Companies learned to count AI tokens the way they learned to count cloud hours—right before realizing they'd been measuring the wrong thing all along.
The Summary
- Amazon killed its internal AI leaderboard after gamification drove token usage up and usefulness down. Uber burned through its 2026 AI coding budget in four months. Google's token usage grew 7x in a year.
- Enterprises are obsessing over tokens, GPU hours, and cost per inference while ignoring whether AI is actually generating revenue, accelerating decisions, or creating competitive advantage.
- The "tokenmaxxing" trend mirrors the cloud overspending disaster from the 2010s: perfect measurement of the wrong metric.
The Signal
Amazon's SVP telling employees "please don't use AI just for the sake of using AI" is the sound of a penny dropping. The company built an internal leaderboard to track token usage, likely thinking they'd gamify their way to AI adoption. Instead, they gamified bullshit generation. More AI-powered tasks, fewer useful results. When you measure the thing instead of the outcome, you get more of the thing.
This isn't an Amazon problem. It's an enterprise AI problem. The "tokens are the new oil" crowd built dashboards that show token burn rates, cost per inference, model utilization percentages. CFOs who got burned by surprise cloud bills in 2015 are now hyper-focused on AI spend. They know what each query costs down to the cent. What they don't know: whether any of it matters.
"They know what the intelligence costs, but not whether the intelligence is useful."
Uber's budget blowout tells the same story from a different angle. Four months to burn through a year's AI coding budget means either catastrophically bad forecasting or runaway usage without guardrails. Probably both. When you tell developers they have AI coding assistants and measure success by adoption rates, you get developers using AI coding assistants. Whether the code ships faster, breaks less, or solves harder problems—that's a different question entirely.
Google's 7x token growth in twelve months sounds impressive until you ask: 7x toward what? Revenue growth? User value? Product velocity? Or just 7x more tokens? Meta, Microsoft, and Salesforce are now scrambling to limit token usage, which suggests they've discovered the same thing Amazon did: measurement without strategy is just expensive accounting.
Key parallels to the cloud era:
- Track spend obsessively, ignore business outcomes
- Measure utilization instead of value creation
- Celebrate adoption metrics while profits stay flat
- Realize too late that efficiency ≠ effectiveness
The tokenmaxxing phenomenon—literally trending for weeks—is Silicon Valley's tell. When an industry coordinates around a metric this hard, it usually means nobody knows what actually matters yet. Tokens are legible, measurable, comparable across teams. Revenue attribution from AI? Decision speed improvements? Friction reduction in customer workflows? Those are hard. So everyone defaults to what's easy to count.
The Implication
If your AI strategy starts with token budgets, you've already lost. The right question isn't "how much AI are we using" but "what can we do now that we couldn't do before." Track token costs if you want. But track them as a denominator, not a numerator. Value created per token. Revenue per inference. Time saved per GPU hour. The enterprises that win the AI transition won't be the ones with the tightest token budgets. They'll be the ones who figured out how to measure whether the intelligence was worth buying in the first place.
Amazon's leaderboard shutdown is a gift. It's permission to stop measuring activity and start measuring results. Most companies won't take it.