1.58 million AMD CPUs, 800 petabytes of RAM and 10,000 petabytes of SSD: that is what it takes to give 100 million Muse users the dedicated cloud computer Meta advertises, and the figure moving the processor names since Sunday. The arithmetic holds. The 2 vCPUs, 8 GB and 100 GB under it are not Meta's — four people read them off their own machines between 14 and 21 September.
Meta's own word for the size of the box is 'enough'
Four Meta documents describe it, and none carries a vCPU count, a gigabyte or a chip vendor. The 8 September newsroom post has Muse "on its own dedicated computer in the cloud, contained so no one else's agent can reach it". Meta Superintelligence Labs, that day: "Each Virtual Machine(VM) is an isolated linux box with a browser and enough storage, CPU, and memory to do real work", on a systemd-nspawn Debian image called Hatch — the runtime, not the product tier reported under that name in August.
Four people read the spec off their own screens, and they disagreed
Between 14 and 19 September three X users posted what their box showed: "98GB and 7.5GB RAM on 2 cores", then "2 CPUs, 8 GB RAM, 7.5 GB disk", then "Mine's 100GB disk". At 00:45 UTC on 21 September @tanayj wrote the version the tape kept: "2 vCPUs (a slice of a 126-core AMD EPYC), 8GB RAM, 100GB persistent disk, Ubuntu 24.04".
Fifteen hours later @demian_ai turned it into a chip ladder and closed: "So either a lot of those VMs get frozen when idle, or this does not scale as advertised." Wccftech's write-up that evening carried the ladder, dropped that sentence, and opened: "Meta has promised each Muse user a dedicated VM, replete with 2 vCPUs, 8GB of RAM, and 100GB of SSD." Meta promised the VM, not the three sizes.
A tenth of the load cannot multiply the chip count by ten
Rebuilt from the observed spec: a 126-core host at 2 vCPUs a guest holds 63 machines, so 100m users need 100m × 2 ÷ 126 = 1.59 million chips, 800 PB of RAM (100m × 8 GB) and 10,000 PB of SSD. All three headline figures survive.
Three others do not. Wccftech's body prints 100 PB of RAM where its own inputs give 800; the 8 GB went missing. It prints 15.8 million CPUs for a fleet 10% active, where a tenth of 1.59 million is 159,000 — its 50% case, 0.79 million, is right. Its headline says Ryzen, AMD's client line; its own body says EPYC. And the 4:1 to 40:1 agent CPU-to-GPU ratio credited on X to TrendForce inverts TrendForce's April research: 1:4 to 1:8 today, 1:1 to 1:2 for agentic work.
Today the fleet is about 16,000 chips
The same spec at roughly one million live VMs — @demian_ai's estimate, nobody's disclosure — gives 15,873 chips, 8 PB of RAM and 100 PB of SSD. The circulating figure is 100 times that, and everything between the rungs is a guess at how many run at once.
R40 arithmetic on the user-observed spec; no figure here is a Meta disclosure
| Live VMs | 126-core chips | RAM |
|---|---|---|
| ~1 million (today) | 15,873 | 8 PB |
| 10 million hot | 158,730 | 80 PB |
| 25 million hot | 396,825 | 200 PB |
| 50 million hot | 793,651 | 400 PB |
| 100 million, all hot | 1,587,302 | 800 PB |
Chips are users x 2 vCPUs / 126 cores; RAM is users x 8 GB at 1 PB = 1,000,000 GB. The spec and the 126-core host are user observations posted on X between 14 and 21 September. The 'today' row uses the one-million live-VM estimate in the 21 September post that built the ladder; Meta has published no VM, subscriber or download count.
Meta has published no download, subscriber or live-VM count; trackers count the app: 642,000 US mobile daily users, per Apptopia via TechCrunch. That undercounts machines if the VM persists after the app closes, overcounts them if idle VMs freeze.
$2.4 billion of seats against $20.6 billion of CPUs
One side has published numbers: Meta's price list is free with an unpublished limit, Power at $20 a month for 500 million tokens a week, Maximum at $100 for 3 billion. Ten million Power seats is $2.4 billion a year — 3.9% of the $60,801 million Meta reported for the June quarter, 97.6% of it advertising, and 1.7% of the top of its $130–145 billion capex guide.
The other side has no Meta number at all. At AMD's $12,984 list for a 128-core 9755 — no 126-core part exists, so the guest is likely seeing one with two cores reserved — 1.59 million chips is $20.6 billion of CPUs before RAM and SSD, 14% of that ceiling. Its ~1.6 GW is 23% of the 7 GW Meta is reported to be building this year, sized here in August.
The rest of the tape predates Muse: Lip-Bu Tan's line about serving half his CPU customers came at a 14 September conference, not on a Muse order for Intel; and 800 PB is DDR, not the high-bandwidth memory constraining Micron.
One Meta sentence with a number in it ends this
The best case for the ladder is Meta's own vocabulary: "dedicated", "persistent", a machine that keeps working after people close the app. Against it, the person who built the ladder does not believe Meta will run it that way; their two rungs are 16,000 chips and 1.59 million, with nothing observed between.
A vCPU count, a gigabyte of RAM, a live-VM count or a paid-seat count, in any Meta document, settles it. None exists. Until one does, every chip figure on the tape is a screenshot multiplied by 100 million.
Sources: Meta's own are the 8 September newsroom post, the Meta Superintelligence Labs engineering post on Muse's security, the product page and the subscription help-centre page — every quoted phrase, the runtime and the tiers come from those, none of which carries a size. The spec, the 126-core host, the chip ladder, its closing line and the live-VM estimate are X posts of 14–21 September, attributed to the accounts that made them, neither disclosed nor verified here. The restatement corrected is Wccftech's of 21 September; ratios are TrendForce's April research; Tan's remark is press coverage of a conference; daily users are a third-party tracker via TechCrunch. Revenue, mix and the capex guide are Meta's June-quarter reporting; the gigawatt plan is press-reported. Chip counts, petabytes, dollars and percentages derived here are ours.
