# AI buyers push back on token spending as efficiency takes priority

Enterprise enthusiasm for AI is colliding with a more practical concern: cost. After a stretch in which heavy model usage was often treated as a proxy for business value, companies and vendors are now taking a harder look at what they are actually getting for the money they spend on tokens.

The shift comes as AI investment continues to surge. Gartner has forecast that worldwide AI spending will approach $2.6 trillion this year, a 47% jump from 2025. But executives are increasingly warning that the economics of broad AI deployment are getting harder to ignore.

Recent comments from major industry players suggest the market is moving away from a simple more-is-better mentality. Microsoft CEO Satya Nadella said in an interview that using large, expensive models for every task can be addictive, but said his company is pushing teams toward more efficient options when the problem does not require frontier-level capability. His remarks followed a Microsoft Build event in which efficiency featured prominently in product announcements.

Budget pressure is showing up elsewhere too. Uber’s CTO said the company used up its annual Claude Code allocation by April, underscoring how quickly AI tools can become expensive at scale. At the same time, other industry voices are trying to reframe the conversation away from raw usage and toward business return. At a recent Nebius event, one executive argued for a focus on extracting as much value as possible from each token rather than celebrating high consumption.

The pushback reflects a growing realization inside enterprises: activity is not the same as impact. Raj Ramanujam, vice president of global alliances and cloud at Dynatrace, said companies often rushed into AI because they feared falling behind, building systems and workflows before fully considering long-term operating costs. He said many are only now coming to terms with those expenses.

Rob May, chief executive of Neurometric.ai, said the earlier fascination with token volume came from a genuine need to measure AI adoption and performance in the workplace. But he argued that counting tokens alone can be misleading because not all usage carries the same value. Simple tasks and highly complex work may both consume tokens, even though their business importance is very different.

That distinction has led some in the industry to call for a more selective approach. May has promoted what he calls "tokenminning," a framework that emphasizes using cheaper or smaller models when they are sufficient for the job. In his view, enterprises do not need the most advanced model for every prompt or workflow.

The renewed focus on efficiency may also affect competition among model providers. Reports have said OpenAI and Anthropic are weighing price cuts on tokens as they compete for customers ahead of planned public listings. That would suggest pricing pressure is building even among the largest AI companies.

For enterprises, the emerging lesson is straightforward. AI may still be moving quickly into business operations, but the real test is no longer how much it is used. It is whether the technology delivers measurable value at a sustainable cost. As more companies examine where frontier models are actually necessary, smaller language models and open-source alternatives may gain a stronger foothold.

If that trend continues, the industry’s most expensive models could face a more selective market than the one many expected during the first wave of AI adoption.