Corporate America is beginning to scale back its unrestricted spending on artificial intelligence, as the skyrocketing costs of AI computing tokens are prompting companies to ration usage and seek out more economical tools. Many enterprises have found themselves exhausting their annual AI budgets within a few months; for instance, Uber Technologies reportedly depleted its annual budget for "agentic" or autonomous AI use by March. This shift follows an initial period where companies encouraged widespread AI adoption, often leading to employees engaging in "tokenmaxxing" – using as much computing as possible to appear AI-forward, even as model providers transitioned to usage-based pricing.
This reevaluation comes as companies like Google report immense growth in AI usage, now processing over 3.2 quadrillion tokens monthly, a sevenfold increase from a year ago. However, data from EntelligenceAI, a startup aggregating data from over 2,000 companies, indicates that only 18% of spending on advanced AI coding tokens translates into shipped products. This suggests AI models still require significant improvement to justify the expense, with executives noting high costs associated with debugging and rewriting AI-generated code. Some top technical executives, including those at Uber, Meta Platforms, Microsoft, Salesforce, and DoorDash, are implementing new strategies to ensure AI use directly contributes to productivity or are limiting access to certain tools for employees.
Companies are exploring various strategies to manage costs. This includes directing employees towards cheaper, in-house tools or less powerful, more cost-effective AI models. Matan Grinberg, CEO of Factory, a coding automator, noted that an executive at a major financial institution reported employees spending hundreds of thousands of dollars a month on tokens, often using premium, expensive models for simple queries. While some critics view this cost-cutting as a potential slowdown in AI growth, investors and tech executives largely believe sales and usage will continue to climb faster than forecasts, indicating that the industry is still in its early stages of adoption and optimization. Firms like Anthropic continue to highlight significant productivity gains, stating their models can help complete complex tasks in less than two weeks that previously took over seven months.
