Meta’s ‘Muse’ AI could require billions in computing power to serve 100 million users, estimate says

Meta’s generative AI offering, often discussed under the name “Muse,” may demand massive computing resources to scale, according to an estimate shared by Ivan Burazin, CEO of Daytona. In a write-up picked up by Yahoo Finance, Burazin’s “just do the math” approach ties the cost of infrastructure to user counts, arguing that supporting 100 million users would take about 65,000 CPUs and 75 petabytes of DRAM, amounting to roughly $2.8 billion in compute capacity.

The same calculation places Meta’s audience advantage front and center by comparing that 100 million user benchmark with Facebook’s reported scale. The article states Facebook has about 3 billion users, implying that serving even a fraction of that population with a high-capacity AI assistant would require a substantial expansion of data center resources.

What the estimate is, and what it is not matters for interpretation. Burazin’s framing is presented as an extrapolation rather than a disclosure from Meta, and it is not accompanied here by a confirmed Meta engineering spec, contract terms, or unit economics. Without Meta’s own numbers, the cost figure is best read as a planning-style model about how infrastructure could scale, not a verified internal budget.

Still, the basic direction is consistent with the reality of today’s large-scale AI systems: training and especially serving advanced models typically require large data center footprints, large memory subsystems, and continual compute provisioning. In that context, infrastructure tied to user experience, latency, and concurrent usage can become a material expense line as adoption grows from pilot deployments to mainstream products.

The estimate also implicitly highlights the operational challenge of moving from “AI that demos well” to “AI that runs reliably for many people at once.” Even if a system can be built, scaling it for wide audiences depends on how often users query it, how the company optimizes model size and memory usage, and how it manages throughput. The numbers in the calculation, including the DRAM requirement, reflect that serving costs can be driven as much by memory and availability targets as by raw compute.

In terms of what Meta itself has communicated, the company’s official newsroom is the place where it typically summarizes product updates, AI research, and infrastructure initiatives. However, the calculation reported by Yahoo Finance is not attributed here to an on-record statement by Meta. As a result, readers should distinguish between an industry estimate and confirmed Meta disclosures about Muse’s infrastructure plan and cost structure.

The uncertainty is straightforward: the calculation does not show the assumptions used to convert user scale into specific CPU counts, DRAM needs, and total infrastructure dollars. It also does not clarify whether the estimate assumes a particular level of concurrent demand, a particular model size, a specific approach to caching or routing queries, or a particular definition of “users” (registered accounts versus active users versus daily active usage). Those choices can swing results dramatically, even if the overall direction is plausible. Until Meta provides more detail, the estimate should be treated as a scenario, not a forecast.

For investors and business watchers, the practical next question is whether Meta can scale Muse-like experiences with lower marginal infrastructure cost through better model efficiency, tighter memory usage, or improved serving architectures. The company’s next product and infrastructure updates, if they include capacity, cost, or performance targets, would be the most relevant indicates to watch. For now, the market takeaway is that scaling generative AI to tens or hundreds of millions of users could translate into multi-billion-dollar compute commitments, depending on the design and demand profile.