The AI boom is threatened by an epidemic of unreliable numbers

Schumpeter (an opinion column) argues that investors in the AI boom are working from shaky numbers. Reported examples: the FT put OpenAI’s revenue near $50bn a week after Axios reported $70bn, and tech stocks swung; Broadcom’s Hock Tan likened markets to a casino. Lab “revenues” are really annualised-run-rate variants with different conventions: Anthropic counts sales through cloud providers gross, OpenAI net of partners’ cut. Anthropic reportedly told investors on 14 October it had $17bn of Q3 revenue and was profitable on its own adjusted measure; its ARR was $65bn in July. Coatue suggests valuing labs on “pre-training EBITDA”, which the columnist compares with WeWork’s “community-adjusted EBITDA”. Amazon, Google, Meta and Microsoft have $2.4trn of off-balance-sheet commitments, more than their bond debt. Goldman Sachs estimates about $1trn of annual AI revenue is needed for clouds to earn 10–20% returns. S&P 500 Q2 earnings growth was 50% unadjusted but 32% excluding Alphabet’s and Amazon’s write-ups of private stakes, and Epoch AI says US GDP growth was undercounted by 0.3pp because of how Nvidia’s cross-border sales are measured. The column blames underfunded statistical agencies and audit watchdogs; the Anthropic and OpenAI figures are reported, not audited.