Meta’s new consumer AI agent, Muse, could need somewhere between one and four gigawatts of average power to serve 100 million daily users, according to a back of envelope model published by an analyst writing under the byline FD on the Substack newsletter Robonomics. The estimate is not a figure Meta has confirmed. It is one outsider’s attempt to reverse engineer the economics of running an agent at consumer scale, built from public specifications, a rival lab’s published infrastructure paper, and a Microsoft energy study.
The stakes are concrete. Power, not chips, has become the binding constraint on AI buildouts in 2026, and any credible estimate of how much a single consumer product could draw matters to every utility, data center developer, and rival lab watching Meta’s next move.
The analyst starts from Muse’s known virtual machine configuration: 2 virtual CPUs, about 8GB of RAM, and roughly 100GB of persistent storage per user, drawn from a third party teardown rather than an official Meta disclosure. The model assumes each active user keeps an agent running for two hours a day, then layers in two adjustments before reaching a peak figure: usage clusters 2.5 times above that daily average during waking hours, and the buildout carries 20 percent of headroom on top of that. Run through those steps, the math lands on roughly 25 million virtual machines live at once during peak hours.
For a comparison point, the analyst cites DeepSeek’s published DSec infrastructure paper: its production agent sandbox spreads 160 nodes across roughly 30,000 physical CPU cores paired with 250TB of DRAM, handling more than 380,000 concurrent sandboxes at peak. Divide those cores by the sandbox count and DeepSeek is running each live virtual machine on about a quarter of one physical core, 0.23 to be precise. Since DeepSeek is treated as the efficiency benchmark, the analyst doubles that figure for Muse, penciling in half a physical core for every live machine as the base case. Multiplied out, that puts the fleet at about 12.5 million physical CPU cores, which translates to somewhere between 50,000 and 65,000 server CPUs and roughly $800 million in hardware at public list prices, plus nearly $2 billion more in DRAM. The entire sandbox layer, on this accounting, draws about 0.1 gigawatts, a smaller number than the analyst says many expected.
A 2026 Microsoft study supplies the inference side of the math: it puts optimized frontier inference at roughly 0.31 watt hours for an ordinary query and about 4 watt hours for a long reasoning query. Building on that, the analyst’s working assumption is that a single Muse user racks up 50 of these “heavy reasoning” events daily, each costing 5 to 10 watt hours. Scaled across the user base, that alone produces roughly 25 gigawatt hours a day, or about one gigawatt of continuous average power, before any growth in how many such calls each agent generates is factored in.
The conclusion cuts against the popular framing that giving every user a virtual machine is what makes agent products expensive to run. The sandbox layer is a rounding error next to the inference bill. As agents take on longer tasks and increasingly spawn other agents to complete them, the number of model calls per user, not the number of users, becomes the variable that decides how much power a product like Muse actually needs.
That distinction should reset how operators read every “X million users” headline about consumer AI agents this year. A user count says almost nothing about power demand or GPU procurement without knowing how many reasoning steps that agent burns per session, and any capacity plan built only on daily active users is likely to miss the real bottleneck by an order of magnitude.
Reported by Robonomics, a Substack newsletter on AI infrastructure economics, on 27 September 2026.
