Uber's AI coding bill is the warning shot: companies gave employees powerful agents before they decided what those agents were allowed to cost.
Uber's 2026 AI coding budget was gone by April. Not squeezed. Not running a little hot. Gone. According to a recent Financial Times report, the company put a $1,500 monthly cap on each employee's use of tools such as Claude Code and Cursor after AI token spending ran ahead of the plan in the first four months of the year.
You can see why every CTO is paying attention. Uber is not a small team discovering a billing page too late. It has thousands of engineers, a serious internal technology organization, and a president and chief operating officer, Andrew Macdonald, willing to ask the question out loud: is all this AI use actually improving the products customers touch?
That is the right question. Frankly, it should have been asked before the budget was opened.
The same pattern is showing up elsewhere. Business Insider reported that Walmart limited use of Code Puppy, its internal AI coding tool, after employee adoption rose quickly and workers started sending repeated requests for similar problems. Walmart Global CTO Suresh Kumar framed the limit as a way to reduce duplication, not as a retreat from AI. That distinction matters. The issue is not whether engineers should use AI coding tools. Of course they should, when the work earns the spend. The issue is whether the company knows the difference between useful automation and expensive noise.
AI agents make that harder than old software ever did. A seat license for a traditional SaaS tool is dull but predictable. A coding agent can make one request, call a model several times, inspect files, retry an answer, invoke tools, and keep burning tokens while the employee sees a single task box on screen. To the user, it feels like one job. To finance, it can look like a meter left running.
Axios reported in late May that one consultant's client spent $500 million in a single month on Claude after failing to set usage limits for employees. We don't know the company's name, and we shouldn't pretend we do. The useful fact is simpler: a cost that large can happen when access is broad, usage is unmetered, and nobody has authority to stop the bill before it arrives.
The bill is forcing discipline
The correction is already underway. The Financial Times reported that companies including Uber, Walmart, Cisco, Amazon, and Meta have started tightening employee AI use as the cost of agents strains budgets. KPMG survey data cited this month found that only 26% of businesses have a clear, comprehensive view of AI costs, while half have only partial visibility and 22% either lack visibility or learn the cost after the invoice lands.
If you run a company, that should bother you. You would not let every employee choose a cloud instance, run it without labels, and explain the purpose later. Yet many companies have treated AI tools exactly that way because the early adoption graph looked good in board slides. Usage is not value. A busy model can still be a lazy investment.
The dumbest waste is also the easiest to understand: employees reaching for the strongest model for work that doesn't need it. A frontier coding model might be the right tool for a risky refactor or a strange production bug. It is not the right tool for renaming a variable, drafting a routine note, or summarizing a short paragraph. You don't need the most expensive system in the stack to do the cheapest job in the queue.
That is where the startup opportunity sits. Business Insider reported this month that OpenRouter raised $113 million at a $1.3 billion valuation after building a business around routing AI tasks across different models. The pitch is practical rather than glamorous: send simpler work to cheaper models, reserve the expensive models for work that justifies them, and give companies a clearer view of where the money is going. Concentrate AI, another routing startup, came out of stealth with more than $5 million in funding in the same report.
Routing will not fix bad management. It gives managers fewer excuses. If a company can see who is using which model, for what kind of work, and at what cost, the conversation changes quickly. A team that can show a coding agent cut a release cycle by two weeks will keep its budget. A team using premium models to generate throwaway boilerplate will have a harder time.
Here is the thing: the companies that handle this well probably will not spend less on AI. They will spend with more intent. Caps, model tiers, usage dashboards, and per-project budgets are not anti-AI measures. They are what serious deployment looks like after the first wave of experimentation.
Uber's mistake was not using AI coding tools. The mistake was letting the spending story get ahead of the product story. Once that happens, the CFO stops hearing about developer velocity and starts seeing a line item with no clean owner. Every founder selling AI tools into the enterprise should learn from that. If your product can burn through a budget faster than procurement can understand it, cost control is not an admin feature. It is part of the product.
Also read: Zoox unveils a redesigned robotaxi but a federal waiver stands between Amazon and its first fare • Vishal Sikka's new AI startup wants to do what Infosys once paid him to prevent • Google delays Gemini 3.5 Pro to July as talent exodus deepens the pressure on its AI ambitions
