META × AMZN / RESEARCH MODEL

Muse Compute Split

Duty-cycle CPU + token burn + VM persistence, with explicit sliders for how much of the runtime lands on AWS.

How this works in 30 seconds

Set users and how hard they use the agent.
The model splits cost into tokens, VMs/CPU, and a small residual (safety, connectors, power markup).
AWS sliders allocate only the runtime/CPU and optional token overflow. They do not assume Muse Spark is hosted on Bedrock.
Compare the Meta subsidy with subscriptions and a hypothetical commerce take-rate.

Muse has two expensive layers. The model burns tokens. The agent also occupies a cloud computer — a browser, disk, and CPU — even when the user is not staring at the app. Published models disagree on whether that computer is mostly idle (duty-cycle) or mostly reserved (always-on). This page lets you sit between them.

Amazon’s disclosed relationship is a large Graviton CPU commitment for agentic work, plus a longer AWS/Bedrock relationship. That is not the same thing as “Muse runs on Bedrock.” The AWS sliders are labeled assumptions so you can test both the conservative and the aggressive read.

Meta all-in annual cost

$2.70B

40.0M actives · 2.1% concurrency

Cost per active / month

$5.62

Outside the $3.50–$4.50 cross-check

Token layer vs VM/CPU

$1.91B

CPU $262M · memory $175M · power $87.3M · residual $240M

Estimated AWS annual bill

$420M

15.6% of Meta Muse compute

Graviton $183MEC2 overflow $71.5MBedrock $166M

Subscription coverage

10.7%

$288M / year from 800.0k paid

Commerce take to break even

40.2%

After subs · GMV $6.00B · take now $120M

Short term, AWS collects a contracted CPU bill and a small token assumption. Meta funds the subsidy. Longer term, Meta needs commerce or a cheaper token factory. Amazon still collects whatever CPU overflows the Graviton commitment.

5 tasks × 6 min. Page default: hybrid VMs, not the $1.3B processor-only line.

Scale

Registered users (people)ASSUME100.0MActives = registered × active rate.Active rate (share of registered)ASSUME40.0%Paid mix (share of actives)ASSUME2.0%Paid ARPU ($ / month)META$30Blend of the $20 Power and $100 Maximum plans.

40.0M actives · 800.0k paid · $288M subscription revenue / year.

Intensity

Tasks / active / day (tasks)BEP5.0Minutes / task (minutes)BEP6.0Peak multiplier (× average)BEP3.0×Tokens in / task (tokens)BEP200.0kTokens out / task (tokens)BEP10.0kCache hit rate (share of input)ASSUME40.0%

Live concurrency 2.08% = tasks × minutes / 1,440. 7.3M tokens / active / week · 15.33 quadrillion tokens / year.

Token pricing

Internal cost is who-pays. List price is API value given away. UA

Internal $ / 1M blended ($ / million tokens)UA$0.12Uncover band is $0.10–$0.20. Contributor $0.10 in / $0.20 out is the public ceiling, not the default.Standard input ($ / 1M)META$1.25Observation $1.25. Override only to test the list-price ghost.Cached input ($ / 1M)META$0.15Observation $0.15.Standard output ($ / 1M)META$4.25Observation $4.25.

Selected token bill $1.84B · list-price ghost $14.93B · contributor-shaped ceiling $1.03B. Internal mode applies the blended rate to every token. Cache only changes the list-price ghost.

VM / CPU hardware

VM persistence. Duty-cycle and always-on are two views of the same layer. They are not added.

Hardware dollars. Full TCO uses the nextsignal schedule on this mode’s machine count, split into CPU and memory. Processor-only prices sockets: the BEP thread formula on duty-cycle, or the fleet’s vCPUs on always-on and hybrid. Processor-only excludes memory.

Warm share of actives (share kept hot)ASSUME15.0%Hybrid only. Default 15%.vCPU / VM (vCPU)NSP2RAM / VM (GB)NSP8Processor $ / 256-thread socket ($ list)BEP$10,931EPYC 9755 list about $10,931. Excludes memory, chassis, power, and ops.CPU utilization (share)BEP70.0%Hardware $ / 100M hot VMs ($ capex)NSP$40BPublished band $30–50B. Default $40B.Useful life (years)NSP5.55.5 years matches nextsignal. Slide to 3 for BEP’s “fleet built over three years.”Facility power markup (share of depreciation)ASSUME20.0%NSP also cites ~1.60 kW IT and PUE 1.3. Dollars here are a percent of depreciation, not a tariff.Safety / connectors ($ / active / month)AA84$0.50

Full VM TCO · 6.0M VM-equivalents · capex $2.40B · CPU $262M/yr · memory $175M/yr · power $87.3M/yr · ~0.10 GW scaled from the 1.63 GW / 100M VM observation. An 8 vCPU box is 2.8× this VM on a 60/40 CPU/memory split of the nextsignal schedule ASSUME. BEP’s ~$1.3B processor line is 1B users, duty-cycle, 3-year life, before power — not this default.

AWS allocation

Graviton dollars are EC2/custom-silicon dollars. Bedrock dollars are managed-inference dollars. This calculator will not collapse them. Default Bedrock share is 5% because Muse Spark is sold on Meta’s own API. Raise it only if you believe Meta will overflow tokens onto Bedrock or call third-party models there.

Graviton / agentic CPU AWS ASSUME

Graviton share of VM/CPU $ (share)ASSUME35.0%April 2026 deal is “tens of millions of cores” for agentic CPU. BEP Base at 1B users is about 45M cores. Default is not “all Muse VMs live at AWS.”Graviton $ / core-year ($ / core-year)ASSUME$80Back-solved from “multi-billion over several years” ÷ tens of millions of cores. Contract terms are undisclosed.Committed Graviton cores (cores)AWS30.0MObservation: “tens of millions.”

Cap Graviton dollars at committed cores × $ / core-year ASSUME

General AWS CPU overflow ASSUME

Overflow to general AWS EC2 (share of non-Graviton CPU)ASSUME15.0%EC2 premium vs owned (multiple)ASSUME1.40×

Bedrock token overflow ASSUME

Most speculative slider on the page. Muse Spark is first-party. Bedrock is more plausible for routing, fallback models, Llama-family calls, or enterprise-adjacent features than for the core consumer agent. Default is low on purpose.

Bedrock share of tokens (share)ASSUME5.0%Bedrock $ / 1M vs Meta internal (multiple)ASSUME1.8×Managed inference. Expect a stack markup if it is used at all.

Bedrock mix

Who pays the subsidy

Commerce GMV / active / year ($)ASSUME$200Meta take-rate (share of GMV)ASSUME2.0%

Amazon blocks Muse shopping on Amazon.com — cut GMV 25%. Other merchants are not blocked. ASSUME

Uncovered after subs and commerce: $2.29B. No ads inside Muse at launch META.

Cost bridge

List-price token value (ghost)

$14.93B

Meta token spend, after Bedrock markup

$1.91B

CPU depreciation

$262M

Memory depreciation

$175M

Power

$87.3M

EC2 premium above owned CPU

$20.4M

VM layer

$544M

Residual safety / connectors

$240M

Meta all-in

$2.70B

of which AWS Graviton

$183M

of which AWS EC2 overflow

$71.5M

of which Bedrock (AWS portion)

$166M

AWS total

$420M

Meta-owned remainder

$2.28B

Coverage

Subscriptions$288M

Commerce take$120M

Uncovered subsidy$2.29B

Per active user / month

Inference$3.99

CPU$0.55

Memory$0.36

Power$0.18

VM layer$1.13

Residual$0.50

Total$5.62

alphaseeker84 published about $3.50–$4.50, with inference near $1.40 and VM near $1.60. This tile turns amber outside that band. Plans on the page are $20 and $100. Wide sanity flag is under $2.50 or over $6.00.

Users × intensity

Meta $B · AWS $B

LightBaseHeavy30.0M actives$1.2B$241M AWS$2.0B$315M AWS$3.7B$567M AWS100.0M actives$3.9B$802M AWS$6.7B$1.1B AWS$12.4B$1.9B AWS250.0M actives$9.7B$2.0B AWS$16.9B$2.6B AWS$31.1B$4.7B AWS1.00B actives$38.9B$6.3B AWS$67.6B$8.8B AWS$124.6B$14.8B AWS

Agentic cores

Implied cores (BEP math)

1.8M

14.0k sockets · util 70%

Committed Graviton cores

30.0M

Observation, not a Muse-only figure

Gap

-28.2M

Commitment covers the duty-cycle need

Share of 120M cores / GW

0.01

Haas rule of thumb, via BEP

Sources

BEP Research, “Meta’s Muse Rally Bought the Wrong Layer,” 21 Sep 2026 — duty-cycle CPU, 1B-user Light/Base/Heavy, 200k/10k tokens per task, EPYC 9755 list price.
nextsignalprediction, “What It Takes to Give Every Personal Agent a Computer,” 23 Sep 2026 — 2 vCPU/8 GB VM, 100M-user hardware/power/depreciation.
alphaseeker84, “Muse Is a Memory Story…,” 21 Sep 2026 — $3.50–$4.50 cost to serve, inference vs VM split, memory-bound view.
Uncover Alpha, Muse launch economics, 8 Sep 2026 — free-tier token value, 100M MAU token volumes, $0.10–$0.20 internal cost band.
Meta Model API pricing and Muse launch posts, Jul–Sep 2026 — list prices, $20/$100 plans, Secure VM architecture.
Meta and AWS Graviton announcements, 24 Apr 2026 — tens of millions of Graviton cores, multi-year multi-billion deal, agentic CPU rationale.
Launch coverage (Axios, TechCrunch, NYT), 8 Sep 2026 — product behavior, paid tiers, no ads in Muse.
Subsequent reporting that Amazon blocked Muse on Amazon.com — used only for the commerce toggle.

Figures above are adapted from those essays and disclosures. This is not their official calculator. The $20 Power plan is reported at 500.0M Muse tokens/week and the $100 Maximum plan at 3.00B / week. Contributor API observations, unused as the default bill: $0.10 in / $0.002 cached / $0.20 out per 1M.

This is a simplified research model stitched from public essays and company disclosures. It is not Meta’s cost accounting and not Amazon’s revenue recognition. Contract pricing for Graviton and any Bedrock use is undisclosed. Do not treat outputs as forecasts. Not investment advice.