Meta Muse Runs a Separate Computer for Every User, and 642,000 of Them Showed Up in 12 Days
Meta Muse hands every user their own virtual machine, so the fastest-growing consumer AI app is buying CPU, RAM and leased capacity while Meta's own gigawatt campuses are still years out.
The first consumer AI product whose unit cost is a data center line item, shipped while the campuses meant to serve it are still pouring concrete.
Meta Muse launched on September 8. It fills out forms, books dinner, cleans your inbox and pays for things. Meta says it hands every user their own dedicated computer in the cloud, walled off from everyone else's agent. Meta calls it the Muse Secure VM, which is a general-purpose machine with an operating system on it, one per person.
Muse has been popular from the jump. Apptopia, an app analytics firm, counted 2.8 million installs in the first 12 days and 642,000 US daily active users, against the 231,000 ChatGPT had at the same point after its mobile debut. Muse went to number one free on the US App Store. So somewhere in $META's fleet, 642,000 virtual machines a day are booting up, opening a browser and waiting on a human.
What Meta Muse actually eats
A chat turn is a GPU problem. You send tokens, the model sends tokens back, the session ends and the memory gets freed. An agent session is a tenancy. The VM stays alive between your prompts because it's holding your files, your logged-in browser and your half-finished task, so it's burning CPU cores, RAM and storage whether or not the model is thinking. Uncover Alpha, an equity research newsletter, put the free tier at 100 million tokens a week with paid plans at $20 and $100 a month, and estimated $2.6B to $10B a year of inference cost if Muse reaches 100 million monthly actives. Meta has published no figure of its own, and Uncover Alpha's math deliberately leaves out the VMs.
The VMs change what Meta buys. Market commentary on agentic workloads now runs anywhere from 4 to 40 CPUs per GPU, an outside estimate rather than any Meta disclosure, but even the low end is a different bill of materials than a training cluster. Training is bursty, batched and tolerant of latency. A person waiting on an agent to book a flight will notice all three.
Meta's data centers are built for the other workload
Meta guided 2026 capital expenditures to $130B to $145B, narrowed up from $125B to $145B, against $72.2B spent in all of 2025. Second-quarter capex alone was $31.1B, up from $17.0B a year earlier.
But that money is buying training campuses, and training campuses take years. The Index tracks Meta Prometheus in New Albany, Ohio at 1,000 MW planned with 562 MW live today, and Meta Hyperion in Holly Ridge, Louisiana at 5,000 MW planned and still under construction, both on a stated basis as of July. Muse shipped anyway, into capacity that already exists.
So Meta rents. CoreWeave, the GPU cloud that leases capacity to hyperscalers by contract, signed a six-year, $14.2B compute deal with Meta and then expanded it by another $21 billion in April, which puts roughly $35B of Meta's compute on someone else's balance sheet. That was signed before anyone outside Menlo Park had heard of Muse.
Who gets paid
$CRWV is the cleanest answer, and the mechanism is boring on purpose. The Meta commitments sit in CoreWeave's contracted revenue backlog and convert to revenue as capacity gets delivered, quarter by quarter, whether or not Muse holds its ranking. Analysts covering the launch named Nvidia, AMD, Broadcom and Micron as the hardware beneficiaries, and $MU is the one I'd underline, because every one of those 642,000 VMs is memory reserved in a server whether the user is doing anything or not.
The CPU names are the ones the market hasn't repriced yet. $AMD and $INTC sell the general-purpose server silicon that runs a browser, while $NVDA is shipping its own server CPU, Vera, into the same gap. For two years those parts were treated as overhead on a GPU invoice, because the workload that pays for them didn't exist at consumer scale until this month.
Agents are the first AI workload where the boring hardware scales with users instead of with model size. So I'm watching whether Meta's next capex guide splits training from serving. If it does, 642,000 is the number that made them do it.