Companies that rushed to put AI tools in their workers’ paws are starting to discover the less glamorous side of the revolution: the bill.
According to the Financial Times, Amazon, Walmart, Cisco, Uber and Meta are among the early corporate adopters now trying to control how much their employees spend on AI. Some are imposing limits, while others are steering staff towards cheaper models when the most powerful systems are not needed.
The mood has shifted because AI use inside companies has moved beyond casual chatbot experiments. Businesses are now deploying AI agents, which can carry out more complex tasks, but require far more computing power.
That extra compute is becoming harder to hide. As providers move from flat subscription models to token-based billing, companies are being charged more directly for each prompt, task and automated workflow.
The result is that AI is no longer being treated as a magic box of productivity but as a cost centre, which is where most corporate dreams go to be beaten with spreadsheets.
Deloitte’s Costi Perricos said finance chiefs and boards are now beginning to think seriously about compute costs. The earlier idea that AI was cheap or free is starting to look like another comforting myth sold during the opening act of a technology boom.
OpenAI chief executive Sam Altman has acknowledged that cost has become a major issue for customers this year. According to him, it was not a serious concern last year, when companies were still more focused on adoption than the scale of the bill.
Uber has already imposed limits on employee AI use, reportedly capping spending on individual tools at $1,500 a month. The company had burned through its AI budget for 2026 by April.
Uber president and chief operating officer Andrew Macdonald said it is becoming harder to prove that token spending leads directly to better products. The challenge is showing that heavy AI use produces measurable improvements, not just impressive internal usage figures.
Walmart has taken a similar approach with its internal AI agent. Its Code Puppy coding platform became so popular that the retailer decided to step back and look more carefully at which tools should be used for which jobs.
Cisco is seeing the problem from the infrastructure side. Jeetu Patel, its president and chief product officer, said AI agents need far more resources than chatbots, especially if each worker is supported by multiple agents running in the background.
Analysts at Goldman Sachs expect AI agents to drive a huge rise in token consumption by 2030. They also expect demand to worsen chip shortages over the next 12 to 18 months.
This creates a problem for the largest AI labs, including OpenAI and Anthropic. Their growth depends on expanding usage, but corporate customers are now becoming more careful about what that usage costs.
China is also gaining an advantage in AI token consumption. Cheaper energy and more efficient models have helped Chinese AI labs offer lower token prices than many leading US rivals.
The corporate AI message is changing quickly. Companies still want the productivity gains, but they no longer want staff throwing premium models at every routine task as if the company card were blessed by heaven. AI adoption is still accelerating, but the free-for-all phase is ending.
