Yeah, It’s a Bubble. Ship Anyway.

“But the most absurd and preposterous of all, and which shewed, more completely than any other, the utter madness of the people, was one started by an unknown adventurer, entitled”A company for carrying on an undertaking of great advantage, but nobody to know what it is.”” Charles Mackay, Memoirs of Extraordinary Popular Delusions and the Madness of Crowds (1841)

Your cousin texts you a chart. Your CFO wants the AI line item frozen “until the market settles.” Somewhere in the same week a startup with $0 revenue and a $2B round quietly dies, and the comment section files the autopsy: see, LLMs don’t work. Meanwhile the boring ticket agent you built is still closing tickets at 4am and has no opinion whatsoever about Nvidia’s stock price. One of those two things is about to get switched off. The wrong one, usually, by someone reasonable, for reasons that sound great in the meeting.

Bottom line: This is a money bubble. Not a “the tech is fake” bubble. The fundraising, the capex dick-measuring, the circular vendor deals, the GPU churches in the desert: that pops. People will stop lighting dumb money on fire. They will stop raising obscene amounts for a wrapper and a deck. The models, the agents, the adoption curve? Those do not go back in the box.

If you cannot hold both of those in your head at once you will make the two classic dumb trades: either you treat every press release as destiny, or you treat a valuation crash as proof the whole thing was a carnival trick.

When it bites

Your board asks “is AI over?” the week Nvidia coughs. They mean the stock. You still have agents in prod writing tickets.
You freeze a working agent program because Twitter said bubble. You just donated the lead to whoever kept shipping. Ask me how I know: I have killed a working agent over a chart I did not finish reading, then rebuilt the same thing later, slower, and called it a new initiative.

What people mean when they say “bubble”

Say the word in a room and you get four different movies:

Tulips: it’s all fake, nothing underneath.
Pets.com: the companies are fake, the internet wasn’t.
Fiber: they overbuilt the pipes. The pipes were real. Somebody else used them cheap later.
2008: debt, fraud, contagion. The houses were real. The paper on the houses was a weapon.

When people talk about this AI bubble they are almost never talking about (1). They are talking about (2) and (3) with a little (4) in the round-tripped cloud credits.

Technical reality: transformers work. Tool-using agents work well enough to replace whole categories of grunt work if you design them like this book says. Context windows got huge. That is not a hallucination of the market. That is 2022-2026 on disk.

Financial reality: the price of the story got stupid. Capex at the five biggest US clouds/AI infra names was being quoted in the mid-to-high hundreds of billions for 2026. Model-lab revenue (even the ones printing real bills) is a rounding error next to the iron they’re burning. Valuations assumed every knowledge worker becomes a seat, forever, at list price, with no DeepSeek in the room and no open weights eating the middle. Fundraising assumed “AI” on the slide was a business model. It is a feature. Sometimes it is a cost center.

A bubble is when the money believes a story faster than the cash flows can catch it. That is this. It is not when the compiler stops working.

The pattern

Every general-purpose tech does this:

It works. Narrow, then less narrow.
Money notices. Capital sprints. You cannot raise a normal round anymore; you raise a narrative round.
Spend goes feral. Everyone has to show a cluster, a deal, a partnership with a lab, a keynote. Circular revenue: I buy your cloud, you buy my model, we both book it.
The dumb money arrives. Wrappers. “AI-powered” toothpaste. Agencies charging $80k to paste a system prompt into a GPT wrapper. I have been on the buying end of that invoice. I read the deck, nodded at the architecture diagram, and paid five figures for a text box with my own instructions in it. Jeremy Schoemaker, veteran of two exits, bought a prompt at retail. Total n00b.
Something breaks the spell. A cheap model from left field. A missed earnings. A round that doesn’t close. Rates. Energy. A fraud. Doesn’t matter which: the permission to be stupid with money gets revoked.
The tech keeps compounding. On cheaper iron. With uglier margins. At companies that had a real customer.

Dot-com: Pets.com is a punchline. Amazon is a country. The fiber they overbuilt is why your house has a pipe. 2000 was a financial event. TCP/IP did not get repealed. The GeoCities under-construction GIF died. The construction did not.

This one will rhyme. The labs that only exist as a story will get merged, acquihired, or quietly wound down. The hyperscalers will write down a data center or three and call it “optimization.” Headcount in “AI strategy” will get a new job title or no job. Tokens will not get dumber. Agents will not forget how to use tools. Your customers will not go back to doing the thing by hand because a Series B exploded.

One worked example

Call this one a composite, because that is what it is. I have lived pieces of it and watched the rest from a folding chair. Nobody handed me an audited spreadsheet on the way out.

You run a company that actually ships. 2024-2026 you got pitched twice a week: “we’ll put an agent on it.” Half those pitches were a chatbot with a logo, and I said yes to one of them because the demo was pretty and I was tired. You still built the boring version: tools, a loop, a stop condition, logs a human can read. My memory of every version of this I have touched is the same: the boring one quietly ate hours nobody was tracking. I cannot tell you how many hours mine saved, because I never wrote a single number down. Not one. Two exits, and I ran my own agents on vibes. That is its own species of dumb, and I am about to spend a whole book telling you to track the thing I did not track. It also showed up on a slide as “AI transformation” so finance would stop asking.

Put illustrative numbers on it, because the shape only argues if you can run the math. Say 900 tickets a week, every one touched by a human. Stand up the boring agent and it closes 300 end to end, leaves 600 for people, and kicks anything it is unsure about to a queue a human reads. Those are scenario inputs I made up so you can follow the arithmetic, not a measurement I took. The threshold matters more than the volume: if the agent is not clearing a fifth of intake, or humans reopen more than one closure in twenty, you shut it off. Write that threshold down before you turn it on, or you will argue about it later with a market chart in your hand.

Then the tape gets ugly.

If you freeze experiments that never had a customer, good. That was tourist money. If you freeze the agent that closes tickets, you just decided the bubble was technical. It isn’t. The ticket agent does not care what OpenAI is worth this week. It cares whether the tool call succeeded.

Same split inside the industry:

Dies in a pop	Survives a pop
$100M seed for a thin wrapper	A workflow with a measurable stop
Training runs as brand marketing	Inference that a customer pays for
Circular cloud credits booked as ARR	Cash from a human who needed a job done
“Foundation model” on the pitch, no moat	Owns the workflow data the wrapper rents
Headcount that exists to have AI	Headcount that uses AI to ship

You want to be in the right column before the music stops. Not because you’re a moralist. Because the right column still has customers when the left column is a Substack post.

The quiet failure

The loud failure is obvious: you believed the deck and bought the GPU church.

The quiet failure is worse and it will be more common:

You treat a financial pop as a technical verdict and you stop.

That’s how companies missed the web after 2001. The magazine said it was over. The survivors just had cheaper servers and less competition. The people who “knew it was a bubble” and sat on their hands got a story they tell at parties. The people who knew it was a bubble and kept building the real thing got the decade. I was in the first group in 2001, posting first-post-grade certainty on Slashdot about how it was all over, and I have the party story to prove it, which is worth exactly what a party story is worth.

Second quiet failure: you keep spending like it’s 2025 after the permission is gone. Same stupidity, opposite direction. When capital gets scarce you do not get extra points for a 70B fine-tune you cannot name a customer for. You get a board meeting. I have sat in that one holding a 70B fine-tune and a customer list of zero, explaining “strategic capability” to people who wanted a number.

Do / don’t

Separate can it do the job from can this round price forever.
Build agents that earn their keep in hours or dollars. That is bubble-proof. That is also just good taste.
Use the cheap models when they’re good enough. The pop makes “good enough” suddenly fashionable.
Keep the loops, the tools, the evals. Those are technical. They compound.
Take the write-downs like an adult when they come. Don’t rewrite history and claim you never believed.

Don’t

Don’t say “AI is fake” because a fund marked a unicorn down. That’s a category error.
Don’t say “it’s not a bubble” because Claude can write a function. That’s the other category error.
Don’t fundraise on the word. Fundraise on a wedge.
Don’t confuse capex with progress. A warehouse full of GPUs is inventory, not intelligence.
Don’t wait for the pop to start doing the work. The work is the hedge. The room that spends 2026 arguing about timing loses to the one kid who already shipped.

Where this sits in the book

Ch. 1 is the history and the next leap (including the quantum brochure, minus the brochure). Ch. 2 is why the money around that history got drunk. Ch. 3 is what an agent actually is, the part that survives the hangover.

If you came here for a hot take that “AI is over”: wrong book. If you came here for permission to light money on fire because the demo slapped: also wrong book.

The tape will pop. The tech will not. Build like you believe the second sentence.

Sources and receipts

Thesis is Jeremy’s (financial bubble, not technical): that stays as argument, not citation. Everything checkable gets anchored or flagged:

Labs booking vendor compute credits as revenue / circular cloud deals: the only first-party numbers are in Microsoft’s FY2026 10-K, filed 29 July 2026 (https://www.sec.gov/Archives/edgar/data/789019/000119312526323660/msft-20260630.htm), read 9 September 2026. It states total funding commitments of $13.0 billion related to the OpenAI investment, of which $11.9 billion had been funded as of 30 June 2026; an equity-method stake of about 25% on an as-converted basis; and fiscal 2026 revenue from commercial arrangements with OpenAI, inclusive of revenue-sharing payments, of $24.1 billion, with $6.0 billion of accounts receivable from OpenAI at year end. Microsoft does not break that $13.0 billion into cash versus Azure compute credits anywhere in the filing, and its own announcement (Official Microsoft Blog, 23 January 2023, https://blogs.microsoft.com/blog/2023/01/23/microsoftandopenaiextendpartnership/) says only “multiyear, multibillion dollar investment” and names no figure. So the split is Jeremy’s read, not a disclosed fact. The point survives: the number everyone quotes is a commitment, not cash, and the same vendor booked $24.1 billion back from the company it funded. Other vendor-to-lab commitments got announced the same way, in blog posts with no figure attached, so this is the one I can put a filing behind and the only one I will.
“Capex in the mid-to-high hundreds of billions for 2026” across big clouds: every per-company guidance figure and industry total I could chase traced back to press summaries, not a filing or an earnings call I could open. So take the size as my read of the direction, not a measurement. The shape of the argument does not need the exact number: it needs the gap between what is being spent and what is being collected, and that gap is not close.
Context-window / tool-use capability claims (2022-2026): uncontroversial directionally; pin with 2 model cards or changelogs at edit time (e.g. function-calling releases, long-context releases) rather than a single paper.
Dot-com / fiber / Pets.com / 2008 analogies: standard history, told as analogy. No citation needed beyond “analogy, not data.”
Worked example (boring ticket agent vs. frozen budget): ILLUSTRATIVE composite, flagged as such in the text. Jeremy’s own memory of shops he worked in, told as shape, not as measured data.
“DeepSeek in the room” line: Cheap open-weights pressure (DeepSeek and others) continues to force cloud providers to compete on inference margins, not markups.
Anthropic: round size, quarterly revenue, and IPO timing all circulated in 2026 coverage; none survived a check against a primary document, so none of those numbers print here. Forward 2027 ARR projections circulate in the press with no named author I could stand behind, so no number for them prints in this book. That is the point anyway: the gap between a reported quarter and a forecast nobody will sign is the whole bubble in one line.