Meta’s Adam Mosseri says AI token budgets may soon be capped for engineers

Meta may soon have to put hard limits on how many AI tokens individual engineers can burn through, according to Instagram chief Adam Mosseri, who said the company could need caps within the next year or two as internal AI usage becomes more expensive.

Mosseri said the growing cost of AI experimentation is beginning to look less like a side expense and more like a core operating constraint, comparable to payroll, cloud capacity, or other budgets that managers already ration across teams. His comments come as Meta and other major tech companies wrestle with rapidly rising AI infrastructure and coding costs.

Why Meta is thinking about AI token caps now

Meta’s internal AI spending has become a serious enough issue that leaders are already changing how employees use tools, and Mosseri said token caps may eventually be part of that reset. The company has reportedly already shut down an internal leaderboard that tracked AI token usage after costs rose sharply.

Token spend refers to the cost of sending prompts to AI models and receiving responses. In a company setting, that can add up quickly when many engineers are using large language models for coding, debugging, experimentation, or workflow automation throughout the day.

Mosseri said he can imagine a near future in which the cost of a strong engineer’s AI usage could be as significant as the salary attached to that employee. In that scenario, he argued, management would likely need to impose limits.

How does Meta think about AI spending?

Meta’s view, according to Mosseri, is that token consumption should be treated like any other scarce internal resource. He compared it to the way the company already manages computing power, storage, and staffing across different teams.

In his telling, AI budgets should be allocated with the same discipline used for operational spending. That means leaders would decide where tokens create the most value, which teams deserve more room to experiment, and where usage should be constrained because the return is unclear.

What Mosseri said about budgeting

Mosseri framed AI tokens as another business resource that has to be deployed carefully, alongside capacity, operating expenses, and payroll. He said token allowances would likely need to reflect how much trust the company places in each engineer’s ability to use them productively.

He also suggested that token limits do not necessarily reflect mistrust or punishment. Instead, they may simply be a practical response to the economics of AI, especially when the models are powerful enough to encourage heavy use and the cost per query can multiply across a large workforce.

What happened inside Meta?

Meta has already started trimming some of the more wasteful internal AI behavior. Mosseri said the company reduced costs by discontinuing what he called “silly” practices, including the internal token-spend leaderboard.

That leaderboard appears to have been intended as a competitive or gamified way to track usage. But if the result was encouraging engineers to maximize consumption rather than maximize value, it likely became counterproductive as costs climbed.

Meta’s move reflects a broader shift across the tech industry: the era of unrestricted internal AI experimentation is colliding with the reality that advanced models are not free to use at scale.

How are other tech companies reacting?

Meta is not the only company trying to tame AI costs after an initially open-ended phase of experimentation. Several large technology firms are now rethinking how employees access external models and how much freedom they should have to use them.

Uber reportedly ran through its 2026 AI coding budget by April, a sign of how quickly usage can spiral when teams adopt AI tools broadly. Microsoft also reacted to mounting costs by canceling some Claude Code licenses and shifting engineers toward its own Copilot CLI tool instead.

These moves suggest that AI adoption inside enterprises is entering a more mature and more constrained phase. Early enthusiasm for productivity gains is giving way to budget discipline, procurement scrutiny, and performance measurement.

Why token costs are becoming a strategic issue

Token costs matter because they are both variable and scalable. A handful of engineers using AI occasionally may not move the needle much, but hundreds or thousands of employees running large numbers of queries can create a significant new operating expense.

That is especially relevant for companies like Meta, which deploy AI across software development, product design, infrastructure, moderation, and internal tooling. As usage spreads, even small per-query expenses can become major line items.

Mosseri suggested that if token prices remain high, companies will need to decide who gets access, how much they get, and for what purpose. In practice, that means AI usage could start to look more like cloud resource allocation than an unlimited employee perk.

What is an ROI-positive AI budget?

An ROI-positive AI budget is one that produces more value than it costs, and Mosseri argued that internal caps should depend partly on whether an engineer consistently uses AI in ways that improve output. If a worker can turn tokens into faster shipping, better code, or stronger products, they may justify a larger allowance.

That logic also creates a natural management challenge. Leaders will need to measure whether AI-assisted work is actually producing better outcomes, rather than simply increasing activity or generating more output with no material business return.

When could caps arrive?

Mosseri did not say Meta has immediate plans to introduce token limits, but he described them as plausible in the near future, perhaps within one to two years. His timeline suggests the company believes the economics of AI will continue evolving quickly enough to force a policy shift soon.

For now, Meta appears to be relying on softer controls: reducing wasteful internal behavior, reassessing tooling, and encouraging more responsible use. But the comments indicate the company is already preparing employees for a more disciplined model.

What does this mean for Meta employees?

If Meta does adopt caps, engineers would likely need to think more carefully about when and how they use AI tools. That could affect coding workflows, internal prototyping, and the speed at which teams test ideas.

It may also create a divide between employees who use AI strategically and those who lean on it heavily for routine tasks. In a capped system, judgment and efficiency could matter nearly as much as raw usage.

Still, Mosseri’s comments imply that Meta is not trying to turn off AI. Rather, it is trying to avoid a world where unconstrained usage becomes wasteful, expensive, and disconnected from business value.

Why this matters beyond Meta

Meta’s internal debate is a preview of a broader corporate challenge. As AI becomes embedded in everyday work, companies will have to decide whether access should be universal or rationed, and whether the people who use the most AI should be rewarded or monitored more closely.

The issue also hints at where the AI market may be headed. If major customers begin pushing back on usage costs, model providers may face pressure to lower prices, simplify products, or bundle more services to keep enterprises from defecting to cheaper alternatives.

Mosseri said he expects pricing to come down over time as model makers compete more aggressively for users. That would mirror other technology markets, where competition often turns expensive capabilities into more affordable commodities.

The bigger picture: AI abundance meets cost control

The current moment in AI is defined by a contradiction. Companies want employees to use the technology everywhere because it can boost productivity, but they also fear the bill that comes with unrestricted access.

That tension is now showing up in budgeting meetings, procurement decisions, and executive remarks. What once felt like a limitless digital utility is increasingly being treated like a metered service.

Meta’s experience suggests the next phase of enterprise AI adoption will not be about asking whether workers can use AI. It will be about how much they can use, who pays for it, and how leaders prove the investment is worth the money.

Key takeaways

• Meta may introduce AI token caps for engineers within one to two years, according to Adam Mosseri.
• The company is already cutting wasteful internal AI practices, including a token-spend leaderboard.
• Mosseri said AI usage should be managed like other scarce resources such as payroll and computing capacity.
• Other companies, including Uber and Microsoft, have also tightened AI spending after cost overruns.
• Lower prices from model makers could eventually reduce the need for strict internal limits.

Timeline of the AI cost squeeze

The recent discussion around Meta’s token budgets fits into a fast-moving series of cost controls across major tech companies.

What comes next?

For now, Meta is still in the phase of monitoring, adjusting, and preparing rather than imposing hard limits. But Mosseri’s comments make clear that the company sees AI token usage as an issue that will need formal management soon, not eventually.

If his forecast proves correct, the most powerful internal AI users at Meta may soon face the same reality that has long governed cloud infrastructure and headcount: access is not infinite, and every resource has to justify itself.

That shift could become a defining feature of enterprise AI in 2026 and beyond.

FAQ

Will Meta cap AI token use for engineers?

Meta has not announced token caps yet, but Adam Mosseri said the company could need them within a year or two. He suggested the limits would likely depend on how much value an engineer’s AI use creates for the business.

What are AI token budgets?

AI token budgets are spending limits tied to how many tokens employees use when prompting AI models. Tokens are the units models process for inputs and outputs, so higher usage generally means higher cost for the company.

Why is Meta worried about token costs?

Meta is worried because internal AI usage can scale quickly and become extremely expensive. The company has already taken steps to reduce wasteful behavior after costs rose enough to threaten billions of dollars in spending in 2026.

Are other companies cutting back on AI spending?

Yes. Uber reportedly blew through its AI coding budget early, and Microsoft stopped buying some Claude Code licenses in favor of its own Copilot CLI tool. Those moves show that high AI usage is becoming a cost-management issue across the industry.

Will AI prices go down?

Mosseri expects AI prices to fall over time as model providers compete more aggressively for customers. If that happens, companies may be able to loosen internal caps or offer broader access without the same financial pressure.