The Man Who Calls BS On AI: AI Is The World’s Greatest SCAM, And They All Know It! | Ed Zitron

Ed Zitron's central charge is that generative AI is sold as something far more capable and economically mature than it is. He calls it a “con,” pointing to claims that AI will replace all jobs, cure cancer, and transform work, while describing the products as costly, unreliable cloud software. The host challenges the strength of that label, noting that he has not previously heard a guest express the view so directly. Zitron grounds his criticism in sixteen years in the technology industry and says his objection is not to technology itself but to being misled. His language is deliberately forceful; the interview presents his judgment, not a legal finding of fraud.

His financial argument is that the infrastructure boom is supported by a small number of unprofitable AI companies and a circular flow of money. Zitron says OpenAI lost $20.9 billion the previous year, that OpenAI and Anthropic account for 70% of AI revenue among major technology companies, and that cloud providers are investing in those same firms as customers. He cites large funding and spending commitments, while arguing that technology companies do not disclose AI revenue clearly and use undefined “annualized run rate” figures. These are specific claims made in the interview; the segment does not provide underlying filings or explain the revenue calculation, so they should be read as Zitron's account rather than independently verified figures.

To make the spending concrete, Zitron describes a planned OpenAI–Oracle data center in Abilene, Texas: 1.2 gigawatts across eight buildings, each expected to contain 50,000 NVIDIA GB200 GPUs. He compares its power demand and footprint with Bristol, arguing that an extraordinary amount of energy, capital, and labor is being concentrated in one location. He says the major firms have spent more than a trillion dollars on capital expenditure and want to spend another trillion the following year, despite the revenue he attributes largely to OpenAI and Anthropic. The host raises the counterargument that adoption is unprecedented and that revenue may lag behind investment. Zitron disputes whether widespread use reflects voluntary demand, given that AI features are inserted into search, office software, and shopping platforms.

Zitron argues that many users treat generative AI as a search tool because they have been repeatedly told they need it, not because it has already demonstrated trillion-dollar value. He says it can be useful for broad queries but unreliable on specifics, and calls the investment scale unprecedented. He also separates GPU data centers for generative AI from ordinary CPU-based infrastructure and earlier high-performance computing uses such as protein research or autonomous systems. The distinction is important to his case: he is not claiming that every form of AI is useless, but that the current buildout is being justified by generative AI demand. The host's claim that AI adoption is broad remains a live counterpoint, while the exchange does not establish how much use is voluntary or how much is driven by product bundling and workplace pressure.

The conversation turns from adoption to what users actually get. Zitron says AI-generated content has not created the problem of low-quality web material from scratch; search optimization had already rewarded generic pages, and generative AI makes that kind of output easier to produce at scale. He also explains token billing as a meter for input and output, including some internal model activity, while noting that monthly subscriptions hide the underlying usage limits. This makes the visible subscription price a poor guide to the cost of serving heavy users. His explanation is a high-level description of token billing, not a complete comparison of every provider's pricing rules.

Zitron argues that flat-rate subscriptions obscure the cost of serving intensive users. He cites a SemiAnalysis estimate that a $200 monthly ChatGPT plan could consume roughly $14,000 worth of tokens, while a similarly priced Anthropic plan could consume about $8,000; even a $20 plan might incur about $400 in usage costs. He says companies subsidize that gap, which helps explain large reported losses and why enterprise customers object when asked to pay closer to actual usage. He gives Uber as an example of a company that exhausted its annual token budget in three months. These figures are claims presented in the interview, and the discussion does not show the underlying calculations or distinguish the cost of compute from subscription pricing policy.

The disagreement is not simply whether AI can be useful. Zitron acknowledges some value, including help with coding and technical support, but says companies do not disclose unit economics and that GPU capacity must be provisioned whether it is fully used or not. He argues that users pay even when a model hallucinates or breaks their code, while subscriptions conceal the relationship between usage and cost. The host asks whether the spending could be an investment ahead of future value; Zitron counters that costs have not fallen as expected and says inference providers and GPU-rental firms also appear unprofitable. The segment offers a critique of the business model, but not a full comparison of each company's revenue, costs, or long-term contracts.