In particular, the CEOs of Microsoft, Uber and Nvidia began to publicly discuss the problem of the cost-effectiveness of mass adoption of AI. Their main message is that AI costs have turned out to be higher than expected, while economic efficiency has been lower than forecast.
The debate about the real cost of enterprise AI intensified after a series of publications in the U.S. technology media and business press. The focus is on the dramatic rise in the cost of computing, API requests and infrastructure for generative AI.
In particular, The Verge reported that Microsoft intends to reduce the use of Claude Code – Anthropic’s AI programming tool – and move some employees to its own GitHub Copilot CLI. According to the publication, the decision is related not only to the unification of internal tools, but also to the high cost of using Claude Code.
Uber faced a similar problem. By April, four months into the calendar year, Uber had exhausted its entire artificial intelligence budget for 2026 after Forbes writes.
Anthropic’s Claude code had spread to about 5,000 engineers faster than the company’s financial models had anticipated, according to the publication. Uber’s total research and development spending will reach $3.4 billion in 2025, up 9% from a year earlier.
Anthropic is changing its payment model
The claims of excessive spending on AI tools come amid structural changes at Anthropic itself that have fueled user frustration.
The company announced on May 13 that paid Claude subscribers will soon face a separate monthly bill for agent tools and third-party solutions, paid at the full application programming interface (API) rate. The new payment model will go into effect on June 15.
Together, these two developments describe the same problem. Consumption-based token pricing doesn’t work as well as the models CFOs in charge of software know how to model.
In this context, the gap between what engineers consume and what finance teams expect is no longer hypothetical.
This was indirectly confirmed by Uber president Andrew McDonald, who admitted in an interview with The Verge that the company is still struggling to directly link the growth in AI spending to a proportional increase in useful features and products for users.
Nvidia has also recognized the problem of the cost of computing
Nvidia’s vice president of applied deep learning Brian Catazaro’s statement caused additional resonance. As Axios reports, the top manager said that for his team “the cost of computing far exceeds the cost of employees.”
This assessment was one of the first public admissions from a representative of the AI industry itself that the current model of using generative AI could be significantly more expensive than traditional labor.
Against the backdrop of rising costs, analysts are increasingly debating whether enterprise AI is actually reducing costs or just redistributing them toward cloud infrastructure, GPUs, and API services.
From enthusiasm to economic efficiency
The situation shows that the generative AI market is entering a new phase – from the period of mass enthusiasm to the stage of evaluating economic efficiency.
For investors and the corporate sector, this means an increased focus on the ROI of AI projects, the cost of tokens, data center costs, and the efficiency of AI infrastructure.
This is especially relevant for companies that have started to implement AI “on an industrial scale” in software development, analytics, customer support and internal business processes.
At the same time, most experts are not talking about the failure of AI as a technology. Rather, they are talking about a revision of expectations: companies are trying to understand which AI application scenarios really give a sustainable economic effect, and which ones so far only lead to an increase in operating costs.
