Uber deployed Claude Code to roughly 5,000 engineers in late 2025. By April 2026, 95% of those engineers were using AI coding tools monthly, 70% of committed code originated from AI assistants, and the entire 2026 AI budget was gone. Uber’s COO, Andrew Macdonald, told Fortune on May 26 that the company still can’t connect that spending to shipped features: “If you’re not actually able to draw a direct line to how many useful features you’re shipping to your users, that trade becomes harder to justify.”
Uber is not an outlier. Microsoft quietly began canceling most internal Claude Code licenses across its Experiences and Devices division (Windows, Microsoft 365, Teams, Surface) with a June 30 deadline. GitHub is shifting all Copilot plans to usage-based billing on June 1. And Gartner’s updated May 2026 forecast projects worldwide AI spending will hit $2.59 trillion this year, up 47% from 2025.
The AI developer tool category has gone from a rounding error to the fastest growing, least governed line item in enterprise IT. FinOps teams that built their practice around cloud infrastructure are about to face a cost category that behaves nothing like EC2 instances or S3 buckets.
What Happened at Uber
The adoption curve was steeper than anyone planned for. When Uber introduced Claude Code in December 2025, 32% of engineers were using it by February. By March, 84% qualified as “agentic users,” meaning they were running autonomous coding workflows, not just asking chat questions. By April, the full year budget was exhausted.
Three structural factors accelerated the blowout.
Token-based pricing with no consumption ceiling. Claude Code charges per token consumed. Unlike a $19/seat/month SaaS license, there is no upper bound on what a single engineer can spend. Monthly per-engineer costs at Uber ranged from $500 to $2,000, which is 5x to 20x what most enterprises budget for a developer tool seat.
Gamified adoption with no cost signal. Uber ran internal leaderboards ranking teams by total AI tool usage. Engineers had every incentive to consume more tokens and no signal telling them what that consumption cost. This is the classic FinOps anti-pattern: the people generating spend have no visibility into what they’re spending.
Agentic workflows that multiply token consumption. CEO Dara Khosrowshahi noted that 10% of committed code is now built by fully autonomous agents. Agentic workflows consume dramatically more tokens than interactive chat. A single autonomous coding session can burn through thousands of dollars in API calls because the agent iterates, retries, and reasons across entire codebases. Goldman Sachs projects a 24x increase in global token consumption by 2030, driven largely by this type of agentic usage.
Uber’s R&D spending hit $951 million in Q1 2026, a 17% year over year increase. The AI tool line item was a meaningful contributor.
Microsoft’s Parallel Retreat
Microsoft’s situation carries a different lesson. The company piloted Claude Code in December 2025 across its Experiences and Devices division. Six months later, it began canceling most of those licenses, steering engineers toward GitHub Copilot CLI instead.
The public framing is strategic: Microsoft owns GitHub and wants its engineers on its own tooling. But the financial signal is harder to ignore. As Nvidia VP Bryan Catanzaro put it in the same Fortune report: “For my team, the cost of compute is far beyond the costs of the employees.”
When the compute cost of an AI coding assistant exceeds the salary cost of the engineer using it, the ROI math breaks. Microsoft apparently hit that threshold and pulled back. Most enterprises will reach the same conclusion; they just don’t have the fallback of owning a competing product.
June 1: GitHub Copilot’s Billing Shift Changes the Math for Everyone
Starting June 1, 2026, every GitHub Copilot plan moves to usage-based billing through AI Credits (where 1 credit = $0.01 USD). Plan prices stay the same ($19/user/month for Business, $39/user/month for Enterprise), but those prices now represent an included credit allotment, not unlimited access.
What consumes credits: Copilot Chat, Copilot CLI, the cloud agent, Copilot Spaces, Spark, and third-party coding agents. What stays unlimited: code completions and Next Edit suggestions.
The implication is significant. Any organization running GitHub Copilot Enterprise at scale will, for the first time, see per-engineer token consumption reflected in their bill. GitHub is giving admins budget controls at the enterprise, cost center, and user level. Enterprise customers get promotional credits for June through August ($70/user/month for Enterprise plans), which softens the transition but doesn’t change the structural shift.
This is the same pricing model that broke Uber’s budget, now arriving at every organization that uses Copilot.
Why Traditional IT Budgeting Fails Here
Enterprise IT budgets are built around predictable categories: cloud infrastructure (forecasted via reservation strategy), SaaS licenses (fixed per seat), and headcount (salaried). AI developer tools break all three assumptions.
They are not fixed per seat. A Claude Code license that costs $50/month for an engineer doing light chat can cost $2,000/month for one running agentic workflows on a large monorepo. The variance within a single tool is 40x.
They are not forecastable from historical usage. Uber went from 32% adoption to 95% in three months. No historical trend line would have predicted that acceleration, because the adoption curve for AI coding tools follows a network effect: once enough engineers on a team use it, holdouts adopt quickly because the codebase starts assuming AI-assisted workflows.
They sit in a governance gap. The engineering team selects the tool. The platform team provisions access. Finance sees an “AI tools” line item three months later. By then, the spending pattern is entrenched. This is the same organizational gap that created cloud cost surprises a decade ago, except AI tools can scale spending faster because there is no provisioning delay. An engineer doesn’t need to spin up a cluster; they just start prompting.
Gartner’s warning is direct: “Chief Product Officers should not confuse the deflation of commodity tokens with the democratization of frontier reasoning.” Falling per-token prices do not mean falling bills. Uber proved that comprehensively.
What FinOps Teams Should Do Now
The organizations that avoided cloud cost blowouts in the 2016 to 2020 era did so by applying three principles early: visibility, allocation, and governance. The same FinOps principles apply to AI developer tool spending, but the implementation details differ.
Build a token consumption dashboard before you need one. Most AI coding tools expose usage APIs. Pull per-engineer, per-team, and per-model token consumption into whatever observability platform your FinOps team already uses. The goal is the same as cloud cost allocation: make consumption visible to the people generating it. If your engineers don’t know they’re spending $1,200/month on Claude Code, they can’t make informed tradeoffs.
Set per-user and per-team spending caps. GitHub’s new budget controls at the enterprise, cost center, and user level are a template. If your AI tool vendor doesn’t offer equivalent controls, build them at the API gateway. Uber’s leaderboard incentivized consumption without a cost signal; the fix is pairing usage visibility with a budget constraint.
Treat AI tool selection as a procurement decision, not an engineering decision. The engineers choosing between Claude Code, Copilot, and Cursor are optimizing for capability. Finance needs to be in the room to evaluate the cost structure. A tool that is 10% more productive but 3x more expensive per token may not be the right choice at scale. This is vendor management, not shadow IT tolerance.
Model the agentic multiplier. If 10% of your code is written by autonomous agents today and that number reaches 40% in 12 months (as Goldman Sachs’ trajectory implies), your token consumption will not grow 4x. It will grow 10x or more, because agentic workflows consume far more tokens per output line than interactive prompting. Budget for the multiplier, not the current run rate.
Benchmark against the $500 to $2,000 per engineer range. If your per-engineer AI tool cost is approaching $500/month, you are in Uber territory. That is not necessarily wrong (Uber’s engineers are shipping AI-generated code at scale), but it requires an explicit ROI case that connects token spend to engineering output. Without that case, the CFO will eventually ask the same question Uber’s COO asked: is this trade worth it?
The Window Is Closing
Uber had the resources to absorb a multibillion-dollar AI experiment. Most enterprises do not. The shift to usage-based billing across the industry (GitHub on June 1, with others likely to follow) means that unmanaged AI developer tool spending will show up on the P&L within quarters, not years.
The FinOps teams that treat AI developer tools as just another SaaS line item will be the ones explaining budget overruns to the CFO. The ones that build visibility, allocation, and governance now will be the ones who saw it coming.