Uber Spent a Year's AI Budget in Four Months
In April 2026, Uber CTO Praveen Neppalli Naga confirmed to The Information that the company had already exhausted its entire annual AI budget, just four months into the year. The culprit was a company-wide rollout of Claude Code to roughly 5,000 engineers, which drove monthly per-engineer costs between $500 and $2,000 for power users. Adoption was explosive. By March, 84 percent of Uber's engineers were classified as agentic coding users, up from 32 percent in February. Nearly 70 percent of committed code was AI-generated. Eleven percent of backend updates were being written entirely by autonomous agents.
By most adoption metrics, the rollout was a success. The budget problem, however, was not a technology problem. It was a planning problem. The teams driving AI usage were not the same teams managing the spend. Usage leaderboards incentivized consumption without any mechanism connecting that consumption to business outcomes. Uber COO Andrew Macdonald said publicly that it was "very hard to draw a line" between the company's AI-assisted code commits and whether it was actually shipping more useful features to consumers. A company spending $3.4 billion on R&D in 2025 ran out of AI budget before Q2 of 2026.
This Is Not an Uber Problem
What happened at Uber was not an outlier. It was the most visible example of a structural failure playing out across the enterprise at the same time.
Meta launched an internal leaderboard called Claudeonomics in April 2026, ranking its 85,000 employees by AI token consumption and awarding badges like "Token Legend" to the heaviest users. The company's top user burned through 281 billion tokens in a single month. Collectively, Meta's employees ran through roughly 60 trillion tokens in 30 days. The leaderboard was shut down within days. Amazon issued internal guidance in May telling engineers to stop using AI "just to use AI." Salesforce was simultaneously projecting a $300 million annual Anthropic bill and shopping for model routing solutions to bring it down. The practice driving all of these outcomes now has a name: tokenmaxxing, treating AI token consumption as a proxy for productivity the same way an earlier era mistook lines of code for engineering output.
The numbers underneath this trend are striking. Per-developer token consumption rose 18.6 times in nine months. Companies were already three times over their full-year AI budgets by April. Gartner forecasts that AI agent software spending will reach $207 billion in 2026, up 139 percent from the $86 billion spent in 2025. And 62 percent of organizations still cannot accurately predict their monthly AI expenses.
Why Enterprise Budget Models Were Not Built for This
The root issue is structural. Traditional enterprise software is priced per seat. A company budgets for a known number of licenses at a fixed rate, and the annual cost is predictable within a narrow margin. AI tooling, particularly the agentic kind, does not work this way.
Token-based pricing ties cost directly to usage, and agentic workflows consume five to thirty times more tokens than simple chatbot interactions. The same tool, the same engineer, the same workday can produce wildly different invoices depending on whether that engineer is using the tool for autocomplete or orchestrating parallel agents across a large codebase. Annual budget models built around predictable per-license costs cannot absorb that kind of variance. Finance teams found out the hard way.
Compounding this is a governance gap that most enterprises have not closed. Only 43 percent of organizations have formal AI governance policies in place at all, according to survey data from the period. Only 21 percent have mature agentic governance. Most enterprises do not yet apply to AI tooling the spending controls that DevOps and FinOps teams routinely apply to cloud compute, including per-engineer usage caps, real-time consumption monitoring, and budget alerts before an overrun rather than after.
The FinOps parallel is instructive. Cloud computing went through the same cycle a decade ago. Initial adoption, an economic wall as untracked usage ballooned costs, and then a disciplined practice around matching infrastructure to actual workload needs. Forrester Research is already predicting that 25 percent of planned 2026 AI spending will get deferred to 2027 as CFOs push back on projects that cannot demonstrate measurable returns. The easy phase of AI experimentation is over. The hard work of governance has begun.
What the Smarter Response Looks Like
The companies beginning to pull ahead are not the ones slowing down their AI adoption. They are the ones changing how they govern it. Uber's post-crisis response is instructive: a hard $1,500-per-tool monthly cap, a real-time usage dashboard so every engineer can track their own spend, and a requirement for executive approval to exceed the cap. Simple controls that should have been in place at the point of rollout.
The broader pattern among companies navigating this well involves three changes in sequence. First, treating AI spend like cloud spend, which means real-time visibility, cost attribution at the team or project level, and consumption alerts before they become overruns. Second, shifting measurement from activity metrics like tokens consumed or code commits to outcome metrics like features shipped, customer impact, or hours of manual work eliminated. Third, building the ROI case before the rollout rather than trying to reconstruct it after the invoice arrives.
The signal from the market supports this shift. Enterprise buyer preference for consumption-based and outcome-based pricing together now exceeds 50 percent, while per-seat pricing preference has dropped to roughly 20 percent. That is not just a procurement preference. It is a reflection of what finance teams have learned about how AI costs actually behave.
How KAIDATA Approaches the Problem
At KAIDATA, the governance infrastructure question comes before any tooling conversation. We work with companies to build the measurement layer that makes AI spend defensible, identifying what business outcomes are being tracked, how AI activity maps to those outcomes, and where the attribution gap sits between what is being spent and what is being proven.
That work is not glamorous. It looks a lot like FinOps for AI before the AI is even deployed. But it is the difference between a rollout that survives CFO review and one that gets quietly shut down at the first budget cycle. The companies that figured this out in 2026 will have a significant structural advantage when AI costs eventually fall further, because they will already know which investments are actually working and which ones were just burning tokens.
If your AI costs are climbing while the business case feels harder to make, the problem is almost never the technology. It is the framework sitting around it.
