AI Agents Drive 24-Fold Increase in Token Demand, Raising Serious Cost Concerns

An Alarming Situation:

2026 Budget Exhausted Within Months While the adoption of AI accelerates, its operational costs are beginning to take a toll on major tech companies. Praveen Neppalli Naga, CTO of Uber, shocked the industry by revealing that the company had already exhausted its entire AI budget for the fiscal year 2026 within just a few months. Andrew Macdonald, Uber’s Chief Operating Officer, told Business Insider that there is no clear correlation between the rise in token usage and the actual improvement in consumer-facing features. After discussions with senior engineers, it was found that while token usage had increased, it did not proportionally enhance features that benefit customers. Macdonald admitted that although the volume of code deployments had risen, “it was extremely difficult to draw a clear line between that and actual software improvements.” This statement underscores a significant issue faced by the entire industry: massive investments in AI are not necessarily translating into tangible business outcomes.

Goldman Sachs Warns of a 24-Fold Increase in

Token Demand Exacerbating the situation is the explosive growth in token demand driven by the rapid adoption of AI agents. According to Goldman Sachs estimates, token usage could increase more than 24 times over the next few years due to the deployment of AI agents. AI agents are said to consume over 1,000 times more tokens than a single AI chatbot. This indicates that transitioning from conventional conversational AI to autonomous task-executing agent-based AI will result in exponentially higher infrastructure costs. Uber’s case demonstrates that this structural cost issue is no longer hypothetical but a current reality. The pace at which annual budgets are consumed within mere months raises critical questions about the sustainability of corporate AI strategies.

Microsoft Moves to Cut Costs Microsoft, too,

has been unable to escape the pressure of rising AI costs. Earlier this month, the company announced plans to discontinue developer access to its Claude Code service and shift to its in-house Copilot CLI tool by June 30. While the move is officially described as part of a team tool integration effort, the timing—coinciding with the end of Microsoft’s fiscal year—has spurred speculation that cost-cutting was a significant factor. Additionally, Microsoft has announced a transition to a token-based pricing model for GitHub Copilot. This move follows a sharp increase in the tool’s operational costs earlier this year, signaling an industry-wide push to reassess AI service cost structures.

The Dilemma of Tokens and ROI Nvidia CEO

Jensen Huang said in March that he would be concerned if an Nvidia engineer earning $500,000 a year wasn’t using at least $250,000 worth of tokens during the same period. This statement reflects the industry’s stance on maximizing AI utilization. However, as the cases of Uber and Microsoft illustrate, heavy token consumption does not always equate to business value creation. A growing gap is emerging between the demand for AI and the costs that companies can realistically bear. With token demand expected to surge further due to the proliferation of AI agents, businesses are being compelled to fundamentally rethink their AI investments. Moving forward, the focus will need to shift from simply adopting AI to measuring its return on investment and optimizing costs, which will become critical components of future AI strategies.

The Future of the Industry Hinges on AI Cost

Management Reports from Uber, Microsoft, and Goldman Sachs reveal that the AI industry is at a critical turning point. The explosive growth in token demand and the inability of corporate budget structures to keep pace create a paradox that could shape the direction of next-generation AI. The issue of cost is not just a business challenge but also a societal one. As AI agents become more widespread, the sustainability of token consumption itself comes under scrutiny. The development of cost-effective, high-performance AI models, advancements in efficient inference technologies, and the establishment of architectures that minimize token consumption could provide the fundamental solutions needed to address the AI cost crisis.