Why companies are starting to ration AI

Major technology companies are beginning to realize that the bill for using powerful AI models is unsustainable. At Meta, executives are debating giving engineers specific, capped budgets for AI usage, much like a travel or equipment budget. Other companies, such as Uber, have already run through their entire annual AI spending plans months ahead of schedule. While some firms rely on top-tier providers like OpenAI or Anthropic, others are increasingly turning to open-source models—versions of AI that are freely available for anyone to download and modify—to avoid being reliant on a single provider’s expensive, restrictive services.

To understand the cost, think of large AI models as high-end consultants. Every time you ask a question or have the AI perform a task, you are using its computational power. These tasks are measured in tokens, which are effectively the digital syllables the AI reads and produces. Each token requires processing time on specialized hardware—powerful computer chips known as GPUs. These chips are expensive to buy, require constant electricity, and generate massive amounts of heat. As demand for AI grows, so does the demand for the power and water needed to run the massive warehouses full of these chips, known as data centers. For a company, every interaction with an AI model represents a tangible cost that adds up to millions, or billions, of dollars.

We are moving past the experimental phase where AI was a free-for-all luxury. Businesses are now focused on the bottom line, trying to determine which tasks are genuinely worth the high cost of a frontier model, and which can be handled by cheaper, internal alternatives. This transition is also sparking a broader debate about control. When a company relies entirely on an outside provider, they are essentially renting their core capabilities. By shifting toward customizable, open-source models, firms hope to retain control over their own data and processes. Ultimately, the question for the industry is shifting from how to build the smartest possible machine, to how to build one that is efficient enough to actually use profitably.