The Four Horsemen and the Half-Built Server Hall — Listening to Ed Zitron

Bare rooms and empty halls, where once the courtiers’ tablets lay in rows; withered grass and dying willows, where once the singing and the dancing were.

— Cao Xueqin, Dream of the Red Chamber, “Commentary on the Won-Done Song” (Qing dynasty, c. 1750; translation mine)

What This Episode Is About

The August 28, 2026 episode of The Compound and Friends features Ed Zitron. He writes a newsletter called Where’s Your Ed At, hosts the podcast Better Offline, and his 2023 essay on the “rot economy” is the piece that made him widely cited. Today he is one of the loudest AI skeptics anywhere.

The hosts framed their own position honestly up front: they’ve been taping for five years, the overwhelming majority of guests have been bulls, and so far the bulls have mostly been right. They’ve had bears on, but non-specific ones — people who are cautious about everything. Someone bearish specifically on AI, with the arithmetic worked out, was a first.

So this plays less like a debate and more like a reconciliation session. The hosts brought concrete bull evidence: operating costs their own firm has cut, and earnings calls where company after company names what AI did for the quarter. Zitron brought concrete bear evidence: revenue concentration, contract commitments, accounts receivable. Neither side was reciting slogans. It sounds like two people holding the same financial statement and pointing at different lines.

The Main Points

1. His case isn’t “AI is useless.” It’s “where does the money come from?” Zitron repeatedly grants that Nvidia is making real money. His question is about the source upstream. His figure: of roughly $34.33 billion in Microsoft AI revenue for the fiscal year just ended, $24.1 billion came from OpenAI alone. Strip OpenAI out and the company he calls “the apex predator of software sales” — hundreds of thousands of resellers, tens of thousands of salespeople — is selling single-digit billions of AI product to every other business on earth, against north of $260 billion in capex. Nvidia has the same shape: one customer at 16% of the most recent quarter, three customers at 44% of the first half, five customers at about 70%.

2. “Run rate” is, to him, the biggest sleight of hand in the whole story. Classic ARR means annual recurring revenue: ten $100 million contracts, a billion a year, done. What’s being reported now is annualized token spend — not a subscription, nothing that renews itself. A customer burning ten million this month may burn half a million next month, especially if they’re migrating to open weights. Nobody defines whether the period is four weeks times twelve or times thirteen, and nobody defines which four weeks. His line: that this gets accepted is an insult to investors’ intelligence.

3. Circular financing, in specific shapes. Nvidia invests in OpenAI; OpenAI buys compute. Nvidia carries roughly $30 billion in cloud compute agreements — renting back its own GPUs — plus about $25 billion in data center leases. His question is hard to answer cleanly: you sell to the data centers, so why are you also renting them? In the same filing, accounts receivable grew 55% sequentially. He didn’t claim to know what that means. He asked it well: why is collection slowing down as revenue accelerates? Are you near the edge of what you can collect without extending terms?

4. The four pale horses. This is where the episode title comes from. Asked what the first sign would look like if he won’t name a date, he listed: a hyperscaler cutting its capex guidance; a debt downgrade at a heavily AI-levered company (he named Oracle, and said the ratings agencies don’t have the nerve); a marquee AI startup running out of money; and a neocloud — he named CoreWeave — failing to roll its debt. A fifth: an IPO that lands badly. What they share is one mechanism: not demand disappearing, but the money tap seizing up.

5. There’s no salvage value this time. The dot-com bust left dark fiber and servers that companies like Amazon later picked up. AI GPUs, he argues, aren’t good for much else; a data center paused for three years costs just as much to finish, and the electricity will cost more, not less. His comparison number: from AWS’s creation in 2003 to its first profitable year in 2015, all of Amazon’s capex, inflation-adjusted, was about $29.7 billion. Not AWS’s — the whole company’s.

6. The bull evidence is solid, and his answer is to move the question. A host cited Airbnb: 45% of customer support calls last quarter closed by AI with no human involved, with specific numbers on cash flow and expenses, achieved after moving to open source models. He also cited their own firm — ninety-odd employees, four thousand client households, finding something new to automate every week. Zitron didn’t claim it was fake. He relocated the question: does any of that saved money flow back to OpenAI and Anthropic? If customers end up on hosted open-weight models, those efficiency gains are unrelated to whether these two companies survive. He also picked apart a fine detail: a report of “80 to 90% savings” at a telecom reads, one paragraph down, as savings in some functions — and one paragraph after that, the company has plugged AI into thousands of things.

7. The details you remember. A report that a private credit firm agreed to invest in a multi-billion-dollar data center project after ten minutes — a host said flatly he didn’t believe it, and Zitron offered to pull up the article. His image for executives claiming AI made them 100x more productive: the opening of The Death of Stalin, everyone standing around the corpse saying he looks great, he looks so healthy. GitHub Copilot letting people burn $5,000 of tokens for $40, then switching to usage billing on June 1 — “an insane thing to do to customers.” And the best analogy in the episode: token prices haven’t risen, but models burn more tokens per task, so it’s a car with the same miles per gallon that you’re now driving two hundred extra miles.

Going Further

”Everyone says it’s a bubble, but my positions keep going up. Should I get out?”

This is the question most listeners are actually carrying, and the most honest passage in the episode is Zitron answering a version of it.

Asked whether he could be wrong, he said: I was wrong, in 2024, and I know why. I was naive. I thought the market was a sensible place, I didn’t think Microsoft, Google, Amazon and Meta would spend hundreds of billions for no reason, and I didn’t think the debt system would support it. What he got wrong wasn’t the arithmetic. It was how long that arithmetic could be sustained. Which is why he no longer gives timelines.

The useful distinction here: being right about the thesis and being right about the timing are independent, and the market only pays for the second one. A flawless bear case can lose you money for two years. A leaky bull case can make you money over the same two years. Confusing the two produces the worst behavior available — selling everything because the argument sounds good, then chasing it back six months later, and losing on both ends.

The workable move is to track the narrative and the funding chain as two separate lines. The narrative line tells you what premium the market is willing to pay right now. The funding line tells you how much longer that premium can be paid. The value of Zitron’s horsemen list isn’t prophecy — it’s that it converts the second line into observable events: a capex guidance cut, a debt round that won’t clear, an IPO that lands cold. Those aren’t feelings; they make the news. You can disagree with every one of his conclusions and still copy the list down as a dashboard.

One counterweight worth writing down: he says he has no money in the market — “I have emotional skin in the game.” A host translated that generously: you’re risking reputation where others risk dollars. Someone with no position never has to solve the genuinely hard problem, which is what day you sell. Borrow his analysis freely. That problem is still yours.

”The numbers look great. How do I tell which demand is real?”

The most practically useful thing this episode does is demonstrate how to read a good-looking financial statement. Zitron isn’t using secret data. Every figure is in a public filing; most people just don’t look at those lines.

Line one: customer concentration. Nvidia discloses single-customer and top-customer percentages. A host made the right bull rebuttal: those are direct customers, and each of them has millions of customers behind it. That’s the best moment in the segment, because Zitron didn’t deny it — he went looking for how big the indirect demand actually is. If the hyperscalers really had that many end enterprises buying, Microsoft’s non-OpenAI AI business wouldn’t be single-digit billions. Concentration alone isn’t damning. Concentration plus missing indirect demand is. The order generalizes to any industry: first ask who’s paying, then ask whether the payer is making money.

Line two: the definition of the revenue. When you see ARR, run rate, annualized, recurring — go find the definition. The test this episode supplies is simple: will this money show up again next year on its own? Subscriptions will. Usage won’t. When a company annualizes usage into a number shaped like a subscription, it isn’t necessarily lying, but it is making an assumption that flatters it — and the assumption isn’t in the headline.

Line three: receivables and inventory. If either grows faster than revenue, the gap between recognized revenue and collected cash is widening. That signal is industry-agnostic; it has nothing to do with AI. Worth noting that this is where Zitron is most restrained — he says he doesn’t know, maybe some customers are just slow, and only then states the suspicion. Naming your uncertainty first is where credibility comes from.

”A headline says a company saved 90% with AI. Can I believe it?”

The episode hands you a clean control pair.

On one side, the “80 to 90% savings” story: a whole-company headline that becomes “in some functions” one paragraph down, and “thousands of things plugged into AI” one paragraph after that. From ninety percent to unknowable, in two paragraphs.

On the other, Airbnb: 45% of support calls closed end to end without a human, with specific figures on cash flow and expense, achieved by moving to open source models.

The difference isn’t honesty. It’s granularity. The first gives you a percentage too large to falsify. The second gives you a scenario small enough to check. A good efficiency claim looks like the second one: it names the process, gives the denominator, and states the cost — Airbnb’s cost being that they stopped paying a model vendor, which is bad news for anyone invested in the AI supply chain.

This test isn’t only for financial statements. Whenever anyone tells you something made them several times more productive, ask the same three questions: which task, how long did it used to take, how long does it take now? Zitron modeled the honest version himself. He used AI to fix his kid’s Minecraft setup; it chased its tail for half an hour, eventually got there, and he probably couldn’t have done it alone. He didn’t say it was useless. He said that for the cost, it didn’t impress him. That’s an answer with granularity. You can disagree with it and still know exactly what he means.

There’s one more distinction worth keeping. He said what actually worries him isn’t people outsourcing looking things up — it’s people outsourcing thinking. A host agreed instantly, which made it the least contested moment in the episode. Interestingly, it has nothing to do with any of the financial argument, and it may be the line with the largest effect on the listener personally.

Worth Reading Further

• The Compound and Friends, August 28, 2026, with Ed Zitron.
• Zitron’s newsletter Where’s Your Ed At and his podcast Better Offline. His 2023 “rot economy” essay is the prototype for everything else he argues.
• Threads mentioned in the episode: a technology consultant’s blog on how AI is eroding corporate decision-making, and several journalists who have spent the last year tracing private credit into data center financing.
• A host mentioned a book on the 1873 railroad boom and bust, and said rereading it made him want to sell everything. He argues it’s a better analogy for this cycle than 2008, because both booms are so physical.
• For the other side, follow the same public material both sides cited here: hyperscaler capex guidance, and what companies actually say when an analyst asks them to quantify their AI return.

The One Thing to Take Away

One idea survives the whole episode: when the money is circular, the number stops being evidence.

Nvidia invests in OpenAI. OpenAI buys Nvidia chips. Nvidia’s revenue rises, its stock rises, and it becomes better able to invest in the next customer. Every number in that loop is individually true, audited, and printed in a filing. Taken together, the loop may contain no dollar that came from outside it. What Zitron does for two hours is hold up a sheet of paper, draw the arrows one by one, and ask: which arrow points in from outside? The answer to that matters more than any single revenue figure.

None of this is confined to investing. Anything you’re evaluating — a job, a relationship, a habit, a plan you badly want to believe in — takes the same question: is some of the evidence supporting this actually produced by the thing itself? “Everyone says this company is great” is one piece of evidence copied many times, if everyone’s source turns out to be the same article. “I’m doing well lately,” supported by the fact that it feels that way, is the same loop.

An exercise for today. Pick the thing you’re currently most certain about — it needn’t be financial. “I’m sure that person doesn’t like me.” “I’m sure this job is a dead end.” “I’m sure this decision was right.” Take a sheet of paper and list, line by line, where the certainty comes from. Then mark each line: OUT for evidence from somewhere you can’t influence (something someone volunteered, an event that happened, a number written down), and IN for evidence your own behavior or inference generated (you believed it, so you acted a certain way, and then took the result as proof).

Count them. If the OUT column has fewer than three entries, what you’re holding may not be certainty. It may be a loop that spins very smoothly. The exercise won’t tell you the answer. It will tell you whether you’ve earned the right to be that sure.

Raw Markdown

This article is an educational discussion of investment method. It is not advice to buy or sell any individual security, offers no target prices, and does not analyze any current holding. Investing carries risk; make your own decisions or consult a qualified professional.