Meta Reins In Employee AI Spending After Costs Spiral

Meta Platforms is reversing course on its internal AI strategy, telling staff in a memo that it plans to impose limits on employee token usage just weeks after pushing them to adopt AI tools more aggressively. Meta Platforms plans to clamp down on skyrocketing AI costs inside the company by imposing limits on employees' token usage, the company told staff in a memo on Tuesday, just weeks after it pushed them to adopt AI tools in their work. The memo, sent to roughly 6,000 employees, flagged an "exponential increase" in AI usage and warned that internal AI costs alone are on track to reach billions of dollars in 2026An internal memo to 6,000 employees reveals Meta is heading toward billions in AI costs from internal use alone.

To address this, Meta is building a centralized tracking system called "AI Gateway" that monitors usage and spending across teams in real time, with automated alerts for unusual cost spikesAI Gateway is an internal platform Meta is building to track AI usage and spending across teams in real time, with automated alerts for unusual spending spikes and planned budget limits tied to employee token consumption. By 2027, the company plans to move to a more structured system involving budgets and formal allocations tied to individual and team token consumption .By 2027, Meta expects to manage AI tokens through a more structured framework that includes budgets, allocation decisions and supporting tools.

The reversal is striking given how recently Meta had gone in the opposite direction. After making AI usage a "core expectation" in performance reviews, the company inadvertently fueled a phenomenon employees dubbed "tokenmaxxing" — using an internal leaderboard called "Claudeonomics," some employees racked up 73.7 trillion tokens in just over 30 days, often inflating consumption purely to improve their rankings rather than to do meaningful workEarlier, Meta had made AI usage a "core expectation" in performance reviews, which led to so-called "tokenmaxxing": employees artificially inflated their consumption through an internal leaderboard called "Claudeonomics," racking up 73.7 trillion tokens in just over 30 days. Meta's CTO Andrew Bosworth had pushed back against this dynamic even before the new memo, stating bluntly that token usage alone is not a measure of impactCTO Andrew Bosworth pushed back in a separate memo: "Nobody should be using AI tools just for the sake of using them. All motion is not progress and token usage alone is not a measure of impact of any kind." 

Alongside cost controls, Meta is also steering employees away from third-party AI tools — implicitly including Anthropic's Claude — toward its own coding assistant, MetaCode (formerly called Devmate), with engineers in the company's "Applied AI Engineering" division working to improve the tool using internal coding dataMeta also wants to steer employees away from third-party tools like Anthropic's Claude and toward its own coding assistant, MetaCode. Engineers in Meta's new "Applied AI Engineering" division are working to improve MetaCode by creating coding tasks as training data. One commentator on X captured the irony succinctly, noting that just two months earlier Meta had been the poster child for aggressive AI spending, on track to spend billions annually on tools like ClaudeMeta is doing a 180, trying to be vanguard of token-minimizing. 2 months ago Meta epitomized tokenmaxxing, on track to spend billions a year on claude etc. 

Alongside cost controls, Meta is also steering employees away from third-party AI tools — implicitly including Anthropic's Claude — toward its own coding assistant, MetaCode (formerly called Devmate), with engineers in the company's "Applied AI Engineering" division working to improve the tool using internal coding dataMeta also wants to steer employees away from third-party tools like Anthropic's Claude and toward its own coding assistant, MetaCode. Engineers in Meta's new "Applied AI Engineering" division are working to improve MetaCode by creating coding tasks as training data. One commentator on X captured the irony succinctly, noting that just two months earlier Meta had been the poster child for aggressive AI spending, on track to spend billions annually on tools like ClaudeMeta is doing a 180, trying to be vanguard of token-minimizing. 2 months ago Meta epitomized tokenmaxxing, on track to spend billions a year on claude etc. 

This internal spending crunch is separate from — and additive to — Meta's massive external AI infrastructure commitments, including a planned $600 billion data center investment through 2028Meta's internal memo stated the company is on track to spend billions of dollars on employee AI use in 2026 alone, separate from its planned $600 billion data center investment through 2028. The cost pressures have coincided with organizational upheaval: earlier this year, Meta was reportedly considering layoffs affecting at least 20% of its roughly 79,000-person workforce, partly attributed to the scale of its AI infrastructure spendingIn 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. 

Meta is not alone in confronting this problem. Amazon experienced a similar tokenmaxxing spiral and shut down its internal usage leaderboard in late May 2026 after discovering employees were gaming the systemAmazon shut down the leaderboard in late May 2026 after discovering the gaming behavior and rising cost pressure, while reports indicate Microsoft is also pulling back on internal AI usage amid similar cost concerns Microsoft is not alone in this, as Fortune reports that other companies are also pulling back on AI usage. The broader pattern reflects a kind of Jevons Paradox playing out in real time: even as the per-token cost of AI falls, total spending keeps climbing because usage — particularly for agentic AI workflows that can consume up to 1,000 times more tokens than standard queries — is growing even faster than efficiency gainsWhile it's true that the cost of training AI models is falling, making AI tokens more affordable, people have started using more tokens in their day-to-day tasks. This is particularly true for agentic AI, which can use a thousand times more tokens compared to querying an LLM. 

The episode offers a revealing case study in how quickly corporate AI strategy can whiplash from "use more" to "use less" once the bill arrives. Having spent months incentivizing employees to demonstrate AI fluency through raw consumption metrics, Meta is now discovering that volume-based adoption targets created exactly the perverse incentives one would expect — employees optimizing for the metric rather than the outcome. The push toward MetaCode adds a second dimension: beyond simply curbing spend, Meta appears keen to reduce its dependency on external AI providers like Anthropic for internal workflows, even as it continues to position itself as an AI leader externally. Whether centralized budgeting and homegrown tooling can resolve the underlying tension — that meaningful AI-assisted work and "produce evidence of AI usage" are not the same thing — may determine whether this becomes a template other Big Tech firms follow, or simply the first lap of an ongoing cycle of AI spending booms and corrections