Remember when burning AI tokens was a badge of honor? When companies built leaderboards tracking who could use the most AI? That era just ended with a whimper and a lot of very large invoices.

Uber's COO Andrew Macdonald just revealed the company burned through its entire 2026 AI token budget in the first four months of the year. Microsoft cancelled Claude Code subscriptions across multiple product divisions. Salesforce is staring down a $300 million Anthropic bill for 2026. And Meta quietly took down the informal tokenmaxxing leaderboard its employees had created.

Welcome to the great tokenmaxxing collapse of 2026.

The Rise and Fall of Tokenmaxxing

Just a few weeks ago, tokenmaxxing seemed like the future of productivity measurement. The idea was simple: track employee token usage to identify your most innovative AI users. Tokens are the units of data AI models process—roughly 1.5 words of English text per token. More tokens burned meant more AI-powered productivity. Or so the theory went.

Meta, Amazon, OpenAI, and others established formal or informal leaderboards. Employees competed to see who could rack up the highest token counts. The practice spread fast across enterprise tech.

Then Goodhart's Law kicked in: any measure that becomes a target ceases to be a good measure. At Amazon, the Financial Times reported employees spinning up AI agents to complete wholly meaningless tasks just to keep their token stats high. Managers were using token counts for performance reviews, so people gamed the system.

The bills started arriving. And they weren't small.

The Sticker Shock Hits

Uber's burn rate was the warning shot. In a podcast interview last week, COO Andrew Macdonald said the ride-hailing giant blew through its entire 2026 token budget in just four months. Heavy Claude Code usage was part of the problem. But the bigger issue? "If you're not actually able to draw a direct line to how much useful features and functionality you're shipping to your users," Macdonald explained, "it's harder to justify."

Translation: all those tokens weren't producing measurable ROI.

Microsoft took action. According to The Verge, the company cancelled Claude Code subscriptions for employees in several key product divisions. The reason? Cost control. When your employees are burning tokens without clear business impact, the CFO eventually pulls the plug.

Salesforce CEO Marc Benioff put a number on it: his company's Anthropic bill will hit $300 million this year. In interviews, Benioff said he wished there were a "smart router" that could determine which queries actually needed expensive frontier models versus cheaper alternatives. That wish list item tells you everything about where enterprise AI economics are heading.

Meta joined the retreat. The company took down its employee-created tokenmaxxing leaderboard. The message was clear: we're no longer celebrating high token counts as a proxy for innovation.

Why AI Spend Isn't Producing ROI

Here's the uncomfortable truth most enterprises are discovering: AI spending and business outcomes aren't correlating the way anyone expected.

Recent data shows 64% of companies say AI is enabling innovation, but only 39% report EBIT impact at the enterprise level. Productivity gains are visible at the individual level. But they're not translating into firm-wide financial returns.

Why not? Azeem Azhar, author of the Exponential View newsletter, argues we're experiencing a classic "productivity J-curve" you'd expect with any general purpose technology.

When electricity first arrived in factories, companies started by replacing gas lighting with electric bulbs. Cost savings, sure, but no productivity boost. Then they replaced steam engines with large electric motors but kept the old drive-shaft factory layouts. Still minimal gains.

It was only when companies redesigned entire factory floors around individual electric motors—fundamentally rethinking their production processes—that productivity exploded.

Most enterprises are stuck in stage one or two with AI. They're using it to replace existing tasks (like electric lights replacing gas). A few are experimenting with workflow changes (electric motors with old layouts). Almost no one is at stage three: fundamentally rethinking business processes and business models around AI capabilities.

The Hidden Infrastructure Costs

Here's what caught most enterprises off guard: AI doesn't behave like software. It behaves like infrastructure.

When you deploy traditional SaaS, costs are relatively predictable. You pay per seat, maybe per feature tier. Usage scales linearly. With AI, costs scale exponentially as adoption spreads across departments.

Every prompt consumes tokens. Every document upload consumes tokens. Every code request, report summary, customer support chat—all tokens. For a single employee, the cost is negligible. For 10,000 employees using AI tools throughout the day? The math gets ugly fast.

And token costs are just the beginning. At scale, you're also managing:

• Cloud infrastructure and GPU capacity
• Data storage and bandwidth
• Cybersecurity and compliance frameworks
• Integration layers across systems
• Governance and monitoring tools
• Training and change management

A VP of Engineering I spoke with recently called it "the infrastructure trap." He said, "We thought we were buying productivity tools. Turns out we're running a compute-intensive infrastructure platform that never shuts off."

The Cost Control Playbook Emerges

Smart enterprises are shifting strategy. Here's what the new playbook looks like:

1. Kill the leaderboards. Stop incentivizing token usage as a proxy for innovation. It's Goodhart's Law in action.

2. Implement smart routing. Don't use GPT-4 for queries a smaller model can handle. Route simple tasks to cheaper models, complex reasoning to frontier models.

3. Set department-level budgets. Give teams token allocations and make them justify increases with measurable business impact.

4. Audit for waste. Find the meaningless AI tasks employees are running to game metrics. Cut them.

5. Measure business outcomes, not token counts. Track revenue impact, cost savings, customer satisfaction—not how many tokens you burned.

6. Redesign workflows, don't just augment them. This is the hard part. Most companies aren't doing it yet. But this is where the real ROI lives.

The Bigger Strategic Question

Uber's COO framed the challenge perfectly: Can you draw a direct line from AI spending to useful features and functionality for your users?

If the answer is no, you're burning money on innovation theater. If the answer is yes, you can justify the spend and potentially increase it.

This is forcing a healthy recalibration. Tokenmaxxing was always a flawed metric. It measured activity, not impact. It rewarded usage, not value creation. The pendulum is swinging back toward disciplined, ROI-driven AI adoption.

What This Means for Your Organization

For CTOs and Engineering Leaders: Stop celebrating token usage. Start measuring whether AI is helping you ship better features faster. If your teams can't connect AI spend to product velocity or quality, something's broken.

For CFOs: AI isn't a cost center you can cap the way you would SaaS licenses. It's infrastructure that scales with usage. You need visibility into where tokens are being consumed and whether that consumption is producing business value. Budget for it like you'd budget for cloud infrastructure, not software licenses.

For Chief AI Officers: The tokenmaxxing era was useful for experimentation. It got people using AI. Now it's time for discipline. You need usage policies, cost controls, and most importantly, a framework for connecting AI investments to business outcomes.

The Silver Lining

Here's the good news hiding in all this bad news: the market is learning. Companies are moving from blind experimentation to disciplined deployment. That's healthy. That's necessary.

AI costs will come down over time. Models are getting more efficient. Open-source alternatives are improving. Edge deployment is becoming viable for more use cases. The economics will improve.

But right now, we're in the awkward middle phase where AI is powerful enough to be useful but expensive enough to require serious financial justification. The companies that figure out how to extract real business value from AI during this phase will have a significant advantage when the costs eventually drop.

Tokenmaxxing is dead. Good riddance. What comes next is harder but more valuable: proving AI actually delivers the business outcomes we've been promising all along.

The bills are here. Now show us the ROI.