Microsoft, Uber find AI tools more expensive than expected
What's the story
Microsoft has decided to discontinue most of its direct Claude Code licenses and shift its engineering team toward GitHub Copilot CLI. The move comes just six months after the tech giant gave access to Claude Code to thousands of its employees, including developers, project managers, and designers. The rapid adoption of this tool by employees is now forcing the company to reconsider technology that its own engineers had grown reliant on.
Strategic shift
Microsoft's commercial relationship with Anthropic remains unaffected
Despite the internal changes, Microsoft's broader commercial relationship with Anthropic remains unaffected. The Foundry deal, which includes an investment of up to $5 billion in Anthropic and access to Claude models for Foundry customers, is still intact. This also includes Anthropic's commitment to purchase Azure compute capacity worth $30 billion.
Budget concerns
Uber's AI coding tools budget depleted in 4 months
Uber's Chief Technology Officer Praveen Neppalli Naga revealed that the company has already burned through its entire AI coding tools budget for 2026 in just four months. The revelation is surprising, especially since Uber was actively promoting adoption by creating internal leaderboards to track teams' use of these tools. This trend at both Microsoft and Uber highlights a potential issue with workplace AI: the more companies encourage employees to use it, the faster costs can pile up.
Cost implications
The token challenge
The problem lies in the pricing of AI computing, where large language models charge per token, the basic unit of text they process and generate. This means more efficiency and use are financially indistinguishable as both increase total spend. Major tech companies like Amazon and Meta have been pushing token consumption higher, with Amazon encouraging staff to "tokenmaxx" or use as many AI tokens as possible.
Future projections
Goldman Sachs's prediction on agentic AI systems' token consumption
Goldman Sachs predicts that agentic AI systems, which act autonomously across multiple steps instead of responding to single queries, could lead to a 24-fold increase in token consumption by 2030. This would mean enterprises deploying AI agents at scale could reach a staggering 120 quadrillion tokens per month. Despite the expected decrease in unit price of these tokens, research firm Gartner warns this price deflation won't necessarily reduce enterprise bills.
Industry acknowledgment
NVIDIA's Catanzaro highlights the cost problem
Bryan Catanzaro, Vice President of applied deep learning at NVIDIA, the world's leading supplier of chips that power AI infrastructure, recently acknowledged the cost problem in an interview with Axios. He said, "For my team, the cost of compute is far beyond the costs of the employees." This statement indicates that the economics of replacing or augmenting human labor with AI could be much more complex than previously thought.