An internal memo warned roughly 6,000 employees about an 'exponential increase' in AI consumption, signaling the end of the company's freewheeling token culture.
Meta is getting serious about its AI bill. An internal memo circulated to approximately 6,000 employees in early June warned of an “exponential increase” in AI consumption across the company, with projected costs potentially reaching billions of dollars by 2026. The response: new spending controls, a real-time usage dashboard, and formal token budgets coming by 2027.
The shift represents a fascinating about-face for a company that previously encouraged employees to use as much AI as possible. Meta had been running what insiders called a “tokenmaxxing” culture, complete with leaderboards tracking which teams consumed the most AI tokens.
From leaderboards to budgets #
The centerpiece of Meta’s new approach is something called the “AI Gateway,” a dashboard designed to give leadership real-time visibility into token usage across teams. It’s expected to launch in the coming weeks. The tool will track consumption and trigger alerts when spending spikes.
Meta’s own models, including Llama and Muse Spark, are designed with efficiency in mind. But the new governance protocols won’t discriminate. They’ll apply across all AI models used within the organization, whether homegrown or third-party.
The formalized token budgets won’t kick in until 2027, giving teams roughly a year to adjust their workflows.
An industry-wide reckoning #
Meta isn’t alone in confronting this problem. Amazon and Uber are also implementing cost-control strategies as their AI-related expenses climb.
Meta’s pivot is particularly notable because the company has been one of the most aggressive AI spenders in the industry. CEO Mark Zuckerberg has repeatedly emphasized AI as a top strategic priority, and the company has poured capital into training infrastructure, model development, and internal deployment.
What this means for investors #
Meta is implementing these controls while simultaneously scaling AI features across its consumer products, including Instagram, WhatsApp, and its family of apps serving billions of users.
Meta’s Llama models are open-source, which means external developers can use them freely. But internally, the company still bears the full infrastructure cost of running these models at scale.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our