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

Ed Zitron Where's Your Ed At Published

Why Are We Still Doing This?

We recorded 33 claims from this post. 17 were checked against evidence. 16 can’t be checked. 30 more are held. Read the post at its source.

Claims from this post (33)

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: Can’t be checkedSaid

    “Let’s start with a simple one: AI boosters are no longer allowed to explain what’s good about AI using the future tense.”

    What we tested: AI boosters should not justify AI by pointing to future capabilities, only present ones.

  2. Assessment: Can’t be checkedSaid

    “OpenAI claims it made $13.1 billion in revenue in 2025 and “only” lost $8 billion — but those numbers seem unlikely considering my report from November of last year that had OpenAI at $4.3 billion in revenue on $8.67 billion of inference costs through September 2025, and this is accrual accounting, which means these are from the quarters in question.”

    What we tested: OpenAI's claim of $13.1 billion in 2025 revenue and only an $8 billion loss is probably inaccurate, given $4.3 billion revenue against $8.67 billion inference costs through September 2025.

  3. Assessment: UnresolvedSaid

    “Anthropic made $5 billion in its entire existence in revenue and spent $10 billion just on compute.”

    What we tested: Over its whole existence Anthropic has made about $5 billion in revenue and spent about $10 billion on compute alone.

  4. Assessment: UnresolvedSaid

    “Saying obvious stuff like “every AI company is unprofitable” or “there is no path to profitability” or “nobody is talking about AI revenues” is considered unfair or cynical or contrarian, even though these are very reasonable and logical statements grounded in reality.”

    What we tested: Every AI company is unprofitable and there is no path to profitability.

  5. Assessment: SupportedSaid

    “Uber is a completely different business to Anthropic and OpenAI or any other AI company.”

    What we tested: Uber is a completely different business from Anthropic, OpenAI, or any other AI company.

    What happened: Uber describes connecting riders, merchants, couriers, shippers and carriers. The report also identifies Uber One memberships and service costs that vary with activity, limiting any stronger interpretation that Uber shares no commercial features or variable costs with AI providers. (Uber Technologies, Inc., )

  6. Assessment: SupportedSaid

    “Somebody spending $20 a month on a Claude subscription can spend as much as $163 in compute.”

    What we tested: A $20-a-month Claude subscriber can use as much as $163 of compute.

    What happened: The tables report a $20 Pro plan, five million weekly credits and approximately $163 of monthly API-equivalent capacity. The author derives limits from usage responses and token weights from collected observations; this is a capacity estimate, not a disclosure of Anthropic's costs. (she-llac, )

  7. Assessment: Can’t be checkedSaid

    “While I must be clear that Anthropic can limit Claude subscriptions, as can OpenAI limit ChatGPT, I doubt either can do so with precision.”

    What we tested: Anthropic and OpenAI can limit subscription usage but probably cannot do so with precision.

  8. Assessment: Can’t be checkedSaid

    “In both cases, Anthropic (and OpenAI, for that matter) is screwed.”

    What we tested: Whether they subsidize subscribers on purpose or cannot control compute costs, Anthropic and OpenAI are in an untenable economic position on subscriptions.

  9. Assessment: UnresolvedSaid

    “In any case, there is no future for any AI company that uses a subscription-based approach, at least not one where they don’t directly pass on the cost of compute.”

    What we tested: There is no future for any AI company that uses a subscription-based approach without directly passing on the cost of compute.

  10. Assessment: Can’t be checkedSaid

    “This is a huge problem for both Anthropic and OpenAI, as their scurrilous growth-lust means that they’ve done everything they can to get customers used to paying a single monthly cost that directly obfuscates the cost of doing business.”

    What we tested: In chasing growth, Anthropic and OpenAI got customers used to flat monthly prices that hide their real costs, which is a huge problem for both.

  11. Assessment: UnresolvedSaid

    “This means it will be very, very difficult to increase subscription costs on users, and near-impossible to convince them to pay the cost of the API.”

    What we tested: It will be very difficult for AI companies to raise subscription prices and near-impossible to get subscribers to pay API costs.

  12. Assessment: ContradictedSaid

    “First, there’s no way to anticipate how many tokens a prompt will actually burn, which makes any kind of budgeting a non-starter.”

    What we tested: There is no way to anticipate how many tokens a prompt will actually burn, which makes any kind of budgeting for metered usage a non-starter.

    What happened: Anthropic documented a way to constrain monthly API expenditure before the post. This contradicts the claim that every form of budgeting is unavailable, but does not guarantee that a task finishes before its budget is exhausted. (Anthropic, )

  13. Assessment: Can’t be checkedSaid

    “Direct labor and parts are considered part of the calculation of gross margin, and spending on training — both the data and the process of training itself — are absolutely meaningful, and to leave them out is an act of deception.”

    What we tested: AI labs' training costs belong in gross margin, and leaving them out (as Anthropic and OpenAI do) is deceptive.

  14. Assessment: UnresolvedSaid

    “Despite everybody talking about it endlessly for three god damn years, LLMs lack the kind of obvious, replicable, industrially-necessary outcomes that make a 3x, 4x or 10x price increase tenable.”

    What we tested: LLMs do not produce the clear, repeatable, industrially necessary results that would justify a 3x to 10x price increase.

  15. Assessment: Can’t be checkedSaid

    “I also think that Anthropic and OpenAI have deliberately used their subscriptions as a means of conning the media into conceptualizing AI as far more affordable than it actually is.”

    What we tested: Zitron thinks Anthropic and OpenAI have deliberately used their subscriptions to con the media into seeing AI as far more affordable than it actually is.

  16. Assessment: Can’t be checkedSaid

    “I think that these articles would’ve been much less positive if the reporters were even aware of the actual costs.”

    What we tested: Zitron thinks the positive articles about Claude Code would have been much less positive if the reporters had even been aware of the actual costs.

  17. Assessment: Can’t be checkedSaid

    “I know, that doesn’t mean anything, and all Block is doing is AI-washing, which is when a company uses AI as a scapegoat to justify laying people off.”

    What we tested: All Block is doing with its 4000 layoffs is AI-washing, meaning it is using AI as a scapegoat to justify laying people off.

  18. Assessment: UnresolvedSaid

    “AI is not actually driving much revenue at all!”

    What we tested: AI is not actually driving much revenue at all at software companies.

  19. Assessment: UnresolvedSaid

    “Even if you are impressed by what LLMs can do, remember that what you’re impressed by is the result of burning more money than anybody has ever burned on anything, including the Great Financial Crisis’ Troubled Asset Relief Plan (a little over $400 billion) and the COVID Paycheck Protection Program (somewhere between $800 billion and $900 billion).”

    What we tested: What LLMs can do is the result of burning more money than anybody has ever burned on anything, including TARP (a little over $400 billion) and the PPP ($800–900 billion).

  20. Assessment: UnresolvedSaid

    “This sounds sarcastic — petty even! — but it’s the truth, and nobody’s margins appear to be improving.”

    What we tested: No AI startup's margins appear to be improving.

  21. Assessment: ContradictedSaid

    “Absolutely none of these companies have a plan.”

    What we tested: No AI startup has a plan to become margin-positive.

    What happened: Replit explains that charges scale with its computing costs, rejects subsidized usage as its business approach, and describes further efficiency improvements. This is a concrete economic plan, although its success is not established by the announcement. (Replit, )

  22. Assessment: Can’t be checkedSaid

    “Four fucking years of this and we’re still talking about this stuff in riddles, mostly because it’s a terrible business.”

    What we tested: After four years, AI revenue at public software companies is still discussed in riddles, mostly because it's a terrible business.

  23. Assessment: Can’t be checkedSaid

    “It’s a deeply cynical con. Hyperscalers like Microsoft and Meta are paying for these contracts because they don’t reflect as assets on the balance sheet, all while moving the risk onto the AI compute company — and if the AI company misses a deadline, the hyperscaler can walk away.”

    What we tested: Hyperscalers sign neocloud compute contracts to keep the assets off their balance sheets and push the risk onto AI compute companies, keeping the right to walk away if deadlines slip.

  24. Assessment: UnresolvedSaid

    “In the end, I bet Stargate Abilene never gets built, but if it does, I’d be shocked if it’s done before the middle of 2027, which would mean it takes about 3 years per Gigawatt of power, or about a year per 293MW of IT load.”

    What we tested: Zitron bets that Stargate Abilene never gets built.

  25. Assessment: Can’t be checkedSaid

    “Every time you report on these deals should include a timeline.”

    What we tested: Every media report on a data center deal should include a realistic construction timeline.

  26. Assessment: UnresolvedSaid

    “The flippancy with which the media reports on these data centers — both in the structure of the deals and the realities of the construction (I go into detail about this in a premium from late last year, but making data centers is hard) — is allowing con artists to get rich and creating the conditions for yet another great financial crisis.”

    What we tested: Careless media coverage of data center deals is letting con artists get rich and setting up another great financial crisis.

  27. Assessment: Can’t be checkedSaid

    “The problem is that the markets have AI psychosis, and thus believe anything to do with data centers is a natural winner.”

    What we tested: Markets have 'AI psychosis' and therefore believe anything to do with data centers is a natural winner.

  28. Assessment: UnresolvedSaid

    “2026 or 2027’s great financial crisis will replace “homes” with “data centers,” and I worry it’ll be calamitous for the pensions and insurance funds that have tied their futures to AI.”

    What we tested: A great financial crisis centered on data center debt will happen in 2026 or 2027 and could be calamitous for pension and insurance funds.

  29. Assessment: Can’t be checkedSaid

    “Nevertheless, AI has become one of the largest cons in history, bought and sold based on stuff it can’t do (but might do, one day, at a non-specific time), constantly ignoring the blatant swindles and acidic economics that are only made possible with regulators and the media and the markets piloted by people that don’t know or want to know what’s actually happening.”

    What we tested: AI is one of history's largest cons, sold on things it cannot do.

  30. Assessment: Can’t be checkedSaid

    “If you’re excited based on the potential, you’re not excited about technology, you’re excited about marketing.”

    What we tested: Excitement about AI's potential rather than its current abilities is excitement about marketing, not technology.

  31. Assessment: ContradictedSaid

    “Know that LLMs cannot do many of the things that people talk about — they do not blackmail people, no GPT-4 did not trick a taskrabbit, and every single time an AI CEO says AI “will” do something you should spit in their fucking face for making shit up not print it without a second’s thought.”

    What we tested: LLMs cannot do many of the things attributed to them; for example, they do not blackmail people and GPT-4 did not trick a TaskRabbit worker.

    What happened: The report attributes this false explanation to GPT-4 after the worker questioned whether it was a robot, and records that the human supplied the results. This contradicts the denial of behavioral deception in the reported interaction. (OpenAI, )

  32. Assessment: Can’t be checkedSaid

    “However you may feel about Amazon as a service, its engineers are likely indicative of corporate engineering on some level, which is making me wonder if we’re not going to have some real problems in software development in the next few years as a result.”

    What we tested: AI-assisted coding may cause real problems across software development in the next few years.

  33. Assessment: UnresolvedSaid

    “To be explicit, allowing LLM to write all of your code means that you are no longer developing code, nor are you learning how to develop code, nor are you going to become a better software engineer as a result.”

    What we tested: If you allow an LLM to write all of your code, you are no longer developing code, nor learning how to, nor going to become a better software engineer as a result.

Held from this post (30)

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