Find this research useful? Add aibubblequestion.com to your Google Preferred Sources, one click, affects only your own results.

The AI Bubble Question
All pages

Weighing It

The Critics

The loudest sustained critic of the AI boom is Ed Zitron, of the Where's Your Ed At newsletter and Better Offline podcast. Some of his numbers were later confirmed; some of his predictions have missed. This page applies his own standard to his record: sourced fact, estimate, and prediction are different things.

Updated

His sourced facts, and what happened to them

Zitron's most consequential work rests on leaked documents, and much of it has been independently corroborated:

  • OpenAI's payments to Microsoft. Documents he obtained showed OpenAI paid Microsoft $493.8 million in revenue share in 2024 and $865.8 million in the first three quarters of 2025, implying (at the reported 20% share rate) revenue consistent with other outlets' figures. TechCrunch verified the documents (TechCrunch).
  • OpenAI's inference bill. The same leaks put OpenAI's spending on running its modelsRunning a trained AI model to answer questions or generate text. Training builds the model once; inference is the ongoing cost of using it, every single time.Full definition in the glossary at ~$3.8 billion in 2024 and ~$8.65 billion in the first nine months of 2025, at times exceeding its revenue (TechCrunch).
  • OpenAI's 2025 results. Leaked audited financials he reported, $13.07 billion of revenue, $34.0 billion of costs, a $20.92 billion operating loss, and a $38.53 billion loss attributable to OpenAI, were subsequently corroborated by the Financial Times (Runtime Wire). More than $2 spent per $1 earned, at scale. These figures also appear on The Demand Question.
  • The announced-versus-built gap. His April 2026 analysis cited trackers showing about 114 gigawatts of announced AI data centers worldwide, accurately reproduced here from his own figure. The paired "under construction" comparator, previously stated as 15.2 GW, has no reproducible methodology behind it in his post or anywhere else located across three independent search passes, so it is unverifiable rather than a confirmed match to independent queue data, and the 15.2 GW precision is dropped. It anchors the grid section of Outside Shocks, marked the same way there.
  • Copilot's old unit economics. He repeatedly cited a 2023 Wall Street Journal report that GitHub Copilot lost over $20 per user per month on average, weakly corroborated, not disproven or proven. GitHub's own 1 June 2026 shift of all Copilot plans, Student, Pro, Pro+, Business, and Enterprise, to usage-based "AI Credits" billing is consistent with a per-user-loss read, the textbook response to one, without itself measuring a current loss rate (GitHub Changelog).

His "AI carousel" framing, suppliers underwriting their own demand, anticipated a structure Bloomberg's later circular-deals reporting also documents, but Bloomberg frames the same reciprocal-investment pattern in its own, more neutral terms, a possibly legitimate demand-and-supply "flywheel," explicitly distinguished from fraudulent round-tripping, not as vindication of "carousel." The specific post or episode where Zitron coined "carousel" was not identified in this pass; until it is named and dated, treating this as settled fact overstates what can be checked (the AI Money Loop covers the same structural pattern independently) (Bloomberg).

His estimates: weaker, and flagged as such

  • ChatGPT Plus subscribers. He did publish this, in his own byline: citing The Information's 28 April 2026 reporting of OpenAI's own internal forecast that Plus subscribers would fall from 44 million (2025) to 9 million (2026), offset by a cheaper tier reported, unverified (Zitron, 28 Apr 2026). "Leaked" is not his word; he attributes the forecast to The Information's reporting, not to a document he obtained himself. The prediction itself, whether Plus actually collapses to 9M, remains open and unverified; treat as report, not fact, but the sourcing chain is now Zitron's own post, not a secondary recap.
  • Energy and water figures. His environmental claims (US data-center load reaching 8-10% of electricity by 2030; millions of gallons of cooling water daily) track the ranges in independent studies, though he consistently emphasizes the worst-case end (academic study).
  • "No ROI in AI." As a blanket claim this sits poorly beside Microsoft's disclosed $37 billion AI run rate and Meta's ad gains; critics engaging his numbers (podcast interviewers, analyst bloggers) press exactly this point (Sloppish scorecard).

His predictions: a mixed record, by his own admission

In 2023-24 he implied OpenAI could face existential funding problems within a year or two. That was refuted on its own timetable: OpenAI closed $122 billion in committed capital at an $852 billion post-money valuation on 31 March 2026, the largest funding round on record. Zitron has not retracted the underlying thesis; his 18 August 2026 post extends it with two new, dated claims, an imminent liquidity need within about three months, and a specific prediction that OpenAI "will almost-certainly have to raise another round of funding by March 2027," likely near $122 billion again, and repeatedly through 2030. Present this as his current, re-dated position, not a retraction, consistent with the historical base rate that even well-sourced bears mistime booms. His GPU-depreciation "fraud" framing has also drawn substantive pushback: the useful-life extensions were disclosed, auditor-reviewed, and mostly predate the boom, as covered in the depreciation controversy. This is the same pattern the historical record shows for bearish callers generally: right early looks identical to wrong, sometimes for years.

His media criticism

Zitron argues that tech media stays credulous because it is financially and access-dependent on the companies it covers: sponsorships, conferences, and exclusive briefings reward friendly coverage, and unverified figures ("$100 billion projections") get repeated without interrogation (Carnegie Endowment interview, The Guardian). The record offers him partial support: several of his leak-based numbers were ignored until larger outlets confirmed them. The interpretive leap, that this amounts to coordinated narrative management, is his opinion, and this site notes it as such.

Other sceptical voices in this window

Zitron is not the only critic to publish new material in this window, and the record shouldn't read as if he were. Michael Burry cited a real, independently retrievable academic paper, "Intelligence per Watt" (Saad-Falcon, Narayan, et al., 7 August 2026), which finds 88.7% of single-turn chat and reasoning queries can be handled by small local models and reports a 5.3x efficiency gain over two years, to argue frontier GPU demand is overstated relative to what inference actually requires. It is a preprint, independently retrievable but not itself peer-reviewed. Jim Chanos, quoted from a 31 July 2026 podcast appearance, put the moment "closer to a '99-type moment than a '97," naming the run-up, not yet the crash, as the closer historical analog. The underlying podcast audio is paywalled and was not independently accessed, so this rests on Benzinga's report of it, journalism about a claim rather than the claim's primary document. Neither is a filed fact, and neither is Zitron's own record above; both are steelmanned, not just logged, on The Case Against page, where a countervailing voice from the same window is tracked so the two pages stay in balance.

How to use a critic

The research corpus's verdict, which this site adopts: treat his leak-based reporting as serious but incomplete, his structural theses as a coherent bear scenario, and his timing record as evidence that even well-sourced bears mistime booms. That last point cuts in every direction at once, which makes it the most useful thing on this page.