The technology industry has long promoted artificial intelligence as a breakthrough capable of dramatically improving productivity and reducing reliance on human labor. However, a growing number of businesses are discovering that adopting AI at scale is far more financially challenging than initially anticipated.
Rising Costs of AI Systems
A recent report by Fortune has brought renewed attention to this issue, noting that the more workers use an AI-powered system, the more costly it becomes. This trend is already observable not only in startups but also in IT giants such as Microsoft and Uber, which appear to face escalating expenses associated with artificial intelligence.
Microsoft Signals a Wider Industry Concern
As reported by Fortune and The Verge, Microsoft has begun canceling most of its direct Claude Code licenses internally and is shifting employees to the GitHub Copilot CLI solution. According to the news, several thousand Microsoft employees were initially encouraged to experiment with Anthropic's coding assistant. This change reportedly comes just months after the company began offering more internal access to Claude Code. Although Microsoft did not publicly disclose that cost was the primary reason for this decision, various sources claim that extensive use of the tool led to cost issues.
Uber's AI Budget Depleted Early
Uber has also been spending heavily on artificial intelligence. According to reports, Uber's Chief Technology Officer, Praveen Neppalli Naga, informed employees that the company had already exhausted its 2026 budget for AI code tools by the fourth month of the year. In addition to encouraging increased adoption of AI coding technologies, the company used internal leaderboards to measure usage of such tools. These developments raise questions about whether the financial burden of utilizing artificial intelligence will pose a long-term problem for firms looking to substitute or complement their workforce with AI.
Why AI Costs Rise with Usage
Unlike traditional software subscriptions, most modern AI systems rely on token-based pricing. In other words, the more an organization uses AI services, the greater the total expense becomes. In its latest predictions for 2030, research company Gartner projects that performing inference on a trillion-parameter model will cost less than ten percent of the same operation in 2025. This means that, due to increased efficiency in semiconductor technology, AI model architecture, and specialized hardware, individual AI transactions will become much cheaper over time.
However, Gartner also warned that falling token prices may not reduce overall enterprise spending on AI. The next generation of so-called "agentic" AI systems, which perform tasks with greater independence, require significantly more tokens to execute a single job than ordinary chatbots. Agentic AI systems could reportedly need 5 to 30 times more tokens per operation. Will Sommer, senior director analyst at Gartner, cautioned that companies "should not confuse the deflation of commodity tokens with the democratization of frontier reasoning." In simple terms, despite the cost reduction of each individual AI task, the sheer volume of tasks performed by enterprises could drive expenses through the roof.
Researchers Warn of Exploding Token Consumption
Scholars have begun exploring how AI autonomy will affect the economy. A recent research paper examined token consumption in agentic coding, implying that AI agents performing sophisticated coding tasks could use up to 1000 times more tokens than code chat or reasoning. However, researchers have established that higher token consumption does not necessarily lead to better performance or increased accuracy. Accuracy sometimes reaches its peak at moderate costs and does not improve despite additional computing power. This finding is relevant as firms increasingly develop AI systems designed to operate somewhat independently in various processes, including customer support, coding, and enterprise work.
Compute Costs Becoming Impossible to Ignore
According to an interview, Bryan Catanzaro, Vice President of Applied Deep Learning at NVIDIA, stated that "the cost of compute is far beyond the costs of the employees." This statement contradicts one of the fundamental assumptions regarding generative AI: that automation through replacement or support from AI will reduce operating expenses. On the contrary, the issue of increased costs related to productivity versus rising expenses on cloud infrastructure, inference, and GPUs keeps emerging. Despite these challenges, technology executives remain optimistic about an AI-powered future. Recently, NVIDIA CEO Jensen Huang said he believes every employee could eventually work alongside around 100 AI agents. However, if the trend of increased token usage persists despite lower per-token costs, the situation may turn out differently.
