Uber gave 5,000 engineers Claude Code. By April, the CTO told the company they had burned through the entire 2026 AI budget in four months. Cost per engineer: $500 to $2,000 per month (briefs.co, 2026). The story everyone tells is about runaway spending. The story nobody tells: Uber has no idea which of those 5,000 engineers compounded that spend, and which wasted it.
This is the new problem. Not whether to give engineers AI tools. That question is settled. The problem is that you are spending real money on AI-generated code and you have no way to measure whether the humans reviewing it are any good at reviewing it.
How Did Token Budgets Become a Line Item?
Jensen Huang said it at GTC 2026: every NVIDIA engineer gets a token budget worth roughly 50% of their base salary. A $500K engineer should spend $250K a year on tokens (CNBC, 2026). "What used to be a thing for engineers is when you come to work, they give you a laptop. Now when you come to work, they give you a laptop and tokens."
That is not a perk. It is compensation infrastructure. And unlike a laptop, tokens get consumed differently by different engineers. Some compound them into shipped features. Others burn through them generating code nobody validates.
At Uber, 70% of committed code now comes from AI (briefs.co, 2026). Adoption went from 32% to 84% of engineers in months. That velocity is impressive until you ask: who is reviewing the output?
What Does the Verification Gap Actually Cost?
Here is the paradox. 96% of developers say they don't fully trust AI-generated code. But only 48% always verify it before committing (SonarSource, 2026). Nearly everyone is skeptical. Half of them commit anyway.
Werner Vogels, AWS CTO, named this "verification debt" at re:Invent 2025. When you write code yourself, comprehension comes with the act of creation. When the machine writes it, you have to rebuild comprehension during review. That takes time. McKinsey found that code review time increased 12% as AI adoption rose (McKinsey, 2025). You are paying for AI to write more code and paying humans more time to review it. The net equation only works if your engineers are excellent reviewers.
Most are not. 38% of developers say reviewing AI-generated code takes longer than reviewing code from colleagues (IT Pro / SonarSource, 2026). AI-coauthored pull requests carry 1.7x more major issues than human-only PRs (CodeRabbit, 2025). And experienced developers who think AI makes them 24% faster are actually 19% slower in controlled trials (METR.org, 2025).
That last stat deserves to sit with you. The perception gap is the risk.
Why Does This Surface at Hire?
Claude Code had a 47% performance regression in April 2026. Dave Kennedy at TrustedSec tracked defects, security issues, and task completion rates falling off a cliff (Fortune, 2026). Most users did not catch the drop for weeks. The model got worse. Engineers who rubber-stamp AI output were invisible until the model failed under them.
That is the hiring problem. A whiteboard interview won't surface this gap. Neither will a take-home that bans AI. The only way to see it is to put a candidate in a realistic environment with AI tools and watch what they do with the output. Do they validate the output? Do they catch when the model drifts, or do they ship it and move on?
Graphite CTO @elmd_ puts it clearly: AI reviewers working from diffs alone miss the global invariants, API conventions, and architectural decisions that exist outside the diff. The tacit context engineers absorb while writing code does not transfer automatically when AI writes it instead. Someone has to reconstruct it. That is the skill.
What Does Measuring This Look Like?
When tokens cost half a salary, "who can use AI well" stops being a preference question and becomes a budget question. Fairground's screener puts candidates in a real IDE with AI tools and captures the process, not just the output. Every prompt, every iteration, every validation decision, every moment they ignored a warning. That scorecard is the pre-hire prediction of how someone will use their token budget.
Start free, 100 credits. No credit card. No sales call. fairground.work
Related: Every Engineer Now Has a Token Budget
Get started with Fairground in just few mins.
Plug and Play. Works well with your existing ATS.
100 Free Credits
