Pundit ClaimsAI claims by Gary Marcus and Ed Zitron, with the evidence

Ed Zitron Where's Your Ed At Published

AI's Economics Don't Make Sense

We recorded 31 claims from this post. 21 were checked against evidence. 10 can’t be checked. 25 more are held. Read the post at its source.

Claims from this post (31)

In the order they appear in the post. Each quotation is the speaker’s words in this post. A label such as “Supported” or “Can’t be checked” is the claim’s current assessment; its page has the evidence and the reasons.

  1. Assessment: UnresolvedSaid

    “While Copilot might not be “...the same product it was a year ago,” very little has changed about the underlying economic mismatch: that Microsoft was allowing users to burn more than their subscription costs in tokens every single month for three years.”

    What we tested: For three years Microsoft let GitHub Copilot users burn more in tokens each month than their subscriptions cost.

  2. Assessment: UnresolvedSaid

    “And that day has finally arrived, because every single AI service you use subsidized compute, and every single service is losing money as a result:”

    What we tested: Every single AI service you use subsidized compute, and every single one is losing money as a result.

  3. Assessment: UnresolvedSaid

    “On the back end, AI startups are annihilating cash, with up until recently Anthropic allowing you to burn upwards of $8 in compute for every dollar of your subscription. OpenAI allows you to do the same, though it’s hard to gauge by how much.”

    What we tested: AI startups are annihilating cash: until recently Anthropic let subscribers burn upwards of $8 in compute for every dollar of their subscription, and OpenAI lets them do the same, though it is hard to gauge by how much.

  4. Assessment: MixedSaid

    “while prices for some models might have come down, newer “reasoning” models burn way more tokens, which means the cost of inference has, somehow, gotten higher over time.”

    What we tested: Although prices for some models might have come down, newer reasoning models burn far more tokens, so the cost of inference has gotten higher over time.

    What held: The study's April 2024–November 2025 data show increasing frontier expenditure alongside longer reasoning and larger models. This supports the cost-escalation mechanism for frontier workloads, although its causal attribution is observational. (arXiv — Gundlach, Lynch, Mertens and Thompson, )

    What did not: The analysis finds falling total inference expenditure at fixed benchmark scores, with halving times measured in months. This contradicts an unqualified rise in the cost of comparable performance, while remaining compatible with more spending on the frontier. (Håvard Tveit Ihle / LessWrong, )

  5. Assessment: UnresolvedSaid

    “Both assumptions were wrong, because the monthly subscription model does not make sense for any service connected to a Large Language Model.”

    What we tested: The monthly subscription model does not make economic sense for any LLM-connected service.

  6. Assessment: UnresolvedSaid

    “There has never been — and never will be — an economically-feasible way to offer services powered by LLMs without charging the actual token burn of each user, and in the process of deceiving said users, these companies have created products with illusory benefits and questionable return on investment.”

    What we tested: There never has been and never will be an economically feasible way to offer LLM services without charging each user's actual token burn.

  7. Assessment: Can’t be checkedSaid

    “Sidenote: In fact, you can burn a great deal of money without ever getting the outcome you desire, because LLMs aren’t really artificial intelligence at all!”

    What we tested: You can burn a great deal of money without ever getting the outcome you want, because LLMs are not really artificial intelligence at all.

  8. Assessment: UnresolvedSaid

    “I think it’s inevitable that the majority of AI subscriptions move to token-based billing, especially as both Anthropic and OpenAI have now done so with their enterprise customers.”

    What we tested: The majority of AI subscriptions will inevitably move to token-based billing.

  9. Assessment: Can’t be checkedSaid

    “The real thing to look out for — a true pale horse — will be a major AI lab like Anthropic or OpenAI moving all of its subscribers to token-based billing. Once that happens, you’ll know it’s closing time.”

    What we tested: If a major AI lab like Anthropic or OpenAI moves all of its subscribers to token-based billing, that will be a true pale horse and you will know it is closing time.

  10. Assessment: Can’t be checkedSaid

    “In reality, that 10-person, $1250-a-month Teams subscription likely burns anywhere from $5000 to $10,000 a month in API calls, if not more.”

    What we tested: A 10-person $1,250-a-month Claude Teams subscription likely burns $5,000 to $10,000 or more a month in API-equivalent costs.

  11. Assessment: Can’t be checkedSaid

    “This will likely lead to more layoffs to keep up with the cost, as was the case with Meta, and then an eventual pullback when somebody asks “does any of this shit actually help us do our jobs faster or better?””

    What we tested: If investors begin to question mounting R&D costs as revenue growth fails to follow, this will likely lead to more layoffs to keep up with the cost and then an eventual pullback.

  12. Assessment: Can’t be checkedSaid

    “And once everybody switches to token-based billing, I’m not sure we’ll see quite as much hype around generative AI.”

    What we tested: Once everyone switches to token-based billing, hype around generative AI may decline.

  13. Assessment: UnresolvedSaid

    “Of the 114GW of data centers supposedly being built by the end of 2028, only 15.2GW is under construction in any way, shape, or form.”

    What we tested: Of the 114GW of AI data centers announced for completion by end of 2028, only 15.2GW is under construction.

  14. Assessment: Can’t be checkedSaid

    “As a result, every AI data center starts millions of dollars in the hole, and even with six-year-long depreciation schedules, takes years to pay off… and with NVIDIA’s yearly upgrade cycle, those GPUs are unlikely to make that much money once you’re done with your first customer contract.”

    What we tested: AI data centers take years to pay off, and their GPUs are unlikely to earn much after the first customer contract because of NVIDIA's yearly upgrade cycle.

  15. Assessment: UnresolvedSaid

    “It’s also unclear whether the customer base for AI compute exists outside of OpenAI and Anthropic, whose demand accounts for 50% of AI data centers under construction, creating a massive systemic weakness if either of them lacks the money to pay.”

    What we tested: OpenAI and Anthropic account for 50% of AI data centers under construction, a systemic weakness if either cannot pay, and demand outside them is unclear.

  16. Assessment: UnresolvedSaid

    “As of April 27 2026, only two buildings are operational and generating revenue, and the third barely has any IT gear in it. I estimate the total cost of Stargate Abilene to be around $52.8 billion.”

    What we tested: As of April 27, 2026, only two Stargate Abilene buildings are operational and generating revenue, the third barely has any IT gear, and the author estimates the total cost at around $52.8 billion.

  17. Assessment: UnresolvedSaid

    “I should also be clear that Oracle is mortgaging its entire fucking future on projects like Stargate Abilene, incurring billions of dollars of costs up front for a business that will take years to turn a profit even if OpenAI makes every single payment in a timely manner.”

    What we tested: Oracle is staking its future on Stargate projects that will take years to turn a profit even if OpenAI pays on time.

  18. Assessment: UnresolvedSaid

    “It is unclear if that’s 200MW of critical IT capacity or the total available power at the Abilene campus, and in any case, this is only enough power for two buildings, which means that Oracle is most decidedly not “on schedule.””

    What we tested: Oracle's Stargate Abilene is behind schedule, with only enough power for two buildings.

  19. Assessment: UnresolvedSaid

    “And if OpenAI can’t pay for that compute, Oracle dies, because it’s taken on around $115 billion in debt just to build Stargate’s data centers, and needs another $150 billion to finish them:”

    What we tested: If OpenAI cannot pay for its Stargate compute, Oracle will die, having taken on about $115 billion in debt and needing $150 billion more.

  20. Assessment: UnresolvedSaid

    “Either way, Oracle is insufficiently-capitalized to finish Stargate Abilene. It will need at least another $150 billion to get this done, and that’s assuming that other partners pick up about $30 billion in costs.”

    What we tested: Oracle is insufficiently capitalized to finish Stargate Abilene and will need at least another $150 billion to do so, assuming other partners pick up about $30 billion in costs.

  21. Assessment: UnresolvedSaid

    “I really need to be clear that Oracle has no other path to making this revenue without OpenAI, and these projects are entirely financed and paid for using the projected cashflow of the data centers themselves.”

    What we tested: Oracle has no other path to making this revenue without OpenAI, and these projects are entirely financed and paid for using the projected cashflow of the data centers themselves.

  22. Assessment: UnresolvedSaid

    “Anthropic and OpenAI both intend to burn tens of billions of dollars in the next few years, and neither of them have a path to profitability.”

    What we tested: Anthropic and OpenAI both intend to burn tens of billions of dollars in the next few years, and neither has a path to profitability.

  23. Assessment: UnresolvedSaid

    “This means that a large part — if not the majority of — AI compute revenue is dependent on a continual flow of venture capital and debt, both of which are only made possible by investors that still believe that generative AI will be the biggest, most hugest thing in the world.”

    What we tested: A large part, perhaps most, of AI compute revenue depends on continued venture capital and debt.

  24. Assessment: Can’t be checkedSaid

    “Members of the media, I say this with abundant respect: Anthropic is mistreating its customers, and it is doing so because it believes it can get away with it.”

    What we tested: Anthropic is mistreating its customers, and it does so because it believes it can get away with it.

  25. Assessment: Can’t be checkedSaid

    “AI is a con, and this is how the con works.”

    What we tested: The AI industry is a con.

  26. Assessment: UnresolvedSaid

    “I truly believe that LLM-based subscriptions are going to go away entirely, at least at the scale of any product that generates code, and in doing so, Amodei and Altman will wrap up their con, or at least believe that they have.”

    What we tested: LLM-based subscriptions will go away entirely, at least for code-generating products.

  27. Assessment: UnresolvedSaid

    “This is a blinking red fucking light, and in a sane market would send Oracle’s stock into a tailspin, because OpenAI’s ascent to over $280 billion in annual revenue is critical to Oracle’s ability to not run out of money.”

    What we tested: Oracle avoiding running out of money depends on OpenAI growing to over $280 billion in annual revenue, so the CFO's worries are a serious warning sign.

  28. Assessment: UnresolvedSaid

    “This is the kind of thing that happens before a company starts dying. OpenAI’s growth is slowing at precisely the time it needs to accelerate. It needs to effectively 10x its current business by 2030 to make its obligations.”

    What we tested: OpenAI's growth is slowing at precisely the time it needs to accelerate, since it needs to effectively 10x its current business by 2030 to meet its obligations, and this is the kind of thing that happens before a company starts dying.

  29. Assessment: Can’t be checkedSaid

    “It’s an insult to both good software and good taste — the most-expensive, least-reliable applications ever invented, their mistakes forgiven, their mediocrities celebrated, their infrastructure hailed as an inert god of capital.”

    What we tested: OpenAI's and Anthropic's products are the most expensive and least reliable applications ever made.

  30. Assessment: Can’t be checkedSaid

    “Generative AI is an insult. It is unreliable, its economics don’t make sense, its outcomes don’t justify its existence, and the perpetrators of its con are boring, oafish and avaricious men disconnected from society and anybody who would ever disagree with them.”

    What we tested: Generative AI is an insult: it is unreliable, its economics don't make sense, its outcomes don't justify its existence, and the perpetrators of its con are boring, oafish and avaricious men disconnected from society.

  31. Assessment: UnresolvedSaid

    “The people that have so aggressively worked together to hold up this industry have only delayed its inevitable fall.”

    What we tested: The AI industry will inevitably fall; its supporters have only delayed it.

Held from this post (25)

These records did not pass every check, so nothing about them is published. Held is not a finding.