A widely shared estimate says Meta’s Muse agent could require 65,000 CPUs and 75 petabytes of memory to serve 100 million users, implying about $2.8 billion of infrastructure. Multiplying that figure by Facebook’s roughly three billion users produces $84 billion, but the arithmetic is a stress test—not Meta guidance or a forecast of actual spending.
Why the linear scenario is too simple
Meta describes Muse as a personal agent running inside a dedicated secure virtual machine. That architecture can be compute-intensive because the system must plan, browse and act, rather than merely return text. Still, not every Facebook user will adopt Muse, use it simultaneously or consume the same resources. Hardware prices, model compression, scheduling and custom chips can also change unit cost.
The useful calculation is cost per active user
The third-party estimate equals about $28 of infrastructure per 100 million-user cohort, before electricity, networking, depreciation and operations. If average utilization halves, effective capacity per user improves; if agent tasks become longer or more autonomous, it worsens. BTI therefore treats $28 as an input for sensitivity analysis, not a stable margin assumption.
Meta’s official material says Muse runs on a secure VM and is being extended to consumer and enterprise uses. That makes inference efficiency a business variable: expensive usage can pressure margins unless engagement, advertising or subscription revenue scales with it.
What investors should watch
The decisive evidence will be Meta’s disclosed capital expenditure, depreciation, AI-product usage and monetization—not social-media hardware estimates. Investors should also watch whether Muse shifts more work to Meta’s custom silicon.
BTI's bottom line
The estimate usefully exposes the cost sensitivity of agentic AI, but converting it into an $84 billion bill assumes adoption and resource use that Meta has not promised.