The AI Reckoning: Anthropic's Trillion Dollar Bet and the Warnings Nobody Wants to Hear
Something strange happened this week in Silicon Valley. A company on the verge of one of the largest IPOs in history filed its paperwork on Monday, then published a report three days later suggesting maybe we should slow down. That company is Anthropic, and the tension between those two moves captures everything chaotic and terrifying about the AI moment we are living through.
The Pause Report That Landed Three Days After the IPO Filing
On June 1, Anthropic confidentially filed for an IPO after closing a $65 billion Series H at a $965 billion valuation. Three days later, the Anthropic Institute published When AI Builds Itself, co-authored by Marina Favaro and Jack Clark, warning that AI systems are approaching recursive self-improvement and calling for a coordinated global pause on frontier development.
The internal metrics they cited are striking: as of May 2026, more than 80 percent of code merged into Anthropic's own codebase was written by Claude, and engineers were shipping roughly eight times as much code per quarter as before 2025.
The catch is that Anthropic would only pause if other frontier labs did so under verifiable conditions. Everybody goes, or nobody goes. And critics were quick to note that a coordinated freeze landing right as Anthropic sits near the top of the competitive landscape is either extraordinary corporate conscience or exceptionally savvy positioning. Both things can be true.
AI Is Genuinely Doing New Things
Whatever you think of the pause proposal, the capability evidence is real. On May 20, OpenAI announced that an internal reasoning model had disproved the Erdős unit distance conjecture, an open problem in discrete geometry that had resisted mathematicians for nearly 80 years. The proof was verified by external mathematicians. Fields Medal winner Tim Gowers called it "a milestone in AI mathematics."
That is not a benchmark score. That is a model solving a problem that was never in any training data.
The Economic Trap Nobody Can Exit
Even if AI never escapes human oversight, a March 2026 paper by Brett Hemenway Falk (UPenn) and Gerry Tsoukalas (Boston University) titled The AI Layoff Trap argues the economic damage may arrive anyway, driven by perfectly rational behaviour.
The logic: workers are consumers. When a firm replaces employees with AI, it captures the full cost savings, but the lost spending is spread across every company selling anything. Each firm automates because the alternative is being undercut. Collectively, they destroy the demand they all depend on. The paper's conclusion: "At the limit, firms automate their way to boundless productivity and zero demand."
The authors argue the only fix is a Pigouvian automation tax, similar to how we tax pollution, making it costly enough to fire people that the math stops rewarding the race. UBI and retraining do not solve it in their model.
Real numbers give the model support: nearly 80,000 tech workers were laid off through April 2026, with roughly half attributed to AI, and business investment overtook consumer spending as the leading US GDP contributor in Q1 2026.
The Third Scenario: Expensive Disappointment
Not doom. Not dystopia. Just a very costly shrug.
MIT's Project NANDA reviewed over 300 enterprise AI deployments and found that despite $30 to $40 billion in collective spending, 95 percent of organisations reported no measurable return on investment. Only 5 percent extracted real value. Critics note the six-month ROI window is narrow, and the internet looked similarly barren on financial metrics for years. Fair point. But more apps are being released than ever on the iOS App Store while reviews and sustained usage are declining. More being built. Less being used.
Where This Leaves Us
Three forces pulling in different directions at once: Anthropic warning AI might escape human oversight while filing for a trillion-dollar IPO on the strength of it; economists proving mathematically that even a well-behaved AI rollout could hollow out consumer demand through individually rational decisions; and enterprise data suggesting most organisations cannot make AI pay off at all.
These outcomes are not mutually exclusive. The benefits may accrue to a small number of companies while the costs distribute broadly and quickly, which is roughly how every transformative technology in history has behaved.
Anthropic's report closes with a line worth sitting with: "The window to investigate the questions together is here. And people outside AI companies should be involved in this deliberation."
That is not a call that should be left to the companies doing the filing.