AI Bubble: How AI's push towards IPOs became a death drive | Ed Zitron

The video argues that the AI boom is economically unsustainable, driven by executives who cannot measure ROI and who treat AI as a status symbol rather than a productive investment. Companies face arbitrary token caps and "token austerity" because LLM usage costs as much as headcount while producing no clear value, making the industry a "financial Afghanistan" with no viable exit. OpenAI and Anthropic's growth depends on customers spending without justification, so they are shifting enterprise clients from flat subscriptions to token-based billing, exposing a pricing model that was subsidized from the start and leaving users with "invisible output" that fails to justify the cost. With Sam Altman admitting that model costs are "a huge issue," the narrative of infinite growth contradicts the reality of price ceilings and oversupplied data centers, revealing a bubble built on executive incompetence and cargo-cult adoption.

• ▶ 1:23 The core problem is that companies cannot measure the cost of a "unit of work" across AI tools like Claude Code and Copilot CLI, making ROI impossible to quantify; Uber's COO said they couldn't track LLM usage to any actual use, leaving AI as "something without ROI."

• ▶ 2:09 Companies are imposing arbitrary token caps—T-Mobile at $2K, Uber at ~$1,500/month—and these will only get tighter; Brex also capped Codex around $2K, while OpenAI's free months of Codex mean many firms are now hitting "token austerity" for the first time.

• ▶ 2:49 Adoption is "cargo cult stuff": companies use AI because investors reward them for it, but no one knows actual costs or whether limits like $1,500 are good or bad; once embedded, AI can't simply be removed—like a tick, "it'll leave bits in there," and has already made codebases like Zillow's worse.

• ▶ 5:20 LLM spending will inevitably be pulled back because companies are spending as much on tokens as on headcount, with no ROI or faster shipping to justify it.
• ▶ 5:50 Lines of code are a terrible success metric—writing more code means less control, less stability, and a more cumbersome system.
• ▶ 6:20 Companies are built on sand: LLM-generated codebases and weakened, LLM-dependent engineers have eroded institutional knowledge, and LLMs create more future problems than they solve.
• ▶ 7:41 Cost reduction is impossible to approach rationally because there is no way to measure ROI or the cost of a task.
• ▶ 7:53 OpenAI and Anthropic's core problem: growth depends on customers spending without justification, creating unsustainable economics like "pay pigs of Anthropic."
• ▶ 8:45 Customers pay for LLM outputs whether good or bad, including wasted tokens from loops and errors; one suggested cost cut is cheaper models like DeepSeek.
• ▶ 9:20 AI lab costs have dropped sharply, with a 75% reduction noted, setting the stage for the dysfunction described.
• ▶ 9:26 American AI labs are described as a "financial Afghanistan" — a costly quagmire with no clear victory or exit, which the speaker calls "remarkable" and "weird."
• ▶ 9:43 Anthropic's growth is driven by a "token firehose" culture of "use as many tokens as possible" — revenue fueled by massive, inefficient token consumption rather than sustainable economics.
• ▶ 9:51 Zitron introduces the "business idiot" theory: executives running companies do no actual work and have no connection to product, productivity, or production; their roles are limited to hiring, firing, spending, lunches, and emails.
• ▶ 10:09 Because they're detached, executives reason that "more AI is good" based only on casual toy experiments, extrapolating that more AI use by engineers means "even more big good" — with no nuance, thought, or planning.
• ▶ 10:24 He concludes the supposed AI boom is driven by executive incompetence, not sound strategy or real engineering gains, as leaders mistake their shallow experiences for evidence that AI is universally powerful.
• ▶ 10:34 Sam Altman is now on record agreeing with Ed Zitron's long-standing criticism about AI costs.
• ▶ 10:40 Altman openly admits that the cost of running these models is "a huge issue."
• ▶ 10:49 The speaker asks whether this admission from the CEO of a leading AI lab should be setting off alarm bells about AI economics.
• ▶ 10:50 OpenAI and Anthropic are pursuing IPOs with massive revenue targets—$284B by 2030 and $174B in 2029—making a slowdown impossible.
• ▶ 11:01 There's a direct contradiction: Altman can’t tell investors about huge future revenues while also admitting customers are “freaking out” about costs.
• ▶ 11:22 The growth narrative demands constant acceleration; neither company can afford to slow down or signal weakness about cost pressures.
• ▶ 11:37 OpenAI leadership is "really freaking out" about costs, signaling an industry-wide price ceiling that may make the AI boom economically unsustainable.
• ▶ 11:43 A massive oversupply of data centers is being built, but saturating that compute would require three or four more OpenAI/Anthropic-scale spenders—which don't exist; hyperscaler spending mainly serves those same two customers.
• ▶ 12:11 The boom cannot slow down: companies must reach hundreds of billions in revenue, and OpenAI's shift to token-based billing with Anthropic ▶ 12:26 shows they are passing unsustainable compute costs onto enterprise customers.
• ▶ 12:26 OpenAI and Anthropic have moved enterprise customers to token-based billing, replacing flat-fee subscriptions.
• ▶ 12:31 This shift ends subsidized subscription pricing, making customers pay the true cost of token usage.
• ▶ 12:40 The change is happening quickly and suspiciously aligns with when companies face the real costs of AI, indicating unsustainable economics.
• ▶ 12:47 The AI industry "conned everyone" with subscriptions that let users burn far more in tokens than plans realistically priced—e.g., a $200/month plan could consume ~$5,000 worth of tokens, making the pricing model unsustainable from the start.
• ▶ 12:59 Journalists have been reviewing AI using a subsidized, fantasy product whose current form "will not exist," so positive coverage is based on unrealistic pricing conditions users won't actually get.
• ▶ 13:10 The coming price shift is like a flat-fare cab suddenly switching to per-mile pricing—but far worse: not Uber going from $10 to $50, but jumping to $500 or $5,000; users were trained to treat AI as all-you-can-eat and are now hitting the wall with GitHub Copilot, where "you can't do anything" once real costs apply.
• ▶ 14:28 The core issue shifts to what AI's massive costs actually produce, framed as an input/output question: "What's coming out the other end?"
• ▶ 14:41 The term "invisible output" is introduced, highlighting that industry watchers see AI's results as intangible and hard to measure despite huge investment.
• ▶ 14:46 The interviewer asks whether there is a "trick" to quantifying AI's real output, raising the possibility that value exists but is being measured incorrectly.
• ▶ 15:00 Zitron argues that AI measurement frameworks like SemiAnalysis's "dark output" are "cope" from a sick capitalism, not legitimate value metrics.
• ▶ 15:28 AI benchmarks are rigged and models are "coddled" because they can't use tools like humans; their performance is contrived.
• ▶ 16:21 The constant need to invent "mystical metrics" for AI value—plus claims like Anthropic's "8x more code"—are red flags, not evidence of real usefulness.
• ▶ 16:48 Real value is self-evident—truly useful products don't require persuasion, and people recognize them instantly.
• ▶ 16:57 Zitron's firsthand iPhone launch experience showed that people reacted with immediate "wow" because benefits like visual voicemail, app-style texting, and email on a touchscreen were instantly demonstrable.
• ▶ 17:25 The core realization was "the internet in your pocket," an intuitive value that needed no pitch and only grew with each model.
• ▶ 17:29 The iPhone 3G generated genuine excitement because there were very tangible reasons to be excited about mobile.
• ▶ 17:37 Cloud computing (AWS) offered clear practical benefits compared to the old burden of running your own Sun Microsystems server.
• ▶ 17:44 With AWS, simply creating a storage bucket and projecting services to the world made the value obvious and easy to grasp.
• ▶ 17:46 LLMs lack the clear, "obvious value" of past tech like cloud services, and while they might let you ship more code, nobody is shipping good software with them.
• ▶ 18:07 Software has become objectively worse across the board, with AI seemingly reducing productivity—evidenced by regressions like BlueSky's inability to crop images.
• ▶ 18:34 Concrete harms include AWS outages from an AI coding tool and AI tools deleting databases, while Anthropic's Claude Code Work shipped in a week but deleted a user's photos—with no counterexample of stable, quality AI-driven products.
• ▶ 19:12 Zitron argues AI "doesn't know anything" and "breaks," producing frustrating results rather than genuine usefulness.
• ▶ 19:16 AI companies are pushing for special, bespoke metrics to make executives "feel so special," instead of being judged by ordinary standards of usefulness.
• ▶ 19:22 Zitron calls AI leaders like Dario Amodei and Sam Altman "rich kids" who were coddled by the press, never had to prove themselves, and now "haven't got anything" when forced to deliver.
• ▶ 19:42 The AI efficiency narrative is framed as the promised driver of economic change, but the host notes "pretty significant issues with youth unemployment across the West" are early evidence of shifts, questioning what actually causes it.
• ▶ 20:08 Ed Zitron attributes labor market problems to massive over-hiring in 2021–2022, broken mentorship structures, and middle managers acting as "cops" who claim credit rather than support workers, making employment for young people "absolute hell" since 2008.
• ▶ 22:26 Zitron argues there is "absolutely zero proof" that AI causes job losses; layoffs happen because bosses "don't understand anything and want to save money," and the AI narrative is a self-serving story pushed by companies and investors to suppress workers.
• ▶ 24:11 Anthropic reportedly becomes the first major AI giant to file for its IPO.
• ▶ 24:24 A significant amount of capital is expected to flow into AI IPOs from both "smart or dumb money."
• ▶ 24:28 Persistent concerns about costs remain unresolved as the IPO push accelerates.
• ▶ 24:34 Guest calls SpaceX "horrible" and "a big lossy mess" of several companies stapled together, arguing a functioning SEC wouldn't let it go public and might try to merge it with Tesla.
• ▶ 25:04 OpenAI and Anthropic are called "absolute dogs" and "awful money burners" — unsustainable, unprofitable, with their non-GAAP financials "deliberately manipulated" to look good; OpenAI is singled out as especially bad.
• ▶ 25:59 The guest says these companies have no path to profitability and shouldn't be allowed public, warning that listing them would be dangerous for retail investors and ordinary people exposed via indices.
• ▶ 26:17 The speaker questions why AI companies would choose to IPO now, given the fragile state of major global economies.
• ▶ 26:21 Global economies face severe supply chain disruptions, war in Iran, and other crises, and at ▶ 26:35 AI is called "absolutely no exception" to this fragility.
• ▶ 26:37 Helium shortages are highlighted as a concrete supply chain risk for AI hardware and data centers, making the current moment dangerous for IPOs.
• ▶ 26:57 IPOs are driven by "exit liquidity for investors"—early backers pushing new companies to go public so they can cash out.
• ▶ 27:14 OpenAI and Anthropic face a trillion-dollar valuation ceiling; to justify an IPO at that level, they'd need to grow to $3–5 trillion, making them bigger than TSMC/Meta and rivaling Nvidia/Google—a highly implausible outcome.
• ▶ 28:00 Anthropic's capital hunger is "insane": despite raising $75 billion in three months, they're going public for even more access to capital, effectively dumping onto public markets.
• ▶ 28:08 AI companies pursue IPOs primarily to gain legitimacy, which is a prerequisite for raising debt in public markets.
• ▶ 28:14 Despite needing debt, these companies can barely raise any due to enormous operating losses—sarcastically cited at $10–20 billion per year.
• ▶ 28:20 Credit markets respond with skepticism, questioning whether AI firms will exist in two years, run out of money, or be able to pay back lenders—creating a vicious financing cycle.
• ▶ 28:30 Lenders are increasingly questioning whether AI companies can actually repay their debts, reflecting deep uncertainty about the sector's creditworthiness.
• ▶ 28:36 A newly reported deal between Broadcom, Google, and Anthropic uses a complex structure: Google buys chips from Broadcom, an SPV buys them from Google, and the SPV leases them to Anthropic.
• ▶ 28:51 A major red flag: Anthropic is borrowing money to fund the arrangement, suggesting the AI industry is being propped up by debt with no clear guarantee of repayment.
• ▶ 28:55 Anthropic reportedly refused to share financial documents with prospective investors, a striking red flag for transparency.
• ▶ 29:06 Lender reactions split: some walked away, while others agreed to proceed without ever seeing Anthropic's financials.
• ▶ 29:09 The speaker harshly criticizes lenders who skip due diligence, arguing that those who lend without reviewing records should lose everything—and at ▶ 29:17 suggests it should perhaps be illegal.
• ▶ 29:18 Calls for making certain corporate debt and borrowing practices illegal, proposing regulation of lending and corporate borrowing.
• ▶ 29:26 Acknowledges the common counterargument that this would cut research and development.
• ▶ 29:30 Rejects that defense, pointing to the poor results of the current permissive approach to corporate capital.
• ▶ 29:32 The speaker challenges whether AI companies' actual fundamentals—products, earnings, or assets—can justify trillion-dollar valuations, concluding "It isn't. Not even close."
• ▶ 29:41 The real motivation for going public is self-enrichment: insiders and investors want to get rich and "socialize the losses."
• ▶ 29:44 The IPO is framed as a mechanism to offload mounting losses from private insiders onto public shareholders, transferring risk rather than marking a legitimate growth milestone.
• ▶ 29:50 Zitron doubts OpenAI can go public, predicting a WeWork-style collapse once its finances are exposed.
• ▶ 29:56 He suggests Anthropic will likely face the same financial scrutiny shortly after.
• ▶ 30:04 He compares OpenAI to WeWork’s failed IPO, expecting even greater losses and investor rejection.
• ▶ 30:25 If OpenAI or Anthropic fail to go public, it would effectively end the AI boom—"bedtime for this whole thing."
• ▶ 30:35 The market is souring and revenues are slowing, making it unclear how the current trajectory continues.
• ▶ 30:45 To meet their promises, both OpenAI and Anthropic would each need to become the size of Google in four years—a projection the speaker begins to question.
• ▶ 30:56 The speaker is skeptical of optimistic projections, saying "I don't think that that's very likely."
• ▶ 31:00 These companies have not found a viable business model.
• ▶ 31:04 They have no path to profitability or sustainability, lacking a clear economic foundation.
• ▶ 31:04 Ed Zitron issues a blunt verdict: certain AI companies have "no path to profitability" and are "dogs."
• ▶ 31:09 The host notes these IPOs are "almost certainly incoming now," and they agree to closely follow developments in the coming weeks and months.
• ▶ 31:23 The host introduces guest Ed Zitron, author of "Where's Your Ed At" and host of the Better Offline podcast, before wrapping up.

• ▶ 0:00 AI Adoption Pullback and Cost-Measurement Problems (0:00 - 5:26) - - Large enterprises are scaling back AI while experts struggle to measure LLM costs and ROI.
• ▶ 5:26 Costs, Hidden Engineering Damage, and AI Token Economics (5:26 - 12:45) - - Code quality and hidden damage compound the unstable business of AI token spending and Altman's cost warnings.
• ▶ 12:45 Token Billing, Subscription Cons, and AI's Unclear Value (12:45 - 19:39) - - Subscription models hide token costs, output is often invisible, and AI lacks the obvious value of earlier tech.
• ▶ 19:39 AI Efficiency, Youth Unemployment, and Layoff Realities (19:39 - 24:14) - - AI efficiency claims collide with a broken job market; layoffs are blamed on AI without proof.
• ▶ 24:14 IPO Risks, Unsustainable AI Giants, and Economic Pressures (24:14 - 31:33) - - Anthropic, OpenAI, and SpaceX face public-market problems amid broader economic and supply-chain risks.