Receipts: The Subprime AI Fact-Check

Ed Zitron has spent the last year building the most widely-read case that AI is a financial fraud. His claim is that the technology itself is fundamentally worthless and the entire industry is a con. His "Where's Your Ed At" newsletters have a large audience, a Webby award, and a Penguin book deal. He bans people from his subreddit for subscribing to AI-positive communities — not for what they say, but for where else they participate. He starts from "AI is worthless" and works backward to find numbers that support it. Some of those numbers are right. Some aren't. The pattern matters.

Sloppish covers the same industry. We think the spending is excessive, the claims are inflated, and the risks are real. We also think AI is useful, that LLMs have genuine applications, and that the technology will continue to matter. We can hold both of those ideas because we follow the data instead of starting with the conclusion. Zitron can't hold that tension — for him it's all fraud, all the time. Which makes him a useful corrective voice but not a reliable analyst. And when I started checking his numbers, a pattern emerged that explains why.¹

The Numbers

Let's start with the biggest miss. In his "Subprime AI Crisis" newsletter, Zitron claims that Cursor's AI coding tool costs roughly $5,000 per month per user to operate.¹ The actual number, based on Kilo Blog's analysis of inference costs for code-completion workloads, is approximately $400 per month per heavy user.³ That's a 12x exaggeration.

This matters because the entire subprime analogy rests on the gap between cost and revenue. Cursor charges $20/month for its Pro tier. If the real cost is $400/month, the unit economics are terrible — a 20:1 loss ratio. That's a strong argument on its own. You don't need to inflate it to $5,000 to make the point. But at $5,000, you've invented a 250:1 ratio that nobody can verify, and the moment someone checks, the entire paragraph loses credibility.

The pattern continues. Zitron writes that credit card debt consumes 20% of household income.² The Federal Reserve's Household Debt Service Ratio, which tracks required debt payments as a share of disposable personal income, stood at 9.92% as of Q4 2024 — and that's all consumer debt, not just credit cards.⁴ Revolving credit (credit cards) accounts for roughly half of that. He doubled a number that was already concerning enough at its real value.

The Small Errors

Some mistakes are smaller but reveal a pattern of not checking. Zitron attributes the coining of "enshittification" to 2022.² Cory Doctorow published the term in his Pluralistic blog on January 21, 2023, and it appeared in Wired a month later.⁵ Off by a year. Not a big deal in isolation. But it's the kind of thing you catch in sixty seconds of Googling, and the failure to do so tells you something about the verification standard.

He writes that Apple TV+ "raised prices 30%." The actual increase was from $6.99 to $9.99 — a 42.9% increase.⁶ This is one of the rare cases where the real number is worse than his claim. He undershot. Which is almost more telling than overshooting: it suggests the numbers aren't being checked in either direction. They're being vibed.

He places the launch of AWS in "2003-2017" for context on cloud computing timelines.¹ AWS launched in March 2006.⁷ Three years off. Minor in a narrative sense, but in a piece that's asking readers to trust its financial analysis, every checkable fact that turns out wrong erodes that trust by a little more.

The Unsourced Claims

This is the bigger problem. Across the two newsletters I audited, at least six major financial claims appear with no source, no link, and no way for the reader to verify them:

Harvey's $11 billion valuation. OpenAI's $850 billion figure. A "27.8% streaming fatigue" statistic — suspiciously precise for a number that appears in no survey, index, or report I could find. Multiple revenue projections presented as fact with no attribution.¹²

Some of these numbers may well be correct. Harvey's valuation has been reported in that range by multiple outlets. But that's the point of citations — they let the reader distinguish between "this is documented" and "I remember reading this somewhere." A publication built on holding others accountable for their claims has to hold itself to the same standard. Receipts require receipts.

The Financial Thesis Holds. The Technology Thesis Doesn't.

The spending argument works on the real numbers.

AI companies have spent over $200 billion in capital expenditure in the last eighteen months, with revenue that doesn't come close to justifying it.⁸ OpenAI reportedly lost $5 billion on $3.7 billion in revenue in 2024. Cursor's unit economics are genuinely unsustainable at actual cost, not just at inflated cost. The comparison to subprime — where financial products were valued far above the assets backing them — is a legitimate analytical framework. BLS data shows tech unemployment ticking upward while companies announce AI-driven headcount reductions and simultaneously increase AI spending.⁹

Zitron is right that the industry is spending money it cannot justify on products whose revenue model is unproven. He's right that the "just wait" defense echoes the "housing prices only go up" logic of 2006. He's right that a correction is plausible.

But "the economics don't work" and "the technology is worthless" are different claims, and Zitron collapses them. The dot-com bubble was real and devastating — and the internet turned out to be the most important technology of the century. The housing crash destroyed the economy — and people still need houses. Financial excess doesn't prove technological failure. It proves financial excess.

Zitron doesn't make that distinction because he doesn't want to. If AI has genuine utility, his narrative gets complicated. Complicated doesn't build a newsletter audience. So the numbers get inflated to make the case simpler than it is, and anyone who sees value in the technology gets banned from the subreddit.

Why It Matters

A critic who inflates numbers hands the target a weapon. Each inflated number and unsourced statistic becomes the response. "Your numbers are sloppy, so why should we trust your thesis?" The dismiss-by-detail strategy. It's what companies do when they can't argue the substance: they argue the footnotes.

Zitron's platform reaches hundreds of thousands of readers who are forming opinions about whether AI is a viable technology or a financial fraud. That's genuine influence. And influence carries a proportional obligation to accuracy — especially when you ban people from your community for disagreeing.

The financial critique deserves better than vibes-based arithmetic, and the technology critique deserves better than predetermined conclusions. And a critic who inflates numbers to 12x their actual value while refusing to engage with anyone who finds the technology useful isn't practicing accountability journalism. He's practicing confirmation bias with a Webby award.

The real numbers are damning enough. The real technology is useful enough. Holding both of those ideas at the same time is harder than picking one and banning everyone who disagrees. But it's the only honest position.