Meta builds AI spending controls after usage spike

A centralized AI monitoring and spending control system is under development at Meta after the company realized it was spending more internally on AI than anticipated. The decision shows that companies are thinking about whether the returns from AI justify the cost involved.

The company has sent out a memo to some 6,000 employees detailing plans for AI spending caps, budgets, and token restrictions. Under the AI Gateway, teams would have access to an overview of AI usage that would automatically send notifications if there are unusual spikes in spending. The structured token management is expected to be fully implemented by 2027.

The memo noted that Meta was seeing rapid growth in internal AI adoption and that it was likely to be spending tens of billions on employee AI usage in 2026.

The after-effect of tokenmaxxing

The shift of focus at Meta from promoting the use of AI to controlling its use shows a recurring theme within corporate America. The firm used to incentivize its employees to use AI by having them establish internal leaderboards (“Claudeonomics,” named after Anthropic’s AI system). Meta no longer runs this particular leaderboard.

The broader trend has a name: “tokenmaxxing,” which is the practice of using the maximum amount of AI tokens possible for whatever reason, whether to inflate internal adoption metrics or just to consume them. The same situation occurred at Amazon after its employees established a leaderboard to track token use, but the firm later took it down in late May over concerns it was driving wasteful spending, reports Business Insider.

Uber’s experience illustrates how quickly costs can spiral. The ride-hailing company burned through its entire planned 2026 AI coding budget by April, just four months into the year. Uber COO Andrew Macdonald told Rapid Response that the company has struggled to connect token spending with measurable output. “That link is not there yet, right?” Macdonald said. “It’s very hard to draw a line between one of those stats and, ‘Okay, now we’re actually producing 25% more useful consumer features.’”

A cost problem the industry hasn’t solved

The budgetary strain goes way beyond Silicon Valley. According to a KPMG survey first reported by The Wall Street Journal, only 26% of companies have a comprehensive view of their AI costs, while 50% have partial visibility and 22% either have no visibility or only discover spending after receiving bills. As noted by Steve Chase, global leader for AI at KPMG, the company has been reportedly helping clients who have already exhausted annual token or cloud computing budgets in a matter of months.

Microsoft recently pulled back almost all the direct licenses of Claude Code and redirected engineers to its own GitHub Copilot CLI, Fortune reported, just six months after making the Anthropic tool accessible to its employees. The move came after employee usage scaled faster than anticipated.

Economic considerations suggest that initial expectations regarding the fast profitability of AI due to labor savings were overoptimistic. NVIDIA’s vice president of applied deep learning, Bryan Catanzaro, revealed to Axios that the compute cost for his group already exceeds the cost of employing people. Goldman Sachs believes that agentic AI might lead to a 24-times rise in token consumption by 2030, with monthly consumption rates hitting 120 quadrillion tokens per month, even as token prices per unit decline.

Moreover, Gartner predicts that declining token costs will not mean cheaper enterprise AI applications because agentic AI algorithms use much higher token counts per task, while providers will probably keep the total savings from their side. “Chief Product Officers should not confuse the deflation of commodity tokens with the democratization of frontier reasoning,” said Gartner senior director analyst Will Sommer. Earlier, Cryptopolitan reported that Zuckerberg admited that Meta made ‘mistakes’ on its AI transformation

What can Meta employees expect?

According to reports, the memo revealed that Meta is going to dissuade its employees from using external AI code-writing software and encourage them to use its very own assistant, MetaCode, which was previously called Devmate. These changes will be implemented in the following weeks.

At the same time, Meta’s efforts to reduce the costs related to AI come along with significant organizational changes. In March of this year, Meta was considering layoffs involving at least 20% of the total number of around 79,000 workers, part of which is caused by investments in AI infrastructure worth around $600 billion until 2028.

The CEO of OpenAI, Sam Altman, has highlighted this challenge in the industry quite well. He stated that “this is the fairest criticism right now of AI,” saying, “You hear companies saying, I am spending a ton of money on AI. And I know some great stuff is happening, but I know there’s a ton of waste.”

For the global economy, the question is whether corporate AI budgets contract before the technology delivers on its productivity promises, or whether falling token prices and better tooling close the gap first. The answer will shape hiring, capital expenditure, and competitive dynamics across industries for years.