In late 2025, four of the most sophisticated companies on earth told their engineers to use AI as much as possible. By the middle of 2026 those same companies were capping their token budgets in a panic. The reversal wasn’t due to a pricing surprise (although many private frontier models have been steadily increasing their token prices in preparation for their now delayed public IPOs) but rather a cognition failure in controlling spend.
Token Tsunamis: FOMO decision making in 4D
The spend mandate came first. Meta pushed AI tooling so hard it gamified consumption, running an internal leaderboard that staff nicknamed “Claudeonomics” which ranked engineers by usage. The result was a textbook incentive: employees burned an estimated 73.7 trillion tokens in roughly thirty days, and a memo to some 6,000 staff flagged an “exponential increase,” with internal AI costs on track to reach billions of dollars in 2026. One of the best examples of turkeys at the trough with trophies for overindulgence. Now Meta is dismantling the leaderboard and replacing it with a centralized “AI Gateway” to track and ration spend by 2027.
Uber ran the same playbook to the same end. Claude Code adoption across its roughly 5,000 engineers jumped from 32% to 84% in months; heavy users ran $500–$2,000 a month; the company burned through its entire 2026 AI budget by April. Amazon shut down its own internal usage leaderboard in late May 2026 after discovering employees were gaming it. Microsoft pulled back too, cancelling internal AI-coding licenses. The mandate to maximize and the order to cap arrived in the same fiscal year.
Fear of Missing Out infects even the smartest and skilled individuals and teams. Look at who owned these rollouts. They were driven by engineering, platform, and growth leadership, people whose entire professional reflex is adoption, velocity, and capability. More usage equaled more leverage right? Ship faster. The instinct is not wrong; it’s just one axis of thought.
You only turn consumption into a competitive game when nobody in the room is pricing the consumption. “Tokenmaxxing” wasn’t an accident, it was the designed outcome of an incentive built by people optimizing for a number that had no cost attached to it. The cognition that was missing wasn’t intelligence or data; every one of these companies could have forecast the bill. What was missing at the moment of the decision was the unit-economics reflex and the habit of asking “marginal cost per engineer, per token, at full adoption” before the leaderboard goes live, not after. It’s the same structural pattern that took down the OpenAI board: a leadership group clustered tight on one axis, with near-zero coverage on the axis that detonated.
Can you Prevent bad decisioning? Yep, Decision Simulation.
Run the tokenmaxxing mandate through a Decision Simulation. Load each decision-maker’s cognitive profile as a vector. The cluster lights up on capability and adoption and goes dark on cost-consequence forecasting. The stress test writes itself: at 84% adoption, with agentic tools that can burn a thousand times the tokens of a single prompt, who in this room is forecasting next quarter’s burn? If the answer is “no one, because no one around the table thinks in those terms,” the simulation flags it before the incentive is launched, not after a finance memo lands on a Sunday.
The gamification itself is the tell. A team that was modeling the maximum would never have incentivized it.
The Self-Fulfilling Prophecy lesson
This is not an argument against spending on AI, but rather intelligent spending given objective and holistic cognition. The failure was governance. And the swing now underway is from tokenmaxxing to what one might call token turtling, capping everything in a defensive crouch which may be applying the same blindness with the sign flipped: one reflex greedy, one scared, both from a room that still can’t price its own decisions. The waste was never the tokens. It was mandating their consumption from a room that could model capability but not consequence.
The fix isn’t to spend less on AI (although that is clearly coming soon); it’s to put the missing cognition in the room before the decision, so you price the call before the invoice does. Every one of these companies had the data to forecast the bill. None had the cognitive diversity at the table to ask for it and that gap, not the price per token, is what blew up the budget.