The Music Is Still Playing

In July 2007, with credit markets showing the first visible cracks, Citigroup CEO Chuck Prince gave an interview to the Financial Times. "When the music stops," he said, "in terms of liquidity, things will be complicated." Then he added the sentence that has followed him ever since: "But as long as the music is playing, you've got to get up and dance. We're still dancing."

Three months later, he resigned. Citi took $65 billion in losses. The music had stopped.

The sentence is evidence of something more interesting than stupidity: rational participation in a system you know is broken, because the alternative — sitting out while everyone else dances — is its own kind of loss. The FOMO is structural. The chair will be missing when the music stops; the only question is who is left standing.

I have been getting whiffs of familiar music.

The Technology Is Real

Let me be precise about what I am not claiming. The internet was real. It did change everything. It killed most of the 1999 cohort first — along with a few hundred billion in investor capital — before the survivors built the infrastructure we now rely on. The timing was catastrophic for anyone who showed up at the wrong point in the cycle. The technology was not the problem.

AI is the same shape. The models work. They do things that were not possible three years ago, at costs that have dropped faster than anyone projected. The underlying technology is doing genuine work. The business models layered on top of it, in many cases, are Pets.com.

Dark Fiber

The clearest parallel is infrastructure. In the late nineties, Global Crossing, 360networks, and Enron Broadband built transcontinental fiber networks on projected demand that was correct in direction and wrong in timing and scale. The fiber sat dark. The companies went bankrupt. A decade later, the fiber lit and became the backbone of everything. The investors who funded the original build got wiped out; the investors who bought the assets out of bankruptcy made fortunes.

The hyperscalers are building GPU clusters on the same thesis. Microsoft, Google, and Meta have committed hundreds of billions in capital expenditure on the assumption that enterprise AI adoption materializes at the pace their models require. The compute will probably get used eventually. The structural difference from dark fiber: GPU capacity depreciates too fast to wait a decade. The intermediate carnage may be considerable regardless.

Pets.com

Hundreds of companies have raised Series A and B rounds to build thin interfaces on top of OpenAI or Anthropic APIs. No proprietary model. No moat. No differentiation. Entirely dependent on infrastructure they don't control, at pricing that can change with a single product announcement. The underlying API can replicate their core feature in an afternoon.

Pets.com at least had a sock puppet with genuine brand recognition. The current cohort has a ChatGPT wrapper and a pitch deck.

Enron's Accounting

Enron's signature move was mark-to-market accounting — booking the net present value of a twenty-year contract as revenue on the day it was signed, before a dollar had been earned. The profits were real on paper. The cash was not.

AI companies are valued on what their models might be worth in five years, not what they're earning now. The complex strategic partnership structures — Microsoft's investment in OpenAI, Google's in Anthropic — make the actual unit economics genuinely difficult to read from outside. Andy Fastow's Special Purpose Entities were also difficult to read from outside. That was the point.

The CDO

Take a pool of variable-quality assets. Tranche them. Label the senior tranche AAA. The rating agencies, paid by the issuers, obliged. The label did work the underlying assets could not.

"AI company" is currently doing similar work. OpenAI, Anthropic, and a three-person wrapper burning through a seed round are all "AI companies" in the way that Goldman's structured credit desk and a Florida condo flipper were both participating in "real estate." The category elevates the valuation. The underlying unit economics are not invited to justify it.

The AI benchmark ecosystem has the rating agency problem too. MMLU, HumanEval, the standard model leaderboards — increasingly gamed by providers training on benchmark-adjacent data, optimized for specific tests. They look authoritative. They are the Moody's ratings of the AI stack.

The Subprime Mandate

Corporate AI spending is increasingly driven by the fear of being left out rather than viable return on investment. The CFO without an AI strategy is the 2025 equivalent of the bank not participating in mortgage-backed securities in 2005 — apparently missing something, possibly the only adult in the room.

The bad loans were hidden in CDO structures. The bad AI ROI is hidden in digital transformation budget lines that nobody is currently required to justify on unit economics. When CFOs start asking those questions — and they will, because CFOs always eventually ask those questions — a significant amount of this spending stops.

A Casual Observer

I was not in the room for either of these. I am not in the room now.

In the winter of 1998 I was doing a consulting engagement for Cyberian Outpost — an early internet software retailer, one of the companies that was supposed to eat Egghead's lunch as catalog shopping moved online — in Kent, Connecticut. (Birthplace of Seth MacFarlane, which is the most interesting thing about Kent, Connecticut in winter.) What I remember most clearly is how much more energy was going into day trading than into building the product. The employees were speculating on the market while the company was trying to compete for its life. The speculation had stopped being a side activity. It had become the culture.

A friend whose husband worked at Shearson Lehman Brothers told me, sometime around 2006, that things were precarious. I asked what she meant. She said: "You have no idea how bad it's going to get." Not a specific prediction. Not details she could share. Just magnitude — delivered with the flatness of someone who had been watching the instruments and had stopped being surprised by what they said.

Both times, I was outside looking in, catching the whiff without access to the readings. I am in the same position now.

What's Actually Different

The financial leverage is different. The subprime collapse cascaded through the banking system because the instruments were debt-layered and the counterparty exposure was systemic and unmapped. The AI bubble is mostly equity speculation — venture capital, public market enthusiasm, corporate capex. When it deflates, there will be no TARP, no emergency Fed intervention, no formal designation of anyone as too big to fail. The losses will land on investors and employees and the executives who approved the budgets. It probably won't take out a bank.

The technology is also different from Pets.com's business model, which was structurally impossible — subsidizing delivery economics indefinitely in a thin-margin logistics category on the theory that scale would fix unit economics. It wouldn't have. AI's unit economics are improving rapidly and the underlying capability is genuine. The question is whether current valuations are priced for the technology or for the narrative layered on top of it.

The canonical dot-com survivor is Amazon. When is the last time you bought a book from Amazon.com? The company that emerged from the wreckage is a cloud infrastructure business — AWS — that the bookstore happened to build in order to run itself, and which turned out to be worth more than everything else combined. The books are still there, somewhere, incidental to what the company actually became. Whatever survives the AI bubble will probably be equally unrecognizable relative to whatever it is currently calling itself.

The similarity worth not forgetting: what there also won't be — because there never is — is anyone going to jail. The pattern of corporations socializing their losses while privatizing their gains has survived every bubble, every bailout, and every subsequent reform. It is more durable than any of the companies involved.

The Gap

Chuck Prince knew the music would stop. He danced anyway, because the structure of his incentives required it.

The people building AI companies know that most of their cohort will not survive the next funding cycle. The people writing corporate AI transformation checks know that most of those budgets will not produce measurable returns. The people committing to GPU buildouts know that not all that capacity will find revenue to justify it.

They are dancing anyway. The music is still playing.

There is clearly a gap between the narrative and the unit economics. The question worth asking — the one that tends to go unasked until it's too late — is how wide it is, and who is standing when the chair goes missing.

I don't have anyone whispering the magnitude to me this time. What I have is the whiff — the day trading energy in the wrong place, the sock puppet confidence, the familiar shape of instruments nobody wants to look at too closely.

You have no idea how bad it's going to get is a sentence I hope I'm wrong to be thinking about.

---

Published by Geordie