Ed Zitron and the Case Against the AI Bubble: Five Messages, a Track Record, a Measured Verdict

Who he is, and how he got here

Ed Zitron’s career is the least likely origin story for the AI industry’s chief sceptic. Per Wikipedia, he was born in London in the mid-1980s, wrote for a video games magazine and PC Zone until about 2008, and moved into public relations. He founded the PR firm EZPR in 2013.

A transparency note, because he asks it of everyone else: his PR business promoted AI companies, including the companion app Nomi, according to one profile, which says he has since parted ways with it. Casey Newton has made the point that he once promoted such companies. Both facts cut in the same direction: he knows how a pitch is built.

The last four months

Here is what we could verify between July and early November 2026. We found no dated appearances for August, and there may be some we missed.

The pattern is consistent. The venue changes, from a markets podcast to a comedy-adjacent talk show, and the thesis does not.

The five messages

1. It is a money story before it is a technology story

Zitron’s most repeated point is that you cannot understand the AI boom by watching demos. Follow the cash. In “Annualized Enron Rate” (13 October) he argues annualised run-rates are “a marketing exercise”, with OpenAI and Anthropic counting revenue differently, and that many AI startups resell lab tokens at a loss funded by venture capital, so lab revenue depends on startup fundraising.

The OpenAI episode in our reality check gives this point teeth. OpenAI’s run-rate was $50 billion, not $68 billion, and the gap was an accounting convention. On the day it surfaced, Nvidia fell 3%, Oracle nearly 6% and CoreWeave nearly 8%. A market that moves that much on a definition is a market that is leaning on two companies.

2. The unit economics do not make sense

His 2024 essay already said generative AI runs “on faith” rather than money, with OpenAI expected to lose over $5 billion that year. In June 2026 he went further: both OpenAI and Anthropic “are dogs”, heading toward IPOs with no path to profitability. The Monetary Matters episode description says Anthropic reportedly spent $2.75 for every dollar of revenue in 2025 against $2.60 for OpenAI. Those are reported figures from a leaked prospectus, not audited numbers.

This is where the story got complicated on 26 October. Anthropic’s public S-1 shows $16.3 billion of revenue in the first half of 2026 and a $974 million operating profit in Q2, per Yahoo Finance’s live coverage, in the same document as an $80.7 billion first-half net loss. His own H1 estimate in the Enron post was $15.5 billion, within about 5%. So the blunt version of message two (“no revenue”) fails. Size is not unit economics: the $2.75 spent per dollar of 2025 revenue from the leaked prospectus is still in the story beside one green operating quarter, and OpenAI’s 2024 losses-on-faith narrative has not gone away either. The sharp version survives: $498 billion of the $657 billion in compute contracts is noncancelable, roughly $70 billion a year by Fortune’s maths, against a Q2 run-rate of about $46 billion by ours.

3. It is a debt bubble, not an equity bubble

This is the message with the most 2008 in it. On The Tech Report he said data-centre projects are being funded on speculative future revenue the way unaffordable mortgages were, and that pension funds and insurers chasing yield have fed private credit that finances them. He cited reports of more than $1.65 trillion of AI data-centre debt kept off balance sheets, special-purpose vehicles such as Meta’s $27 billion Hyperion deal, and long-term debt financing GPUs that must be replaced within a few years. His forecast: “slow, slow, and then all at once.”

The trillion-dollar figure is a secondary-source estimate, and he is a commentator, not a regulator. But the shape of the argument is sound. Whether it is a bubble or not, off-balance-sheet debt backed by one customer is exactly the structure that makes risk hard to see.

4. It is worse than the dot-com bust

Zitron rejects the comfortable dot-com comparison. His case, in a premium post and the Better Offline series “AI Is Worse Than The Dot Com Bubble”: GPUs are not fiber, AI startups are worse than Pets.com, and the crash threatens market contagion. Laying fiber made sense even then, and it kept being useful. A GPU is worth less every year.

On Bloomberg in June he added the retail angle: “people are conflating a semiconductor rally with an underlying successful business”, and ordinary investors could be left holding the losses.

5. Silicon Valley has run out of ideas, so it invented a god

The broadest message is cultural. In his June essay he describes “the Rot-Com Bubble”: hypergrowth ideas are exhausted, venture capital follows the crowd, and “Silicon Valley stopped building the future, and started investing in its own destruction.” The talk of recursive self-improvement, in his view, is a narrative that justifies spending and delays scrutiny. In September he extended it to the existential-risk story itself.

You do not have to buy the cult framing to notice the commercial fact behind it. Doom and utopia are both fundraising decks.

Dot-com, 2008 and now

Our reading: the dot-com analogy fits the sentiment and the valuations. The 2008 analogy fits the plumbing. Both labs can be large and still levered: billions in quarterly revenue does not cancel a roughly $70 billion-a-year compute bill or an $80.7 billion first-half net loss on the same filing. A business that owes $498 billion it cannot cancel has a Lehman-shaped question to answer if revenue slows. Both things can be true.

The track record, honestly

Where he has been right, or close.

• Run-rate opacity: the OpenAI gross-versus-net gap and the market reaction.
• Commitment risk: he cited $518 billion on 1 October and $413 billion uncancellable in his newsletter. The S-1 shows $498 billion noncancelable of $657 billion. He pointed at the right number before the filing.
• Revenue size: his $15.5 billion H1 estimate versus $16.3 billion reported.
• Concentration: a single revenue report moved the chip complex by 3 to 8 percent.

Where he has been wrong. Dan Luu reviewed roughly 30 dated predictions from February 2024 to late 2025, as summarised by the secondary sources we could reach, and found most did not come true:

• Generative AI as a “dead-end technology” (August 2024).
• OpenAI’s revenue forecasts as implausible to the point of “financial crime” (October 2024). OpenAI met or exceeded them.
• Cursor facing a “firesale” and unable to be worth $10 billion (July 2025). It is reportedly valued around $60 billion.
• Anthropic’s $34.5 billion 2027 revenue projection as “laughable” (February 2025). H1 2026 revenue alone was $16.3 billion, an annualised pace close to that target.
• The bubble popping “no later than Q2 2026”. It has not.

Luu is explicit that this is not a defence of AI hype, and that he has also scored futurist optimists badly. The more serious critique, from Kelsey Piper, is that Zitron’s 2024 economic case was plausible but his 2026 case drifted toward alleging fraud, and that he doubts nearly everything instead of the weakest links. She also concedes that the build-out may well be too aggressive, and her line is the right one: “we desperately need better skepticism.”

There is a fair defence too. One commentator argued that Zitron’s sharpest question, who pays for the build-out, survives his bad forecasts and abrasive style. Critics can be right about his errors while the question stays open.

Our measured verdict

• As a forecaster of timing and technology, he is poor and overconfident. A prediction stated with the highest possible confidence and missed repeatedly is a signal to stop treating him as an oracle.
• As a questioner of financing structure, he has been right more often than is comfortable. The S-1 took away the “no revenue” version of his argument and documented the leverage he had already flagged. It is not a clean bill of health.
• The right way to read him is as a risk analyst. Take the questions (who is the counterparty, what is the cancellation clause, gross or net), leave the verdicts (when it will break).
• Neither camp gets to be lazy. “It works” and “it is financed on faith” are compatible. The technology can be useful and the capital structure can still hurt people.

What would change our mind, in dated, falsifiable form:

If the financing sceptics were early:

• Anthropic’s first public quarters show operating margins compressing as new compute lands, as the S-1 warns can happen before revenue catches up.
• A gross-versus-net revenue restatement at either lab moves chip names the way OpenAI’s $50 billion versus $68 billion report did.
• Credit spreads on data-centre debt widen. Zitron says some SPV debt already trades below par; verify that against market data, not his podcast.

If they were late:

1. Anthropic’s first quarterly report as a public company shows operating income holding as new compute lands. (The S-1 itself warns margins can compress first.)
2. Revenue keeps outrunning a roughly $70 billion-a-year compute bill through 2027.
3. AI-startup gross margins improve. He reports sharp swings at some well-funded startups.
4. OpenAI discloses audited numbers that reconcile gross and net revenue.

If the second list holds, he is wrong again on timing. If the first list holds, he was early, which is what most bubble sceptics turn out to be.

What founders should do with this

• Price at the unsubsidised cost. Zitron’s 2024 speculation that API prices could rise 10 to 100 times if subsidies end is unproven, but a 3x stress test on your AI feature margins is cheap insurance.
• Do not depend on one model or one lab. Keep a second provider warm, and abstract the interface. See our bubble survival playbook.
• Buy audited numbers, not slides. Run-rate is a marketing term. Ask for revenue, margin and renewal data.
• Read the S-1. It is the first real document in this story, and it carries liability that a slide deck does not.
• Use sceptics as a red team. Make one person argue Zitron’s case against your roadmap, and another argue Piper’s.
• Keep building. The models work for real tasks. Our durable AI value versus speculation piece covers where value survives either outcome, and the sourced bubble sceptic case and circular financing go deeper on the numbers.

Bottom line

Ed Zitron is a flawed forecaster who keeps asking the right question. He has been wrong about what the technology can do, wrong on timing more than once, and sometimes wrong on the numbers. He has also been right that AI’s financing is opaque, concentrated and leveraged, and Anthropic’s own S-1 now documents a $498 billion noncancelable compute line under that sentence without resolving whether the IPO math works.

The honest position is the dull one. Take the questions, test the claims, ignore the verdicts, and plan as if both the technology and the balance sheets could surprise you. That is better skepticism than either side of the argument is currently selling.