External estimates suggest scaling Muse could cost up to $84 billion

Meta has a new AI agent, and it is hungry. CFO Susan Li has flagged that computing costs for Muse, the company’s personal AI agent, are expected to climb as more people use it. The counterweight is a bold efficiency target. Meta expects to improve compute cost efficiency by more than ten times over the next year. What Muse is, and why it costs so much Muse went live on September 8, 2026. It is pitched as a personal AI agent that carries out complex, multi-step tasks for users across various apps and websites. That distinction matters for the cost story. A conventional chatbot answers a question and stops. An agent keeps working: it plans, clicks, checks, and tries again until the job is done. Li’s message centers on exactly that dynamic. As the user base grows, the infrastructure needed to keep all those agents running grows with it. Meta’s plan to offset the pressure relies on better utilization of its hardware and optimizations to the models themselves. The numbers outside Meta are doing External estimates suggest serving 100 million Muse users would require approximately 65,000 CPUs and 75 PB of DRAM. That works out to around $2.8 billion in costs, according to those estimates. ...