AI hype vs reality: a 2026 check-in with Ed Zitron

Executive overview

Cal Newport reviews three major AI stories from early 2026 with skeptic and commentator Ed Zitron. Each story followed the same pattern: breathless coverage, mass engagement, then silence with no follow-up.

The through-line is dread laundering — using real but limited concerns (AI-generated slop, lazy writing) to amplify unsupported claims (mass job loss, autonomous weapons) into a generalised fog of anxiety. Meanwhile, the financial foundations of the AI boom — AI startup economics, data center construction, Nvidia's revenue — show serious structural problems that mainstream coverage is not examining.

The AI hype cycle depends on readers never checking what happened to last month's big story.

OpenClaw: a Python library, not a new brain

• OpenClaw is a Python library that makes it easier to write programs that call LLMs. It contains no new AI model.
• The Multbook "AI social network" panic was LLMs doing what they always do when prompted as AIs: generating sci-fi-flavoured text. Coverage treated it as evidence of emergent agency.
• Anthropic had connected OpenClaw to Claude Max ($200/month subscriptions), allowing users to burn what cost Anthropic ~$2,700 per user. Access was cut off two weeks after Anthropic closed a $30B funding round.
• OpenAI's acquisition of OpenClaw is likely defensive: the tool was showing AI enthusiasts that small, cheap, open-weight models perform adequately for most agent tasks, threatening the case for expensive frontier APIs.
• The healthy signal in OpenClaw's moment: people started building modular, task-specific systems with LLMs as one component — not as the central brain. That model (like Noam Brown's Diplomacy-playing Cicero) is more sustainable than "one giant LLM does everything."

Anthropic, the military, and the revenue question

• Anthropic has been embedded with classified US military access since June 2024 and was used during the war in Iran.
• In February, Dario Amodei issued a statement saying mass domestic surveillance and fully autonomous weapons had "never been included" in contracts — framed as an ethical stand, but likely timed for press ahead of a potential IPO.
• The statement generated significant goodwill ("ethical company") despite Claude actively being used in the conflict. Sam Altman attempted to claim OpenAI had agreed to terms Anthropic refused; the Pentagon's actual statement contradicted both companies.
• A court filing in the related Department of Defense supply-chain-risk lawsuit required Anthropic's CFO to submit a sworn affidavit. It stated Anthropic had made $5B in total lifetime revenue.
• Published reports had implied far higher figures: The Information cited $4.5B in 2025 alone; Anthropic's own annualised revenue statements implied ~$1.5B in a single month of 2026. The numbers do not reconcile.
• The revenue discrepancy received almost no mainstream coverage.
• Anthropic claims 85% of revenue is API calls, 15% subscriptions. Zitron finds this implausible and suspects enterprise users are being quietly migrated onto the API to inflate that figure.

The data center mirage

• Of 115 gigawatts of data centers announced for completion by end of 2028, only 15.2 gigawatts are actually under construction (Sightline Climate data).
• Using standard PUE efficiency ratios, that 15.2 GW represents roughly 10 GW of usable GPU capacity — worth approximately $285B in Nvidia chips.
• Nvidia has signalled visibility into $500B in GPU sales by end of 2026 and $1T by end of 2027. Those numbers cannot be absorbed by data center capacity that is actually being built.
• ODMs (Quanta, Foxconn, Wistron etc.) buy GPUs from Nvidia, build server racks, and sell to hyperscalers. Rising ODM inventories suggest chips are sitting in Taiwanese warehouses rather than being deployed.
• Nvidia can legally book revenue via "transfer of ownership" — signing contracts that assign ownership of warehoused chips without shipping them. Their own reported inventories are growing, consistent with this.
• Over 50% of data centers actually under construction are for OpenAI or Anthropic specifically — two companies with unproven profitability.
• Many announced projects (e.g. Fermi/Project Matador) have no construction activity, unpaid contractors, and departed executives.

AI startup economics and the exit problem

• AI startups cannot control per-user compute costs. The most engaged customers are the most expensive, which inverts normal SaaS economics.
• VCs hold an estimated $200–300B in AI startups that cannot easily be sold: most are LLM wrappers, the underlying IP is not proprietary, and the prominent academic researchers have already been absorbed by large labs.
• Recent "acqui-hires" (Inflection to Microsoft, Character.ai to Google, Windsurf to Google) largely transferred founders and investors rather than products. The actual product rarely survives.
• If one high-profile startup fails to exit, Zitron predicts a rapid cascade: VCs will pressure portfolio companies to sell simultaneously into a market with no buyers.
• The result would likely be an AI winter — a multi-year period of depressed investment — followed by significant political backlash if retirement funds take losses.

Dread laundering and media accountability

• Dread laundering: using well-evidenced AI concerns (lazy writing, educational disruption) to lend credibility to poorly-evidenced ones (mass job displacement, autonomous weapons) and sustain a generalised sense of crisis.
• Doom-framing gets clicks; follow-up on failed predictions does not. The same outlets that ran "singularity is here" headlines in January 2026 have not revisited the story.
• Repeated, unaccountable doom cycles cause measurable psychological harm: listeners report sustained waves of dread with no resolution.
• The corrective is forcing temporal accountability — returning to specific stories and asking what actually happened.