AI's biggest critic has lost the plot

Kelsey Piper · article · 2026-04-01 source ↗ Reader 🖥️

AI Backlash AI Bubble Debate AI Unit Economics

Ed Zitron's AI-bubble thesis has deteriorated from a defensible economic argument in 2024 into unfalsifiable fraud accusations in 2026, because he refused to update his conclusion when costs fell and adoption surged, and the public discourse desperately needs better-calibrated skepticism in its place.

practical · argumentative — A sustained critique of another writer's analytical framework, evaluating the quality of AI-skeptic argumentation against economic evidence. Practical because it concerns real investment decisions and the quality of public discourse about them.

Structure

Part	Role
Introduction of Zitron and his 2024 thesis	Establishes the original argument and its plausibility
Evidence that the world changed (cost collapse, adoption surge, capability gains)	Provides the factual basis that should have forced an update
Zitron's pivot to fraud accusations	Documents how the argument shifted when economics stopped cooperating
The silliness-as-argument problem	Identifies a rhetorical sleight-of-hand: anecdotes substituted for economics
Contradictory arguments (overbuild vs. build-out not happening)	Exposes internal inconsistency in Zitron's framework
What better skepticism would look like (Epoch AI analysis, competitive dynamics)	Constructive alternative showing real remaining risks

Key terms

bubble AI Bubble Debate
A dot-com-style scenario where valuations vastly exceed the economic value companies will ever deliver, leading to crashes and lost capital — distinct from 'overvalued'
radical skepticism AI Bubble Debate
Dismissing all available evidence as potentially fraudulent, leaving no data that could falsify the thesis — 'perhaps nothing we see is real'
principled skepticism AI Bubble Debate
Skepticism that accepts established facts and targets the weakest links in a company's case, producing actionable analysis
circular dependencies (in fraud accusations)
A reasoning pattern where each accusation of fraud is supported primarily by other accusations of fraud, creating an unfalsifiable loop
life cycle (of a model) AI Unit Economics
The full economic arc from training cost through revenue generation to retirement — used by Epoch AI to assess profitability

Propositions

explicitAI progress from 2024 to 2026 has been much faster than AI progress from 2022 to 2024, contradicting Zitron's 2024 prediction that progress had peaked.conf: high
explicitGPT-4-level intelligence costs approximately 1/1000th what it cost when GPT-4 was released.conf: high
explicitAbout 30% of Fortune 500 companies have enterprise deals with leading AI startups.conf: high
explicitMore than half of Americans use AI chatbots weekly or more often.conf: high
explicitZitron's shift from economic arguments to fraud accusations is an implicit admission that the economic case can no longer be made.conf: medium
explicitThe bubble case must ultimately turn on revenue projections, not on anecdotes about executive silliness or product failures.conf: high
explicitOpenAI made money serving GPT-5 but did not earn enough to recoup training costs before the model was replaced by the next generation.conf: high
explicitZitron's simultaneous claims that the build-out is dangerously aggressive AND that the build-out is not actually happening are internally contradictory — the second mitigates the first.conf: high
implicitThere are legitimate remaining grounds for AI skepticism — competitive pressure forcing unprofitable model release cycles, winner-takes-all dynamics — but Zitron's framework cannot access them because it dismisses all financial data as fraudulent.conf: high
implicitA critic who does not update their thesis when key premises are falsified is no longer doing analysis — they are doing advocacy.conf: medium
explicitThe coding market alone, worth roughly $250 billion annually in US salaries, represents a concrete economic basis for AI value, even without speculating on future capabilities.conf: high

Arguments

Zitron's AI-bubble thesis has deteriorated from defensible economic analysis into unfalsifiable conspiracy thinking. strength: high

Explicit premises:

In 2024, Zitron argued AI would not improve and businesses were not using it — both empirically checkable claims.
By 2026, AI had dramatically improved and business adoption was widespread.
Instead of updating, Zitron shifted to claiming companies are committing FTX-level fraud with their revenue numbers.
Fraud claims by definition dismiss the evidence that would otherwise refute the bubble thesis.

Implicit premises:

A thesis that cannot be falsified by any available evidence is not a useful analytical framework.
The quality of an analyst should be judged partly by their willingness to update when predictions fail.

Inferences:

The shift to fraud accusations is a retreat from empiricism, not an evolution of the argument.

AI has nonspeculative economic value in 2026, unlike the speculative value it had in 2024. strength: high

Explicit premises:

Over half of Americans use AI chatbots weekly.
30% of Fortune 500 have enterprise AI deals.
GPT-4-level costs have fallen 1000x.
The author personally derives more than $100/month of value from a $100/month Claude subscription.
Software engineering is a $250B annual salary market and AI demonstrably increases programmer productivity.

Implicit premises:

People paying for a product month after month are receiving value commensurate with the price.
Mass sustained adoption is evidence of economic value, not just hype.

Inferences:

The 2024 argument that AI was 'mostly a curiosity' is no longer plausible in 2026.

Zitron's contradictory claims about the data center build-out reveal undisciplined analysis. strength: high

Explicit premises:

Zitron argues the capital build-out only makes sense if demand grows unrealistically fast.
Zitron also argues the build-out is not happening as planned — data centers are delayed or cancelled.
A smaller build-out mitigates the overcapacity problem from the first argument.

Implicit premises:

A coherent analytical framework should check whether its separate claims are consistent with each other.

Inferences:

Zitron treats both facts as independent 'strikes against AI' rather than recognizing they partially cancel out.

OpenAI and Anthropic face genuine competitive risks even if AI has real value. strength: medium

Explicit premises:

Epoch AI found OpenAI does not recoup model training costs before releasing the next generation.
Both companies are releasing models at a pace driven by competition rather than profitability.
Google or open-source alternatives could take the lead.

Implicit premises:

A market with real value can still destroy the companies competing in it if margins are negative.
Winner-takes-all dynamics make individual company bets risky even in a growing market.

Inferences:

Better skepticism would focus on these competitive dynamics rather than denying AI has value.

Narratives

The dot-com bubble analogy: the internet was real and transformative, but Pets.com and similar companies went bankrupt because their specific business models did not capture the value the technology eventually delivered.
→ Frames the AI bubble question precisely — the issue is not whether the technology works but whether specific companies will capture the value.
The all-you-can-eat restaurant analogy: most customers are profitable, but some eat so much they are individually unprofitable — this does not make the business 'deeply unprofitable.'
→ Rebuts Zitron's claim that the existence of heavy users burning $1,000-2,000/week on a $200/month plan proves the entire subscription business is unviable.
Piper's personal list of tasks she uses Claude for (reproducing academic papers, finding books, soccer club registration, kindergarten handwriting activities, coding).
→ Provides concrete first-person evidence that AI delivers real utility across a variety of everyday tasks, countering Zitron's 'What tasks, exactly? Who knows!' challenge.
The hypothetical Amazon article: one could write an article listing Amazon's missteps, dumb executive quotes, and product recalls, but none of that would prove Amazon is a bubble because it makes money.
→ Illustrates that anecdotal silliness is never sufficient to establish a bubble — only revenue analysis can do that.

Questions addressed

⭕ Is AI a bubble in the dot-com sense?
Piper argues the strongest version of the bubble case has weakened significantly since 2024 but does not claim it is fully resolved — real risks remain around competitive dynamics and training cost recovery.
✅ Has Ed Zitron's analysis held up as conditions changed?
No — his specific 2024 predictions (progress peaked, businesses won't adopt) were falsified, and he shifted to unfalsifiable fraud claims rather than updating.
✅ What would good AI skepticism look like in 2026?
Accept that AI has real economic value and adoption, then focus on whether the competitive dynamics (model release pace vs. training cost recovery) are sustainable and which companies, if any, will be profitable.
⭕ Are OpenAI and Anthropic's revenue numbers credible?
Piper treats them as credible on the basis that fabricating them would be a straightforward crime, but does not independently verify them.

Gaps

incompletePiper argues Zitron should focus on competitive dynamics and model lifecycle economics but does not herself develop this argument beyond citing the Epoch AI analysis. The crucial question — whether any individual company can become profitable given the training-cost treadmill — is raised but left open.
Read next: Epoch AI's 'Can AI Companies Become Profitable?' analysis (cited in the article) for a rigorous treatment of the model lifecycle economics.
incompleteThe article treats AI's economic value as largely established via adoption numbers and personal testimony, but does not engage with the question of whether current valuations ($100B+ rounds) are justified by the actual revenue and margin trajectory. Real value does not preclude overvaluation.
Read next: Financial analyses that compare AI company valuations to projected free cash flow — e.g., Aswath Damodaran's work on tech valuation, or Epoch AI's cost-revenue analysis applied to specific companies.
incompletePiper addresses the consumer and enterprise market for AI but does not discuss the open-source dynamic in depth, which she acknowledges in a footnote could undermine both OpenAI and Anthropic even if AI itself succeeds massively.
Read next: Coverage of open-source LLM development (Llama, Mistral, DeepSeek) and whether open-weight models are converging on frontier capabilities.
incompleteThe article mentions that AI companies are 'run by people who are not very trustworthy, who often contradict each other or oversell their products' but does not reconcile this with her dismissal of fraud concerns. If the executives are untrustworthy, some middle ground between 'total fraud' and 'fully credible numbers' might warrant exploration.
Read next: Investigative reporting on AI company financial disclosures and accounting practices — how revenue recognition works for API credits, free tiers, and enterprise commitments.

AI's biggest critic has lost the plot

Structure

Key terms

Propositions

Arguments

Narratives

Questions addressed

Gaps

Related documents