Banks Just Found Out What AI Debt Is Virtually Worth: Zero

A 25-minute essay built on the line her channel keeps coming back to: the technology is real, and the way it is being run is not. The title promises worthless debt. The video itself puts the loans at about 90 cents on the dollar and tells viewers nobody should panic about their index funds.

It moves in three steps. The first is trust, told through three stories she presents as one pattern. Flock, the licence-plate camera company, is losing city contracts after reports of police misusing its system, and is offering staff buyouts. Employees of the big labs are resigning and signing an open letter. And a US strike on a school in Minab, Iran, in February was tied by a Pentagon review partly to stale intelligence and reliance on AI targeting. In each case, she argues, the system did what it was asked, and the failure was in deployment, speed and oversight.

The second is whether AI pays. She sets the encouraging studies (productivity gains of 15 to 50% in writing, support and coding work; Erik Brynjolfsson’s estimate that US productivity grew about 2.7% in 2025) against the discouraging ones (only a fifth of firms seeing revenue from it, MIT’s 95% of pilots with no measurable impact, METR’s finding that experienced developers were 19% slower with AI tools). Her reconciliation is Paul David’s 1990 essay on the electric dynamo: factories took decades to reorganise around a new power source before the gains showed. Small models handle most narrow tasks, so routing each request to the smallest model that works is good engineering, but it doesn’t make the large models redundant.

The third is the money, drawn as layers of borrowing. Bond buyers lend to SoftBank at 8.6 to 9.75%, SoftBank funds OpenAI, OpenAI commits to computing from Oracle and CoreWeave, who borrow to build, and Nvidia sometimes guarantees the value of the GPUs pledged as collateral, so the risk loops back to AI spending. Her signs that money is getting dearer: roughly $18 billion of loans for Oracle’s Project Jupiter campus in New Mexico quoted at 89 to 91 cents, AI-linked bonds paying about 115 basis points over Treasuries against 78 for the wider market, and some $420 billion of hyperscaler debt expected in 2027. She closes with a call for financial, product and safety regulation, and with the gap between the $1.2 trillion valuation investors offered OpenAI and the $1.5 trillion it wants, which she calls “the whole story”.

The financing figures check out against the reporting: the Jupiter loan quotes, SoftBank’s coupons and its $64.6 billion total in OpenAI, the spreads and Goldman Sachs’s $420 billion forecast, the $1.2 and $1.5 trillion reports, and OpenAI’s forecast of $856 billion on computing and $278 billion of cash burn from 2026 to 2030, for revenue of $350 billion in 2030. Two need correcting. OpenAI did not publish those projections: they come from a private presentation that the Financial Times reported. And the March round is reported at $852 billion after the money; the $730 billion she gives was the pre-money figure when the first $110 billion was announced. The employee letter (“Pacing the Frontier”, July) had about 1,100 signatures when it was published, where she says nearly 1,400, and it asks Washington to back an international effort to pace frontier development rather than for regulation in general. The Nvidia collateral guarantees, the link that closes her loop, are described with no named deal.