▚ STOUT STREET · PAPERSMMXXVI
WORKING NOTES · ESSAY II

The Meh Case for AI

Neither transformation nor catastrophe. A real technology, a financial unravelling, and a decade between the capital and the payoff.

John Francis · Working Notes · PDF · 17 pages · 5,625 words · 300 KB
$ read the pdf → ▚ save the-meh-case-for-ai.pdf
≈ $600B
gap: AI revenue vs capex
2.2%
of social value innovators capture
4.3%
AI share of 2025 online-fraud losses
≈ 40 yrs
dynamo → productivity statistics
[ abstract ]

What it argues.

The third story nobody argues for: AI is a real general-purpose technology whose benefits arrive roughly a decade after the capital that funded them. The gap in between produces a financial unravelling of some significance · telecom in 2001, railways in 1846 · the industry consolidates, and by the mid-2030s the world is meaningfully better off. Sixty-three quantified claims, a section on why crime adopts first, and a falsification list.

[ figures ]

The evidence, drawn.

fig. 01 · the point the buildout stopped being funded from cash flow
HYPERSCALER BOND ISSUANCE2021-24 average$28B / yr2025$121BBOND COVERAGE · ORDERS ÷ AMOUNT OFFERED5.0×Feb 2026< 2.0×Jul 2026demand fell while spreads held at their tightest since 1997
Issuance up 4.3× on the preceding average; demand depth collapsing while price holds. The essay's leading indicator is not rates · it is the market repricing risk before any central bank does.
fig. 02 · four percent of the economy, ninety-two percent of the growth
Share of US GDPinfo-processing equipment + software≈ 4%Growth share, H1 2025≈ 92%growth excluding the category: 0.1% annualised
Jason Furman's decomposition, as carried in the essay. An economy this dependent on one investment category for its measured growth has concentrated, not diversified, its exposure to the buildout's financing.
fig. 03 · discovery scaled; verification did not
380,000crystal structures predicted stable · DeepMind GNoME41synthesised by Berkeley's A-Lab, in 17 days≈ 430 years of verification at that paceautomation raised throughput and left the backlog untouched · the bottleneck did not move
The essay's deepest objection to the acceleration case, in one ratio. Pharmaceutical research automated its experiments once already · high-throughput screening, the 1990s · and development productivity kept falling on trend.
[ contents ]

Inside the paper.