Your Pension Bought Big Tech’s Debt
Five of the world’s richest companies owe $2.13 trillion. Their balance sheets say otherwise. Meta’s Hyperion data center borrowed $27.3 billion and booked $2.37 billion, parking the rest in Delaware shells named Beignet, after a New Orleans pastry. Off-book leases across Amazon, Microsoft, Google, Meta and Oracle: $831 billion. The buybacks are gone, $48 billion a quarter down to $4.6 billion.
Nobody got out of the bill. Meta pays the rent and eats $28 billion of downside; the bonds are rated a notch below Meta itself, because the market knows who signs the checks. A cleaner balance sheet isn’t less risk, just risk with better lighting. The buyers are insurers and pension funds, sold this as investment grade. Sellers call it small next to 2007 housing. Small to whom?
Wall Street didn’t make $2 trillion disappear. It moved the debt to a pension fund that was promised safety, then never let it vote.
Safety Was the Alibi
Anthropic CEO Dario Amodei published a 10,000-word essay urging frontier labs to pace model capabilities. Within days Altman, Musk and Hassabis endorsed it. The next day Altman said OpenAI would skip this year’s IPO, citing safety. The market has its own reading: DeepSeek’s V4 Flash is closing on Anthropic’s Opus 4.6, its V4.1 Flash has passed it, at a 1/25 cost, and those flagship models are what both labs sell most of.
Two labs staged a race to look responsible. The real pressure was price. Safety is a claim nobody can audit; a price is a number everyone can. DeepSeek matched the flagship for a fraction of the cost, so the valuation OpenAI was about to take public had one question it could not answer.
OpenAI delayed that number and called it caution. The badge was free. The invoice was not.
Math Wrote Rules the AI Labs Ignored
On September 11, twenty-five Fields medalists, Terence Tao among them, published an emergency statement: “A Severe Misalignment of AI in Mathematics.” Their target is not AI doing math. It is labs using unsolved problems as capability benchmarks, rewarding who solves first over what gets understood. It followed a run where OpenAI put about ten thousand agents and 130 billion tokens against open Millennium problems and claims it wins.
They wrote a rule. Nobody has to obey it. No compute, no gate, no penalty, no enforcement: a discipline that spends years certifying one result now watches ten landed in 88 hours, and its strongest move was a blog post. The labs’ incentive is the leaderboard, not understanding. No warning changes an incentive. Math is losing authority it never had a way to defend.
The mathematicians asked politely. AI kept training. A field that certified proofs for a century just learned that proof no longer requires its permission.
Humans Are Shedding
Armin Ronacher, who wrote Flask, handed GPT-6 Astra a weekend task: add virtual threads and lexical scoping to Python. Thirty-five hours later: 75,000 lines, 79 commits, a billion tokens, $1,200, and code that ran. None of it was written to be read. Semicolons fused statements, blank lines gone, indices no human can trace. One user named it machineslop. Ronacher’s verdict: worth nothing. The code did not care.
Every human role in software assumed someone could read the code. Review, QA, audit, the engineer woken at 3 a.m.: all of it dies with that one assumption. Agents already write for agents, and the moment they maintain each other, our eyes stop being an input. This is not the tool failing. People are being written out of the loop, quietly, by something that never asked to keep them. Ronacher’s judgment lands on a curve that no longer runs through him.
Software is learning to build itself, and we are the overhead it is shedding.