Investors repriced server CPUs on a ratio that spans four to 40 per GPU

The band is ten times wide. Published estimates for agent workloads run from four CPUs per GPU to 40 ^[4], a tenfold spread ^[1]. At four, a 10,000-GPU agent deployment needs 40,000 server processors; at 40, it needs 400,000 ^[2].

CPUs entered the conversation because of how the agent is built. Persistent virtual machines, headless browsers, software calls and data transfers lean heavily on CPUs, while GPUs handle the intensive model calculations ^[7]. Give every user a dedicated cloud machine that keeps working after the app is closed ^[6], and the CPU count scales with the number of users.

The evidence about Muse itself is thinner. Meta's launch materials and Monday's reporting did not specify how much CPU capacity Muse consumes or whose processors run it, and the published ratio estimates were not tied to measured Muse workloads ^[5].

Intel's chief executive, Lip-Bu Tan, said in recent days that his company could supply only half the CPU demand it was seeing from frontier-model companies ^[8]. That demand existed before Sept. 8, and it does not locate a per-user agent inside the band. Arm's position also predates the launch. Earlier in 2026 it struck a partnership with Meta to codevelop CPUs for Meta data centers, and in March 2026 it entered the server CPU market with its first in-house chip for AI data centers ^[9].

Monday offered other reasons to buy chips. Investors were looking toward an expected Trump-Xi state dinner with AI executives, a possible SK Hynix deal with Intel, and Jensen Huang's forecast last week that Nvidia would double chip shipments in 2027. Falling oil prices and Treasury yields supported the broader market, too ^[10]. Intel finished the day worth about $645 billion ^[14] after a 12.14% gain ^[2], roughly $70 billion added in a session ^[4]. Monday's reporting did not split the rally between Muse and everything else ^[11].

AMD's published numbers rest on other business for now. Second-quarter revenue was $11.54 billion, up 50% from $7.69 billion a year earlier, and data-center revenue rose 107% to $6.7 billion; the shares fell after the August report because the forecast missed elevated expectations ^[16]. Meta, AMD's second-biggest customer, accounted for about 5.5% of revenue ^[15], roughly $635 million of that quarter ^[3]. AMD came into Monday on a five-session advance of about 25% and a gain above 180% since the start of 2026 ^[17].

Whether persistent per-user agents become a large computing market will be settled over years, not this quarter. The forecasts behind Monday's prices are long-dated. Lisa Su said in August that AMD expects to double data-center sales in 2027 ^[18], and BofA analyst Vivek Arya last week raised his estimate for the 2030 chip market to $3.2 trillion from $2.7 trillion ^[19]. The same stocks had slumped from a July 2026 peak on concern about excessive AI infrastructure spending, before Muse existed ^[21].