Ed Zitron: The AI Bubble is Bleeding Cash, Here Are The Receipts

Ed Zitron argues that the ongoing debate about AI return on investment in 2026 is itself evidence that AI is not a real industry with the total addressable market being sold to investors. Anthropic and OpenAI together account for approximately 89 percent of revenue among top AI companies, and the vast majority of other AI companies barely exceed $100 million in annual revenue, meaning two firms essentially represent the entire demand side of the market.

OpenAI's 2025 financials show $13.07 billion in revenue against total costs and expenses of approximately $34 billion, producing an operating loss of roughly $21 billion. Zitron dismisses reporting that true losses are closer to $8 billion after netting out stock-based compensation and cloud computing credits, calling it la-la land accountancy. He argues cloud credits from Microsoft are not a permanent operating condition and must be counted as real costs. OpenAI's sales and marketing expense of $5.73 billion exceeds Coca-Cola's annual advertising budget, yet OpenAI only launched its first major ad campaign in September 2025 and had roughly 500 salespeople whose total compensation would reach only about $125 million, suggesting a large portion of that line item represents inference costs or free credits rather than traditional advertising. The linear increase of sales and marketing costs alongside revenue is treated as a sign the business model does not scale.

OpenAI spent approximately $17 billion on Microsoft Azure, at least $1 billion renting H100s and H200s from Oracle, and paid CoreWeave a net $360 million. Google is reportedly renting capacity from CoreWeave to sell back to OpenAI. OpenAI holds approximately $6.4 billion in stock-based compensation and $1.2 billion in share-based compensation for compute provided by a related party, suggesting equity is being exchanged for compute. Zitron identifies several circular financing structures as evidence of systemic distortion: a roughly $35 billion arrangement in which a joint venture buys TPUs from Broadcom with Broadcom backstopping $30 billion and Google renting that compute to Anthropic; a $6.3 billion Nvidia backstop to CoreWeave whose revenue flows principally through Microsoft and Google back to OpenAI and Anthropic; and a $17 billion Nebius deal with Microsoft confirmed to be used for OpenAI. He estimates that when compute revenue from non-OpenAI, non-Anthropic, and non-Meta sources is stripped away, real remaining demand is roughly one to two billion dollars.

OpenAI and Anthropic together have approximately $1.1 trillion in compute commitments through 2030. OpenAI projects $284 billion in revenue by 2030 and Anthropic projects $174 billion by approximately 2029. Zitron argues those projections require revenues of roughly $15 to $20 billion per month within a couple of years, with no evidence or plan for how that happens. Anthropic's annualized run rate figures have been reported as $9 billion, then $14 billion, then approximately $47 billion, but Zitron notes the metric is calculated by multiplying that day's subscribers by 12 and the last four weeks of API spend by 13. He illustrates the manipulability by noting that one enterprise company accidentally spent $500 million in a single month on Claude by failing to set spend controls, which would extrapolate to $6.5 billion in annualized run rate under Anthropic's methodology.

Enterprise cost controls are tightening in ways that constrain the demand outlook. Uber burned through its entire AI budget in approximately three months and capped engineer usage at $1,500 per period. Zillow burned through its entire Cursor budget by the end of May, and its AI deployment increased human review hours by thousands per month rather than reducing headcount. Brex set limits of approximately $1,500 to $2,000 per engineer per period and about $5 per week for non-engineers. Zitron argues organizations are moving from no cost controls to active cost controls and those controls are likely to tighten further.

Zitron contends that hallucination in large language models is mathematically certain, making agentic reliability impossible, and that he cannot find evidence of a single dollar actually being spent autonomously by an AI agent despite multiple companies claiming this capability. He also argues that both OpenAI and Anthropic have recently pivoted to promoting recursive self-improvement because they are running out of original ideas, and that there is no proof recursive self-improvement will actually occur.

Zitron warns that by next year approximately 98 percent of all hyperscaler cash flows will go into capital expenditure, forcing them to take on debt. He argues this is structurally dangerous because the investment case for the Magnificent Seven has been premised on those companies being cash-rich and asset-light, and that S&P 500 investors have indirect exposure because lenders are extending credit backed by hyperscaler balance sheets. He draws a parallel to Lucent Technologies financing Winstar with $2 billion in a circular deal before Winstar ran out of money, and notes the real test of current hyperscaler commitments has not yet occurred because data centers are still being built. He predicts 2027 is the most likely year for the AI bubble to pop, with 2026 also possible, and that retail investors will be the primary victims when OpenAI and Anthropic are eventually taken public.