May 25, 2026 Companies are pushing employees to use artificial intelligence tools at unprecedented levels in an effort to unlock productivity gains across engineering and operations. But early signs from major tech firms suggest that rapid adoption is creating unexpected cost pressures that are forcing some companies to scale back.
Microsoft has reportedly begun cancelling most of its direct Claude Code licences and redirecting employees toward GitHub Copilot CLI, just six months after rolling out the tool widely across its workforce. The reversal follows a surge in internal usage, as developers, designers, and project managers quickly adopted the AI coding assistant.
The move does not affect Microsoft’s broader commercial relationship with Anthropic, including its multibillion-dollar Foundry deal and access to Claude models. However, it signals a shift in how the company is managing internal AI usage after what appears to have been rapid and widespread adoption.
Microsoft is not alone. At Uber, chief technology officer Praveen Neppalli Naga said the company exhausted its entire 2026 budget for AI coding tools within the first four months of the year. The company had previously encouraged usage through internal leaderboards that ranked teams based on how heavily they used AI tools.
Similar patterns are emerging across the industry. At Meta, employees created internal systems such as “Claudeonomics” to track AI usage, while Amazon has reportedly encouraged staff to maximize token consumption, a metric tied directly to AI computing costs.
At the centre of the issue is how AI services are priced. Most large language models operate on a token-based system, where costs are determined by how much data is processed and generated. As usage increases, so does spending, even if the cost per token declines.
Forecasts suggest that this dynamic will intensify. Goldman Sachs estimates that agentic AI systems could drive a 24-fold increase in token consumption by 2030, reaching 120 quadrillion tokens per month. These systems, which perform multi-step tasks autonomously, require significantly more compute than traditional AI applications.
At the same time, the cost of individual tokens is expected to fall. Research firm Gartner projects that by 2030, inference costs for advanced models could drop by nearly 90% compared to 2025 levels. However, analysts warn that lower unit costs will not necessarily reduce overall spending.
“Chief Product Officers should not confuse the deflation of commodity tokens with the democratization of frontier reasoning,” said Gartner analyst Will Sommer. The firm notes that higher usage, combined with more complex models and limited cost pass-through from providers, could lead to rising total expenses.
Industry leaders are also acknowledging the imbalance. NVIDIA executive Bryan Catanzaro said in a recent interview that compute costs for his team already exceed the cost of employees, underscoring how expensive large-scale AI usage has become.
These developments complicate broader industry ambitions around “agentic AI,” where digital systems perform tasks alongside or in place of human workers. NVIDIA CEO Jensen Huang has said he expects a future where hundreds of AI agents operate alongside each employee. But if usage costs scale faster than efficiency gains, that vision could carry significant financial trade-offs.
The pattern emerging from companies like Microsoft and Uber suggests that the economics of AI adoption are still unsettled. While firms continue to invest heavily in the technology, the balance between productivity gains and operational costs remains unclear.
