Corporate America Is Getting Tired Of Paying Top Dollar for AI. OpenAI And Anthropic Have A Problem.

Uber used up its entire artificial intelligence (AI) budget for 2026 in just four months. Now, the company is limiting how much employees can spend on each AI coding tool to $1,500 a month. This is a big change for an industry that spent years encouraging workers to experiment with AI as much as possible. Other large companies are reaching a similar conclusion: the most powerful AI models are not always worth using for every task.

This change creates an unexpected problem for OpenAI, Anthropic and other companies that build advanced AI models. Their customers are not giving up on AI. Instead, they are learning to shop more carefully. They are combining expensive models with cheaper options and increasingly asking whether the extra computing power is worth the additional cost.

Amazon, Walmart, Cisco, Uber and Meta are among the companies putting tighter limits on AI use. Some are restricting which tools employees can use. Others are setting spending limits or sending simpler tasks to cheaper models. What started as a race to show that companies were adopting new technology has become a matter of controlling costs.

One immediate problem is that spending on AI works differently from spending on traditional business software. A company used to buying a fixed number of software licenses may now have to pay for every extra task an AI model performs. As employees take on more ambitious projects and automated AI agents handle more complex jobs, costs can rise much faster than expected.

Costs can be especially high when AI agents carry out a series of steps instead of simply answering one question. A single task may cause a model to run many times, search for information, revise its work and perform other operations. Each step uses computing power, so the total cost can grow far beyond what the job first appeared to cost.

In a survey of 2,145 business leaders conducted by KPMG in May, nearly half said their organizations had reduced their use of AI agents because the costs were greater than the benefits. This suggests that the problem affects more than just a few technology companies whose employees use AI especially heavily.

Amazon Chief Technology Officer Werner Vogels sees a wider change in how companies approach AI. Instead of automatically choosing the most powerful model, businesses are starting to ask whether a cheaper one can do the job well enough.

"Do you really need to have the biggest, highest-end model to solve this?" Vogels asked in an interview at the United Nations' AI for Good summit. His answer was simple: "No, you don't."

This view challenges one of the beliefs behind the huge valuations of leading AI companies. OpenAI and Anthropic have attracted large investments based on the idea that increasingly powerful models will become essential tools for businesses around the world. But if customers can use cheaper models for everyday tasks, they may be less willing to pay top prices just because a company offers the most advanced technology.

The competition is changing as a result. Companies are testing open-weight models, including those made by Chinese developers, while keeping access to privately developed models for tasks that need greater ability. Their goal is not necessarily to replace one provider with another. It is to avoid paying the highest price for work that does not need the highest level of performance.

This creates an opportunity for businesses that help companies manage several AI models. Rather than sending every task to one provider, software can choose a model based on the task's cost, difficulty, performance needs and security requirements.

This difference is becoming more important as companies move beyond simply testing AI. A business may still need a leading, highly advanced model for difficult analysis or complex software development. But it can use a smaller model for tasks such as sorting information, writing summaries or handling other routine work.

Ensemble Health Partners provides an example. The healthcare software company planned to spend as much as $100 million on AI this year. It reported success after moving some tasks to an OpenAI model that costs one-twentieth as much as the company's more advanced models. This offers a major reduction in running costs without forcing the company to stop using its existing provider.

These savings matter especially to companies that build AI into products and services that customers use repeatedly. Even a small reduction in the cost of one task can add up to a large amount when that task is performed millions of times.

However, there are limits to how much companies can save. Open-weight models may be available without the usual licensing fees, but businesses still need computing equipment, skilled workers and safety measures to run them. The model with the lowest advertised price is not always the cheapest to operate reliably.

Lower cost also does not mean that a model will perform just as well. Mistakes, delays and the need for extra human checks can wipe out the savings from using a cheaper model. This is especially important in industries with strict rules or in tasks where decisions can have serious consequences.

That is why the market is starting to look less like a complete rejection of expensive AI and more like a division of work. High-end models may remain useful for difficult tasks, while cheaper models take on a growing share of everyday work.

The effects could extend beyond AI model developers. Microsoft may benefit from its position as a business software and cloud services provider if more companies need systems that coordinate different AI models instead of simply giving them access to one provider.

A Morgan Stanley survey of business technology leaders found that 42% could not yet measure the returns from their spending on AI tokens, the units used to measure the text and other data models process. Another 28% said the returns were lower than the costs. Forty-three percent already had spending controls, while another 31% planned to introduce them before the end of 2027.

These figures show the financial challenge behind the industry's growth. Companies still want to use AI, but the people who approve their budgets increasingly want to know what they are getting for their money.

This puts AI companies that build the most advanced models in a difficult position. They must keep spending heavily on computing equipment and research to stay ahead of competitors, even as customers become more willing to choose cheaper options.

This financial pressure does not necessarily mean that demand for AI is falling. Companies can keep using more AI while spending grows more slowly if they negotiate lower prices, choose cheaper models or find more efficient ways to use the technology.

Recent spending data from Ramp suggest that this is already happening. The amount of AI processing, measured in tokens, has continued to rise even as spending has grown more slowly. The company says falling prices and better efficiency are partly responsible for the difference.

The result is a market where success in technology may not always lead to the same level of financial success. More businesses can use more AI to do more work without spending a similar amount more on the models themselves.

This is good news for corporate customers, especially those that have struggled to show that their early AI investments are paying off. But it creates a more difficult situation for AI providers whose financial plans depend on continued revenue growth and high profit margins.

This change also marks a shift from the early days of businesses adopting AI. Back then, company leaders worried that if they did not embrace the technology quickly, their businesses would fall behind. Encouraging employees to experiment made sense while companies were still looking for useful ways to apply AI.

But once those uses become part of everyday business operations, the financial questions change. The main goal is no longer to get employees to use as much AI as possible. It is to get the greatest possible value from every dollar spent.

Uber's exhausted budget captures this change. The company did not lose its enthusiasm for artificial intelligence. It simply ran out of the money set aside to use it.

For OpenAI, Anthropic and their competitors, that may be the most important development. Corporate America still wants AI. But businesses are becoming much less willing to pay whatever the leading providers charge.