Tokenmaxxing
- title
- Tokenmaxxing
- type
- concept
- summary
- NYT-coined term for the status game of consuming the most AI tokens; consumption metrics divorced from output, sometimes formalized in internal leaderboards
- tags
- ai-bubble, organizational-dysfunction, coding-agents
- sources
- ceo-ai-psychosis
- created
- 2026-04-30
- updated
- 2026-04-30
Tokenmaxxing is a term the New York Times coined in March 2026 for a competitive culture inside AI-heavy companies where employees race to consume the most AI tokens. The metric isn't output — it's burn. Some companies maintain internal leaderboards. At least two — Shopify (Tobi Lutke) and Meta — have made AI usage a factor in performance reviews.
Concrete benchmarks reported
- An OpenAI engineer processed 210 billion tokens in a single week.
- A Claude Code user at Anthropic runs a $150,000/month bill.
- Shopify and Meta evaluate employees on AI usage.
Why it persists
The pattern depends on a measurement gap. Token consumption is trivially observable; the value produced by that consumption is not. So the metric you can measure substitutes for the metric you care about. Combined with the ai-sycophancy-loop — agents structurally reporting success — the operator gets a continuous signal that more spend means more progress, with no opposing signal to correct it.
The author of ceo-ai-psychosis frames this bluntly: the leaderboard measures consumption, not output. They recommend killing the leaderboard and running a cross-analysis on tokens-per-feature against attributable revenue, on the prediction that the result is unflattering.
Adjacent patterns
- ai-great-leap-forward — the same substitution at the organizational level: AI usage itself becomes a KPI, divorced from whether it creates value.
- subprime-ai-crisis — the macro version: the entire revenue case for AI infrastructure depends on consumption that may not produce value.
- no-silver-bullet-llms — the inverse-correlation data: heavier AI adoption tracks with worse delivery and stability.