Muse Agent Compute Demand Estimate: How Much Power Is Required to Support 100 Million Daily Active Users?

Author: Robonomics

Compiled by: TechFlow

TechFlow Summary: When Meta embeds the Muse agent into every user's computer, the real bottleneck isn't the CPU, but the inference layer. Using sandbox data from DeepSeek as a benchmark, this analysis reveals a counterintuitive breakdown: the sandbox layer consumes only 0.1GW, while the inference layer could potentially draw 3 to 4GW. For those focused on AI infrastructure investment and energy sectors, this offers a rare look at the compute ledger.

Rough Conclusion

In the base case, supporting 100 million daily active users would require an average total power of 1 to 2GW, with the CPU/VM layer accounting for only about 0.1GW. Depending on the number of "inference equivalents" generated per Muse DAU daily, 3 to 4GW is also entirely reasonable.

The CPU procurement cost for the sandbox layer is under $1 billion, with DRAM costs around $2 billion, significantly lower than many expected.

1. Muse VM / Sandbox Infrastructure

Each user is allocated 2 vCPUs, approximately 8GB of memory, and around 100GB of persistent logical storage. This reflects the observed configuration of the Muse VM.

A. Sandbox CPU

The first mistake is assuming that 100 million DAUs means 100 million VMs are simultaneously consuming compute.

Assuming each Muse DAU's agent actively works for two hours per day:

100 million users × 2 hours / 24 hours = approximately 8 million simultaneously active VMs

Meta obviously cannot provision resources based solely on daily averages. Usage will cluster during waking hours and exhibit burstiness.

Assuming a peak-to-average ratio of 2.5x:

8 million × 2.5 = approximately 20 million peak-active VMs

Add another 20% capacity headroom: approximately 25 million always-ready online VMs. Thus, the core assumption is that Meta's compute infrastructure needs to support roughly 25% of its DAUs being online simultaneously.

A key distinction lies between virtual CPU allocation and physical CPU requirements. Agent sandboxes are highly suited for CPU oversubscription. They spend most of their time waiting. During wait states, VMs may still be alive but consume almost no CPU.

DeepSeek's recently released DSec infrastructure provides a useful benchmark. Its production-grade agent sandbox platform runs on approximately 160 nodes with around 30,000 physical CPU cores and 250TB of DRAM, handling over 380,000 concurrent sandboxes at peak. DSec also demonstrated stable operation capabilities: roughly 800 micro-VMs per node. Calculating at ~188 physical cores per node: 188 physical cores / 800 micro-VMs = ~0.23 physical cores per live VM.

DeepSeek is clearly the efficiency leader. Muse's figures likely fall between 0.3 to 0.75 physical cores per live VM, with the base case set at 0.5 physical cores per live VM. This translates to each physical CPU core supporting roughly two simultaneously live Muse VMs.

Plugging into the base assumption: 25 million live VMs × 0.5 physical cores per VM = 12.5 million physical CPU cores.

On a 256-core CPU: 12.5 million cores / 256 cores per CPU = approximately 50,000 CPUs; on a 192-core CPU, it would be 65,000.

At current publicly listed prices, this amounts to roughly $800 million.

B. Sandbox Memory

Muse exposes approximately 8GB of memory to the user environment, though one observed instance actually utilized only about 3GB when measured.

25 million live VMs × 3GB = 75PB of physical DRAM, establishing a reasonable baseline range of 75 to 100PB.

At current publicly listed prices, this amounts to roughly $2 billion.

C. Sandbox Power Consumption

At a scale of 100 million DAUs, the entire Muse sandbox/VM layer requires approximately 0.1GW.

2. Muse Spark Inference

Model inference is handled through Meta's external inference infrastructure. This is Meta's description of the Muse architecture.

Inference Layer

The user's PC executes tools and manages state locally, but actual model execution occurs on separate inference infrastructure. This is Meta's description of the Muse architecture.

Energy Consumption per Inference Event

Microsoft's 2026 research estimates that the median energy consumption for optimized frontier-scale inference per standard query is approximately 0.31Wh.

However, a long-form inference query contains roughly 15 times the tokens of a standard query, consuming about 13 times more energy, amounting to roughly 4Wh per long inference query.

The study specifically notes that inference and agent workloads exhibit significantly higher energy consumption. Refer to Microsoft Research's paper on AI inference energy consumption.

For workloads like Muse, we assume the energy consumption per heavy inference-equivalent event falls between 5 to 10Wh, serving as a rough sensitivity range.

Sensitivity Analysis: Inference-Equivalent Events per DAU Daily

Assume each DAU generates 50 inference-equivalent events per day.

Assume each active Muse user generates a workload equivalent to 50 heavy inference events per day.

Calculated at 5Wh per event:

100 million users × 50 events/day × 5Wh = 25 million kWh/day

25 million kWh/day / 24 hours = approximately 1.0GW average power

Thus, inference alone may require an average power of roughly 1 to 2GW.

Therefore, a 3 to 4GW figure for Muse is entirely reasonable. Perhaps this explains the rumors that Meta plans to add 7 to 10GW of new compute next year.

Conclusion

The prevailing narrative regarding Muse is that allocating 2 vCPUs and 8GB of memory per user creates massive CPU demand. However, such simplified calculations significantly exaggerate CPU requirements by treating logical VM allocations as dedicated physical infrastructure.

A more intriguing takeaway is that consumer-grade agents could become a meaningful new demand driver for CPUs and traditional DRAM, but inference remains the true compute bottleneck. Moreover, as agents perform more tasks, execute longer trajectories, and increasingly spawn other agents, inference demand could grow at a pace far exceeding user growth itself.