In February 2026, Uber built a leaderboard.
It ranked engineering teams by AI tool usage — how many tokens they consumed, how many sessions they ran, how much of their workflow passed through Claude Code and Cursor. The goal was to accelerate adoption. It worked.
By March, 84% of Uber's 5,000 engineers were classified as agentic coding users. The leaderboard had turned token consumption into a competitive sport. Uber's COO, Andrew Macdonald, would later coin a term for what happened: tokenmaxxing.
By April — four months into the fiscal year — Uber had burned through its entire 2026 AI coding tools budget.
"I'm back to the drawing board because the budget I thought I would need is blown away already."
— Praveen Neppalli Naga, CTO for Mobility and Delivery, The Information
Six Months in Six Lines
| Date | What Happened |
|---|---|
| Dec 2025 | Uber rolls out Claude Code and Cursor to 5,000 engineers. 32% adopt in the first month. |
| Feb 2026 | Adoption hits 63%. Internal leaderboard goes live, ranking teams by AI usage volume. |
| Mar 2026 | 84% agentic users. 95% monthly. CTO spends $1,200 in tokens during a two-hour demo. 70% of committed code is AI-generated. |
| Apr 2026 | Entire 2026 AI budget exhausted. Heavy users costing $500–$2,000/month per engineer. |
| May 2026 | COO publicly questions ROI. Microsoft cancels Claude Code licenses. Copilot announces token billing. |
| Jun 2026 | Uber imposes $1,500/month cap per engineer. Copilot billing goes live. Walmart caps internal AI tool. |
The leaderboard measured adoption. It got adoption. What it didn't measure — and what Uber's COO eventually admitted it couldn't connect — was whether any of that adoption produced useful features for users.
Three Companies, Two Weeks
Uber's story broke in isolation. Then it became a pattern.
Microsoft opened Claude Code access to its Experiences + Devices division — Windows, Office, Teams — in December 2025. By April, adoption had climbed past 84%. Per-engineer costs hit $500–$2,000/month. On May 14, EVP Rajesh Jha issued a memo: transition to GitHub Copilot CLI by June 30. The company that sells Copilot to the world couldn't afford to let its own engineers prefer a competitor's tool — but the real problem was simpler than strategic alignment. Fortune reported the AI was costing more than the engineers using it.
Walmart built an internal tool called Code Puppy and offered employees unlimited tokens. Demand exceeded every projection. By June, Bloomberg reported Walmart had shifted to fixed per-employee allocations.
Three of the largest employers in America. Three independent experiments with flat-rate AI coding tools. Three identical outcomes: adoption soared, budgets broke, caps followed.
The Billing Shock
While enterprises were discovering what AI coding costs at scale, GitHub made the lesson universal.
On June 1, Copilot switched from flat-rate plans to token-based billing. The official announcement drew 900 downvotes and 400+ comments. TechCrunch headlined the backlash: "What a joke."
The numbers justified the reaction. A Copilot Pro user ($10/month) gets $10 in credits. A single agentic coding session can burn $30–$40. Visual Studio Magazine documented a $180 bill on Day 1. Community reports ranged from 10x to 50x cost multipliers for power users.
GitHub's CPO Mario Rodriguez framed it as inevitable: "A quick chat question and a multi-hour autonomous coding session can cost the user the same amount." True. But the same company had spent two years marketing Copilot as an unlimited assistant for $10–$19/month. The backlash wasn't about the pricing model. It was about the bait.
GitHub reportedly spent $20 per user per month in compute while charging $10–$19. The flat rate was never the price. It was the acquisition subsidy.
The Staircase
None of these companies are abandoning AI coding tools. Uber didn't cut access — it capped at $1,500/month. Microsoft didn't ban AI assistance — it redirected to its own tool. Walmart didn't kill Code Puppy — it allocated tokens. GitHub didn't raise prices — it removed the ceiling.
Each response recreates the subsidy in miniature. Uber's $1,500 cap is still a subsidy — just a smaller one with a shorter horizon. GitHub's promotional credits for enterprise ($30–$70/user/month, June through August) are still a subsidy — just a three-month one. The subsidy isn't ending in a cliff. It's ending in a staircase.
The next step down is September, when promotional credits expire and enterprise bills arrive at actual consumption rates. A Mavvrik survey found 85% of companies already missed AI cost forecasts by more than 10%. That was measured before token billing went live.
What the Leaderboard Measured
Uber's leaderboard tracked token consumption. It didn't track features shipped, bugs introduced, review time consumed, or code maintained. It measured the thing that was easiest to count — the thing the billing dashboard already showed — and turned it into a competition.
This is the meter made by the electricity company. Token consumption looks like productivity because the tool that generates tokens is the same tool selling them. The leaderboard didn't create the confusion — it formalized it. It told 5,000 engineers that using more AI was, by definition, better engineering.
Macdonald called it tokenmaxxing. The industry will need to remember the word.
"It's very hard to draw a line between one of those stats and 'Okay now we're actually producing like 25% more useful consumer features.' ... If you're not actually able to draw a direct line to how useful features and functionality you're shipping to your users, that trade becomes harder to justify."
— Andrew Macdonald, President and COO, Uber. Fortune, May 26, 2026
They built a leaderboard. They measured adoption. They got it.
The link to useful work was not there yet.