Ed Zitron, a 16-year tech industry veteran and critic, argues that generative AI is fundamentally a con: companies overstate capabilities, financials, and autonomy while misleading investors, journalists, governments, and the public through the largest non-consensual technology push in history.
The Core Thesis: Generative AI as a Con
Generative AI is sold as magic — autonomous, job-replacing, cancer-curing — but functions as expensive, unreliable cloud software that cannot operate profitably without massive subsidies from Big Tech.
The six leading AI companies (OpenAI, Anthropic, Nvidia, Microsoft, Google, Amazon) derive 70% of AI revenues from OpenAI and Anthropic, both unprofitable and dependent on funding from the same hyperscalers reporting their growth.
Hyperscalers refuse to disclose actual AI revenues, using undefined “annualized run rates” that can mean anything (month × 12, 4 weeks × 13, last 4 weeks × 13) and change each quarter.
When companies have good news they announce it; silence on AI revenue speaks volumes — Microsoft’s FY2026 AI revenue was ~$10B excluding OpenAI pass-through, against $115B capex (rising to $175B).
The Financial Reality: Circular Funding and Massive Losses
OpenAI lost $20.9B in 2025; Anthropic and OpenAI raised $217B in H1 2026 alone, mostly from SoftBank, Nvidia, Amazon, Microsoft — a circular flow where funders are also customers.
Nvidia sold $215.9B in GPUs in its last fiscal year, largely to support ~$22B total global AI revenue outside the two anchor labs.
Inference providers and GPU lessors (e.g., CoreWeave) also appear unprofitable; Nvidia financed CoreWeave’s GPU purchases, then signed a $1.3B contract to rent them back so CoreWeave could secure debt.
Private credit (Blackstone, BlackRock) funds data centers because they know AI labs drive demand — not because of independent customer pull.
Unit economics are inverted: a $200/mo ChatGPT Pro user can burn $14,000 in tokens; $20/mo users burn ~$400. Enterprises balked when asked to pay per-token — Uber exhausted its annual token budget in three months.
The Adoption Myth: Forced, Subsidized, Non-Consensual
Adoption is not organic: Google forces Gemini into Search/Docs, Microsoft pushes Copilot into Office, Amazon inserts Rufus into shopping — the largest non-consensual tech push ever.
Media hysteria (“AI will take your job”) pressures individuals and employees to use tools; “AI-washing” resumes becomes a professional survival tactic.
88% of orgs report “regular use” but mostly for subsidized, low-stakes tasks (search, summarization, boilerplate) — not transformative workflows.
Real cost per million tokens (input + output + reasoning) is hidden behind flat subscriptions; if priced honestly, demand would collapse.
Hallucination rates on simple summarization have dropped from ~22% to ~0.7% on benchmarks, but benchmarks are rigged for LLMs and don’t reflect complex, high-stakes tasks (financial models, medical transcription, security-critical code).
Errors compound: a coding agent that breaks a codebase while “vibe coding” creates multiplicative debugging work for engineers who may have atrophied skills.
LLMs don’t learn like humans — they access context files (claw.md) but lack memory, mood, situational awareness, or genuine understanding; trust is built on shared experience and accountability, not file retrieval.
Video generation (Sora, Veo, Seedance) produces impressive clips but cannot replace the coordinated craft of filmmaking (lighting, continuity, direction, gaffers, ADs).
Training runs cost hundreds of millions and can fail (GPT-5 had a $500M failed run); progress requires ever-larger data and compute with no guarantee of breakthrough.
The Job Displacement Narrative: No Evidence
OpenAI’s own study found zero correlation between token spend and revenue per employee; Oxford Economics’ “AI killing entry-level jobs” claim rested on a single vague correlation line.
Displacement hits vulnerable contract labor (art directors, translators, transcribers) whose bosses already sought cheapest options — digital globalization, not AI novelty.
White-collar professionals (lawyers, accountants) report no productivity revolution; partners tout AI, associates (who do the work) don’t.
“Beat China” rhetoric ignores that China already has restricted GPUs (Blackwell) and functional LLMs; the race is a pretext for unlimited US capex.
Analysts (Kash Rangan, Jastrow) note China’s GPU access for years; the real winners are Nvidia and hyperscalers selling the arms race.
Cyber risks (AI agents finding/exploiting vulns) are real but stem from misconfigured sandboxes and massive compute brute-forcing — not emergent autonomy; regulation of compute access is the lever, not fear-mongering.
Comparison to Dot-Com Bubble: The “Rotcom” Theory
Dot-com had two bubbles: trash websites (Excite@Home buying BlueMountain for $1B) and dark fiber overbuild (demand doubling every 90 days vs. 6–12 months). Post-crash, fiber enabled real demand.
AI differs: demand is almost entirely subsidized; no independent revenue base; data centers are GPU-specific (1.2 GW Stargate Abilene = Bristol’s power in 1/1172nd the space), not general-purpose.
190 GW of data centers planned (Sightline Climate, Feb 2025) = $1.6–3T annual demand needed; actual AI revenue ~$130B — a 10–20x gap.
Hyperscalers’ core businesses (ads, cloud, e-commerce) grow via price hikes and ad-load increases, not AI; they buy GPUs to signal growth to markets (“kick the can”).
Tech CEO Rebuttals and the “Risk of Underinvesting”
Sundar Pichai: “Risk of underinvesting dramatically greater than overinvesting.”
Andy Jassy: “$200B capex not a hunch; investing to be meaningful leader.”
Mark Zuckerberg: “Rather build capacity early than be late” — Shrek/Farquaad energy (“some of you may die, but that’s a risk I’m willing to accept”).
None disclose AI revenue breakdowns; all use basis-point retention gains (Meta: 15 bps = 0.15%) to justify $10B+ spend.
The Coming Crash: 2027 Timeline and Cascading Effects
OpenAI plans $750B compute spend through 2030; delayed IPO (targeted $1T valuation) after audited financials showed $20.9B loss; Amazon’s $35B investment contingent on IPO.
Anthropic likely IPOs first, making OpenAI’s public debut near-impossible — SoftBank holds ~$100B paper value in OpenAI, reliant on liquidity events.
Oracle building 7.1 GW for OpenAI ($400B+); Oracle revenue flat 15 years (inflation-adjusted) — without OpenAI, Oracle faces existential risk.
Nvidia revenue could drop 50–70% (back to 2022 single-digit billions); S&P 500 ~7–8% concentrated in Mag 7 — retirement accounts face 20–40% haircuts.
Venture capital: >50% of 2024 dollars into AI; average TVPI 0.8–1.21x (losing money on paper gains); LLM startups (Cognition $2.6B valuation) have no acquirers but Musk.
Result: tech depression — hiring freezes, layoffs, margin compression, equity re-rating to “airline multiples” (mature, no growth premium).
Environmental and Social Costs
Gas turbines powering data centers (e.g., Musk’s Louisiana site) poison Black communities; behind-the-meter power bypasses grid regulation.
Water consumption, noise, and planning-board capture (Violent, NJ residents opposed; board approved) exemplify inequity: regular businesses denied loans, unprofitable Neoclouds get billions from Jensen Huang.
AI slop floods the web (SEO slop weaponized at scale); Google search degraded after Prabhakar Raghavan (ex-ads head) took over Search in 2020, relaxed spam suppression to juice query volume.
Personal Philosophy and Closing Thoughts
Zitron’s skepticism rooted in contempt for misleading power: “I don’t like being misled… ultra-rich, ultra-powerful people lie through their teeth.”
Community and relationships are the antidote: uplifting peers (editor Matt Hughes, critics Gary Marcus, Molly White, Brian Merchant), expressing appreciation to creators, staying human amid systemic gaslighting.
Financial advice (non-advice): be suspicious of tech promises, trust actions over narratives, conserve cash, treat markets as a casino pumped by media.
Final message: reach out to people you love, tell them they matter — “we don’t do this enough and we need to do it more.”
