and we don’t know what comes next - this is a contested point as it depends on the benchmarks you use but there are question marks over just how much better the current approaches can get.

AI is NOT a bubble

• Compute is still the constraint - demand is significantly outstripping supply today. Current lead times for Nvidia chips are around 1 year. We’re still a long way off supply catching up with demand so this build out isn’t so much about speculative futures as it is about current demand.

• The hyperscalers are trading at reasonable multiples of their revenue - Microsoft trading at around 30x revenue is modest compared to a company like Cisco in 1999 which was trading at 100x. This suggests valuations aren’t all that inflated and therefore we may not see some significant correction.

• Agentic AI has changed the game - simple chat interactions with LLMs had a distinct ceiling but agents are already proving highly useful in a range of domains that will only increase over time.

• AI like electricity - if AI becomes something that permeates everything we do, like electricity, then the applications are endless and we’ve only just scratched the surface of what’s to come.

• Revenues are growing faster than any companies in history - although spending is unprecedented, so is revenue growth. Anthropic has gone from estimated annual revenues of $87m in January 2024 to $30bn in April 2026.

• Productivity numbers are starting to show progress and we may be early on the J curve - much has been made of the fact that productivity numbers have yet to show a major step change. But some studies are starting to show the impact and whilst initial progress may be slow for lots of human and regulatory reasons, we’ll soon be on the steep part of the curve.

• These are smart people spending the money - many of the world’s sharpest business and technical minds are at the forefront of this capital expenditure. Are they all wrong?