AI Token Tsunami: Microsoft and Uber Face Exploding Costs, Reshaping Developer Tooling

Microsoft has initiated a significant internal change, canceling Claude Code licenses for thousands of its developers within the ‘Experiences and Devices’ division. This move mandates a transition to GitHub Copilot CLI, a decision driven primarily by operational cost control ahead of the fiscal year-end and a strategic push to enforce the adoption of their in-house tooling. Microsoft had initially rolled out Claude Code access in December, not solely for developers but also for project managers and designers to experience coding, indicating a rapid four-month turnaround on their investment.

Simultaneously, Uber has also faced an unprecedented surge in AI expenditure, reportedly exhausting its entire 2026 AI budget in just four months due to Claude Code usage. Uber’s 2025 R&D spend was $3.4 billion, a 9% increase from the previous year. By March, 95% of Uber engineers were using AI tools monthly, with 70% committing AI-generated code. This led to an average cost of $150 per engineer, though some individuals racked up bills as high as $2,000. Notably, Uber CTO Praveen Nepali Naga confirmed personally spending $1,200 in a single two-hour Claude Code session. The underlying issue for both companies is the industry’s shift from predictable per-seat licensing to variable per-token billing models for AI services. Uber is now exploring OpenAI Codex alongside Claude Code to mitigate costs.

This trend, dubbed the ‘end of the all-you-can-eat token buffet,’ by Javier Pastor, highlights a critical market miscalculation. Many enterprises had anticipated a continuous decline in AI computation costs, echoing historical trends in hardware. However, recent data, including OpenAI’s revenue growth falling short of projections, indicates that high-performance AI models, such as Anthropic’s Opus 4.7 and Gemini 3.5 Flash, are exhibiting significant price multipliers (15x and 14x, respectively) compared to earlier versions. This unexpected price escalation is forcing a reevaluation of AI adoption strategies, posing the fundamental question of whether the accelerated code generation and increased development speed provided by these tools genuinely deliver a proportional return on investment, a metric often overlooked amidst the excitement of rapid deployment. While some Chinese AI models like Deepseek and Qwen 3.7 Max offer potentially more competitive pricing, they too are beginning to reflect similar cost pressures.