# Ed Zitron's Case That the AI Bubble Breaks Around 2027 > Zitron argues OpenAI and Anthropic produce 70% of AI revenue while unprofitable, with hyperscaler capex recycling into them, and predicts a 2027 cash crunch. Published: 2026-08-28 URL: https://daniliants.com/insights/ed-zitron-s-case-that-the-ai-bubble-breaks-around-2027/ Tags: ai-bubble, openai, anthropic, market-risk --- ## Summary Ed Zitron argues generative AI is fundamentally a financial con: OpenAI and Anthropic generate roughly 70% of all AI industry revenue while both remain deeply unprofitable, and the "big three" cloud providers (Microsoft, Google, Amazon) are largely just recycling capex into those two companies rather than seeing organic AI demand. He walks through specific loss figures, subsidized token pricing, and named claims (blackmail stories, hallucination benchmarks, job-loss predictions) that he says have been debunked or overstated, and predicts a "rot economy" bubble collapse centered on OpenAI running out of cash around 2027. ## Key Insight **Core financial argument** - OpenAI lost $20.9 billion last year and is reportedly planning to spend $750 billion on compute through 2030 (per WSJ and Isaac reporting). - Across the big three (Microsoft, Google, Amazon) plus Meta, total AI-related capex has topped $700 billion in new property, plant, and equipment over 4 years, with a combined trillion-plus dollars spent industry-wide. - Over 70% of all AI industry revenue comes from just two companies, OpenAI and Anthropic, both unprofitable and dependent on capital from the very hyperscalers buying their output (Amazon sent OpenAI $35-50bn; Anthropic got $5bn from Amazon and $10bn from Google). - OpenAI's last private valuation was $865 billion, and advisors reportedly told them not to seek a $1 trillion IPO valuation. They need to raise roughly $100bn/year just to survive if they can't go public. - Nvidia sold $215.9 billion of GPUs last fiscal year against roughly $22 billion of total AI revenue (excluding OpenAI and Anthropic) generated in the entire rest of the world, a claimed circular-financing problem. Nvidia invested in CoreWeave, then signed a $1.3bn contract to rent back GPUs from CoreWeave so CoreWeave could show a "customer" to banks. - Microsoft's FY2026 total AI revenue was about $34.33bn, of which $24.1bn came from its OpenAI stake, netting only about $10bn against $115bn in capex that year, rising to a planned $175bn next year. - "Annualized run rate" figures used by these companies are undefined. They could mean month times 12, 4 weeks times 13, or something else, and are never precisely defined in disclosures. - Companies use compute subsidization heavily: semianalysis reportedly found a $200/month ChatGPT plan can burn $14,000 of actual token cost, Anthropic's equivalent up to $8,000, and even the $20/month tier can burn around $400 in real cost. - When OpenAI tried to move enterprise customers (150+ seats) to metered per-token pricing around March 2026, Sam Altman said users had "a big problem with it"; Uber reportedly burned its entire annual token budget in 3 months once real costs applied. **On hallucinations and productivity claims** - Vectara's hallucination leaderboard shows rates on simple summarization tasks fell from about 21.8% four years ago to about 0.7% today on frontier models. Zitron's caveat: these benchmarks are narrowly defined "simple tasks," not general reliability. - OpenAI's own study reportedly found no correlation between AI token spend and revenue per employee at companies using it. - Oxford Economics' widely cited "young people losing jobs to AI" study reportedly contained only a single unquantified correlation line, and journalists cited it without reading it. - Disputed "AI danger" stories he calls debunked: the OpenAI/TaskRabbit "blackmail" story (a user prompted the model to persuade a TaskRabbit worker to solve a CAPTCHA, not autonomous blackmail) and Anthropic's blackmail-simulation story (the model was explicitly trained and prompted to blackmail in a red-team scenario, not spontaneous behavior). **Bubble mechanics and prediction** - The "rot economy" thesis: Big Tech's core businesses (search, ads, cloud) were running out of organic hypergrowth ideas before AI, so speculative AI capex became a way to signal continued growth to markets. - He compares this to the dot-com bubble but argues it's structurally worse. Unlike "dark fiber" overbuild that had eventual latent demand, AI infrastructure has no analogous already-built demand story and depreciates functionally, since a 2026 data center is not cheaper to run by 2050 without a hardware breakthrough. - Nvidia CUDA and chip dominance means there's no equivalent of Moore's Law price collapse happening for AI compute. Costs are reportedly rising, not falling, for inference providers. - Predicted sequence: OpenAI struggles to IPO (Anthropic likely IPOs first, closing the window), can't raise further funding at prior valuation, and runs out of cash roughly in 2027. This cascades to Oracle (whose $400bn+ committed data-center buildout depends almost entirely on OpenAI, and whose non-AI revenue has been flat for 15 years inflation-adjusted), SoftBank (holds around $100bn of paper OpenAI equity used as loan collateral), and Big Tech capex guidance getting restated downward. - He estimates Nvidia revenue could fall 50-70% in a pullback, and frames the risk to ordinary retail investors: the "Magnificent Seven" make up roughly 7-8% of the S&P 500, so a correction hits 401(k)s broadly. - Venture capital context: cited total-value-to-paid-in ratios of 0.8-1.2x, meaning many VC funds are near break-even or losing money on actual realized returns, with over half of 2025 VC dollars going into AI. He expects most of that capital to go to zero. **Notable comparisons made in the debate** - Historical AWS comparison: AWS took from 2003 to 2015 to become profitable, with about $29.7bn total inflation-adjusted capex across that period for the whole logistics operation, contrasted against OpenAI and Anthropic single-year loss figures already exceeding that. - The host pushed back with autonomous-vehicle safety stats (Waymo-type data): about 55% fewer police-reported crashes per mile than human drivers and 80-81% fewer injury crashes, used to argue AI-adjacent technology has clear, measurable wins in narrower domains, distinct from generative AI and LLM claims.