Estimating Infrastructure Needs for Meta Muse at 100M DAU

Serving 100 million daily active users (DAU) for Meta's Muse agent requires an estimated 1–2 GW of average total power in the base case, with only ~0.1 GW attributed to the CPU/VM sandbox layer. The remaining power demand is driven by model inference, which could scale to 3–4 GW depending on the volume of reasoning-equivalent events per user.

Sandbox Infrastructure Costs

The CPU and memory requirements for the sandbox layer are significantly lower than naive calculations suggest due to high oversubscription rates.

• CPU Demand: Assuming 25 million live VMs at peak (25% of DAU) and 0.5 physical cores per live VM, the system requires 12.5 million physical CPU cores. At current public pricing, this translates to ~$800 million in CPU hardware.
• DRAM Demand: With ~75–100 PB of physical DRAM required (based on 3GB actual usage per VM), the cost is approximately $2 billion.
• Power: The entire sandbox layer consumes only ~0.1 GW.

Inference as the Primary Bottleneck

Inference remains the dominant compute cost. Based on Microsoft’s 2026 study, long reasoning queries consume approximately 4 Wh each.

• If each user generates 50 heavy inference events per day, the total daily energy consumption is 25 GWh, requiring ~1 GW of average power.
• As agents perform longer trajectories and spawn sub-agents, inference demand scales faster than user count, making it the critical infrastructure constraint.

Conclusion

While consumer agents drive new demand for CPUs and conventional DRAM, inference remains the real compute bottleneck. The popular framing that each user needs dedicated high-end hardware exaggerates CPU needs; the actual bottleneck is the energy required to serve model responses.