News
Tech Firms Pull Back From 'Tokenmaxxing' After AI Spending Spirals
Corporations that were praising employees for maximizing AI token usage as recently as this spring are now sharply cutting spending after Uber and Microsoft overpaid for excess licenses.
Contents
Just a few months ago, executives at the biggest tech companies were bragging about how many thousands of dollars their engineers were spending each month on AI model tokens. Today, that same trend, dubbed tokenmaxxing, is being called one of the costliest mistakes in corporate AI adoption, and companies like Uber and Microsoft have already felt the bill.
Where Tokenmaxxing Came From
The term took off in early 2026, as tools like Claude Code and Codex became capable enough that companies began treating token consumption as a measure of employee engagement and innovation. The push came in March from Nvidia CEO Jensen Huang, who said he would be deeply concerned if an engineer earning 500,000 dollars a year wasn't spending 200,000 dollars on tokens.
I would be deeply concerned if an engineer earning 500,000 dollars wasn't spending 200,000 dollars on tokens - Jensen Huang, CEO of Nvidia
Databricks CEO Ali Ghodsi went a step further, publicly praising an employee who spent over 7,000 dollars on model queries in just two weeks. Attitudes like these from industry leaders were meant to encourage engineering teams to make maximum use of premium models, with little reflection on whether each query actually delivered value.
A Bill That Arrived Faster Than the Payoff
The first serious crack in the trend appeared at Uber, which rolled out Claude Code to roughly 5,000 engineers and had already burned through its budget planned for all of 2026 by April. Microsoft took a different route, scaling back Claude Code licenses after costs outpaced the tool's expected benefits.
Karthik Sj of LogicMonitor, quoted by Forbes, notes that high token consumption alone doesn't mean tokens are being used sensibly. He pointed to cases where employees fired up expensive, premium language models for tasks as simple as checking the contents of an email, tasks that required no advanced computing power at all.
From Maximizing Usage to Maximizing Value
After the reports about Uber and Microsoft, companies began moving away from relying on a single premium provider toward a mixed portfolio of cheaper and pricier models, matched to the specific task. Carmen Li of Silicon Data describes it as a clear shift toward more deliberate token consumption among development teams.
Token spend was like the wild west, with little insight into what real value that spending was delivering - Mike Sinoway, Lucidworks
Mike Sinoway of Lucidworks notes that companies are only now starting to treat visibility into token costs as a management priority rather than a technical detail. Previously, AI spending grew without systemic controls, since sheer volume of usage mattered more than the business outcome it produced.
New Billing Headaches
Mixing multiple providers and models within a single organization also creates new accounting headaches. Michael Hahn of auditing firm Vaudit points out that billing errors arise, among other things, when unfulfilled API requests are still invoiced, when retries generate duplicate charges, or when token rates are calculated incorrectly against a provider's price list.
Debo Dutta of Nutanix notes that current engineer spending, ranging from a few thousand to over 50,000 dollars a year per person, is just the beginning. The shift from coding assistants to fully agentic systems that carry out chains of tasks on their own, without human oversight at every step, is expected to increase token consumption by an order of magnitude.
What It Means for Polish Companies
For companies in Poland that are only now scaling up deployments of tools like Claude Code or Copilot, the experiences of Uber and Microsoft are a warning to measure the return on token spending from the outset, not just the raw level of consumption. Without that kind of cost control, a budget planned for an entire year could vanish within a few months, before the tools have had time to deliver measurable time savings.
The analytics industry signals that the next stage will be tying token spending more tightly to concrete business metrics, such as time saved on a task or the number of features shipped, rather than treating query volume itself as a measure of progress in AI adoption.
