Token-adjusted EBITDA and the return of nonsense metrics — are we in bubble territory?
Oct 9, 2025
Key Points
- OpenAI's $650 per-user valuation multiple sits below the $700 per-eyeball benchmark from the 1999 dot-com peak, suggesting AI valuations remain defensible by legacy bubble standards.
- Sell-side analysts quoting 'token-adjusted' metrics and mainstream media bubble profiles would signal late-stage speculation, but neither has emerged yet.
- Sora reached 1 million app downloads in under five days despite invite-only access, illustrating why AI valuations feel stretched despite tame multiples.
Summary
Nonsense metrics offer a window into bubble risk. One host invented 'token-adjusted EBITDA' by dividing OpenAI's $500 billion valuation by its 6 billion tokens generated per minute, yielding a 22x multiple. The exercise is deliberately absurd—tokens have wildly different values, profit streams, and production costs—but it frames a real question about whether the AI industry is entering bubble territory.
The comparison to the dot-com boom is instructive. In 1999, 'eyeballs' became the defining non-GAAP metric for tech valuations. Mary Meeker, the Morgan Stanley analyst who essentially invented the concept in a 1996 Internet trends report, watched valuations climb to $700 per monthly unique visitor, heavily skewed by Yahoo. By 1999, The New Yorker profiled her as 'The Woman in the Bubble.' She knew it was a bubble even as it happened. The stock market didn't peak until eleven months later. Bubble awareness doesn't stop bubbles. Animal spirits override individual analysis.
OpenAI's current metrics look tame by comparison. At 800 million weekly active users (actual accounts with behavior, not casual visitors) and a $500 billion valuation, the per-user multiple is $650, well below the $700 eyeballs benchmark from 1999. Using 6 billion weekly visits yields $83 per visit, an order of magnitude lower. Even the token framing doesn't trigger obvious alarm. At current wholesale pricing, $100 of equity buys an annual stream of 2.3 million tokens worth roughly $23. The math is admittedly absurd since token prices, volumes, and generation costs are all in flux.
Two warning signs deserve monitoring. First, whether sell-side analysts actually start quoting 'token-adjusted' metrics in research. That would signal late-stage bubble language. Second, whether analysts start pushing raw site visits as a valuation lever. One analyst was already doing this. Neither has happened yet. Legacy media hasn't written bubble profiles on AI the way The New Yorker did on dot-com. Social feeds are months ahead of headlines. But The New Yorker's eventual AI bubble piece may signal that mainstream awareness has caught up to speculation.
Sora downloaded 1 million times in under five days, faster than ChatGPT despite an invite-only rollout limited to North America. Growth metrics alone don't prove sustainability, but they illustrate why valuations feel stretched.