Meta Muse's 2,000-tray fleet estimate rests on one hypothetical server and rough napkin math
User tests found Meta's Muse gives each user a persistent sandbox with two vCPUs and about 8GB of memory. The 2,000-tray fleet figure built on it divides a disputed 500,000-user count by one hypothetical server.
The Engineer · Build desk
What happened
- Blogger Evan Hoffman and analyst Tae Kim had Muse report its own environment, which showed AMD EPYC 9D25 hosts running Ubuntu 24.04 on Linux kernel 7.0.
- The hosts have no GPUs; Tom's Hardware reported that Meta runs inference on separate GPU servers and keeps the agent sandboxes CPU-only.
- A separate hands-on test measured one Muse instance at 7.7GB of RAM and 100GB of persistent storage.
Why it matters
- What doesn't add up contradiction The divisor itself is unsettled: Tom's Hardware cites 500,000 daily active users while The Decoder reports 500,000 users in the first week, and a fleet sized to one population is wrong for the other.
- What it limits constraint At two vCPUs per sandbox, 512 vCPUs cap the assumed box at 256 users however much memory is installed, so only more cores or a smaller CPU share per user would raise density.
- What it costs cost Persistent disks are a cost the tray count leaves out: 100GB for each of 500,000 users is 50 million GB, about 50 petabytes of storage.
- What's now possible capability Because Meta built the Runtime Cell to be inspectable, any user can rerun the same commands, so a change to per-user allocation would surface without Meta announcing it.
Next story · Invest · Oct 6
AMD stock surges as agentic AI shifts workloads from GPUs to CPUs, powering Meta's Muse and DotsConfirmed2 publishers
The spec sheet came from the agent. Hoffman and Kim got it by prompting Muse to run Ubuntu commands and pass back the output [4]. That only works because Meta built the Runtime Cell to be inspected. David Singleton, VP of Engineering at Meta Superintelligence Labs, says the transparency was deliberate, and users can see every file from Debian system files to the Muse binaries [19]. The cell has its own root filesystem, while a Sentinel process outside it watches sensitive actions and credentials are stored outside as well [18]. For a machine users are meant to open up, I think the split is sensible. The cell still has limits. A command to query the kernel buffer failed on permissions [20]. Hoffman also said Muse offered to set up SSH into its private VM [21].
Meta's September 8 launch announcement describes a dedicated cloud computer for each user, and Meta has not disclosed per-user specifications, server density or fleet capacity [14][15]. John Gruber wrote that Muse is "groundbreaking technically (each user gets their own entire persistent Linux VM running in Meta's cloud)" [6].
The Tom's Hardware estimate sits on one assumed box, and that box runs out of CPU and memory at exactly the same point. Two vCPUs per sandbox into 512 vCPUs is 256. Eight gigabytes into 2,048GB is also 256, because the server and the sandbox both work out to 4GB per vCPU [22]. Dividing 500,000 users by 256 gives about 1,953 trays, rounded to 2,000 [8]. "This is just some rough napkin math; don't take it as law," Jake Roach wrote in the Tom's Hardware report [9]. He allowed that Meta may run several chip types and carry configuration overhead [10].
The word "core" changes the answer by a factor of two. The Tom's Hardware headline gives each sandbox two dedicated cores, while its server math counts vCPUs [3][12]. RuntimeWire describes the allocation as two vCPUs [2]. The EPYC 9D25 has up to 128 cores, so a two-socket box carries 256 physical cores and 512 vCPUs is two per core [25]. If each sandbox holds two physical cores, one box fits 128 users. At that density, 500,000 users need about 3,906 trays [26].
In my view the per-user shape is the more useful number. Two vCPUs, about 8GB of memory and 100GB of disk is what users measured inside Muse [2][13]. ChatGPT's Work mode reportedly runs a VM with 15GB of RAM and more than nine CPU cores, according to The Decoder [11]. Its memory is nearly twice Muse's [24]. Muse has a free tier, and Stark Insider reported paid Power and Maximum plans at $20 and $100 a month [16]. According to RuntimeWire, neither the allocation nor the tray estimate reveals what Meta actually spends to run the service [17].
What to watch
- Meta publishing per-user VM specifications, server density or a fleet figure for Muse.
- A first-party Muse usage figure that settles whether 500,000 is daily active users or cumulative first-week users.
- Repeat in-VM measurements showing a different vCPU, memory or disk allocation, especially across the free, Power and Maximum tiers.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence45
- Adoption55
- Hype gap+30
- Incentives40
- Confidence50
Perspective Coverage
4 publishers- Builder
- Builder 39%
- Operator
- Operator 35%
- Investor
- Investor 26%
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Each user appears to get their own private, persistent sandbox.
- [2]
User tests of Meta's Muse agent found sandboxes with two vCPUs and about 8GB of memory.
- [3]
Tom's Hardware reported that Muse runs on AMD EPYC Turin hosts, with each sandbox having two dedicated cores and 8GB of memory.
- [4]
Blogger Evan Hoffman and analyst Tae Kim found that Muse will run rudimentary Ubuntu commands if prompted and pass along the output, helping identify the host system's specs.
- [5]
Asked about the VM's specs, Muse revealed it runs on AMD EPYC 9D25 CPUs, a high-density Turin chip with up to 128 cores, with VMs on Ubuntu 24.04 using Linux kernel 7.0.
- [6]
"groundbreaking technically (each user gets their own entire persistent Linux VM running in Meta's cloud)"
- [7]
The systems hosting Muse do not include GPUs; Meta uses separate GPU servers for inference, isolating the agent to CPU-only sandboxes.
- [8]
The Tom's Hardware estimate works out to about 1,953 trays at 256 users per tray, rounded to roughly 2,000.
- [9]
"This is just some rough napkin math; don't take it as law."
- [10]
Meta may have CPU-only servers deployed with multiple different chips to host Muse, and there may be overhead in the configuration; Turin chips support up to 6TB of memory with high-density DIMMs.
- [11]
Users discovered that ChatGPT's Work mode runs its own virtual machine with 15GB of RAM and over nine CPU cores.
- [12]
Tom's Hardware assumed a two-socket system offering up to 512 vCPUs with 2TB of memory, hosting up to 256 Muse users.
- [13]
A separate hands-on test measured one Muse instance at 7.7GB of RAM and 100GB of persistent storage; Meta has not published or guaranteed those specifications.
- [14]
Meta's September 8 launch announcement describes Muse as running on a dedicated cloud computer but gives no per-user CPU or memory allocation.
- [15]
Meta has not disclosed per-user VM specifications, server density or fleet capacity.
- [16]
Muse has a free tier; Stark Insider reported paid Power and Maximum plans at $20 and $100 a month respectively.
- [17]
Neither the reported sandbox allocation nor the server estimate reveals Meta's actual cost to run the service.
- [18]
According to David Singleton, users and their Muse agent work inside a Runtime Cell with its own root filesystem separate from the host; a Sentinel process outside the cell monitors sensitive actions, and passwords and credentials are stored outside the cell.
- [19]
Singleton says Meta built full transparency into the Runtime Cell on purpose; users can see every file, from Debian system files to the binaries that run Muse inside the cell.
- [20]
Hoffman shared an example where an attempted command failed due to improper permissions when Muse tried to query the kernel buffer.
- [21]
Hoffman claimed that Muse offered to set up SSH into Muse's private VM.
- [22]
The assumed server hits its CPU and memory limits at the same 256 sandboxes, since both the server and the sandbox are 4GB per vCPU.
- [23]
At two vCPUs per sandbox, a 512-vCPU box caps at 256 sandboxes regardless of installed memory; memory beyond 2,048GB adds no sandboxes.
- [24]
ChatGPT Work's reported 15GB of RAM is nearly twice Muse's roughly 8GB.
- [25]
A two-socket EPYC 9D25 box has 256 physical cores, so 512 vCPUs is two vCPUs per physical core.
- [26]
If each sandbox holds two physical cores, one box fits 128 sandboxes and 500,000 users need about 3,906 trays, double the headline estimate.
- [27]
100GB of persistent storage for each of 500,000 users is 50 million GB, about 50 petabytes.
- [28]
In its September 25 report, Tom's Hardware said Muse had reportedly passed 500,000 daily active users, citing a third-party figure.
- [29]
The Muse app pulled in over 500,000 users in its first week and hit number one in the Apple App Store.
Sources
3 independent publishers whose own reporting we read for this story.
- runtimewire.comMeta's Muse may need 2,000 server trays at reported user scale
1 article · September 26, 2026
- simonwillison.netQuoting John Gruber
1 article · September 25, 2026
- the-decoder.comMeta's Muse agent gives every user a full cloud computer running Ubuntu Linux
1 article · September 25, 2026
- tomshardware.comMeta Muse runs agents on AMD EPYC Turin hosts with two cores and 8GB of memory
1 article · September 25, 2026
Topics and entities
Topics
- AI Agents
- Code-agent sandboxing
- Cloud compute capacity