According to external sources, as businesses turn more work processes over to AI, the billing is shifting from experimental inputs to visible operating costs. According to windman Chamath Palihapitiya, many corporate executives may not yet have a full grasp of the size of their internal AI calls, and such expenditures may be directly reflected in their profit performance in the future.

Enterprises start to face AI usage costs

In an interview with CNBC, Palihapitiya stated that some companies encouraged their employees to use the AI tool as much as possible, in the hope of improving output and efficiency. Under this approach, however, the number of model calls is rising rapidly, and the costs paid by enterprises to model service providers are magnified simultaneously.

He mentioned that Silicon Valley referred to this tendency as “tokenmaxing”, meaning to consume AI token as much as possible in exchange for more work. Token is the base unit for model data processing and costing, so the higher the usage, the larger the business bill.

Senior management or underestimation of internal call size

According to Palihapitiya, some CEOs and CFOs may not know exactly how much token is consumed within the organization. If this situation continues, some United States enterprises may even lose profits in the future because AI costs exceed expectations.

Companies have begun to speak openly about such pressures. Uber's chief technical officer, Praveen Neppalli Naga, said in April this year that the company's annual Claude Code budget was running out early. Adam Mosseri, director of Instagram, also said that the team had shut down some of the “unutilized but consuming token” projects.

  • Uber, full year, Claude Code budget is running out early.
  • Instagram indicates that high consumption has been reduced token
  • Enterprise AI expenditure is moving from pilot phase to operating subject

Low-price model speeds up the head product.

In parallel with rising cost pressures, Palihapitiya also mentioned that the high-end models of OpenAI and Anthropic were facing competition for cheaper products. According to him, models launched by companies such as Meta, Google and XAI were cheaper, but performance was close to the head product in most of the use scenarios.

In his view, the current iterative model was continuing, but the performance of a single update had not been as high as it had been earlier. For most enterprises, the rationale for continuing large-scale procurement of high-price models is diminished if cheaper models already cover major needs.

This means that the next time an enterprise deploys AI, the focus may no longer be just “more needed”, but rather a reassessment of the relationship between model selection, budget allocation and actual output.