Meta Muse Puts CPUs Back in the AI Spotlight as Agentic Workloads Expand
Meta's new Muse personal AI agent is drawing fresh attention to a part of AI infrastructure that has spent much of the current boom outside the spotlight: the CPU. Unlike a conventional chatbot that primarily sends prompts to an inference model and returns a response, Muse can operate a persistent cloud computer, navigate websites, manage applications, fill forms, organize information, and continue working after the user leaves the app. Those workloads still rely on GPU acceleration for AI inference, but they also create substantial demand for conventional server compute to run operating systems, browsers, APIs, memory management, orchestration, and background processes.
Meta explains that each Muse user can be assigned a dedicated Muse Secure VM containing its own browser and environment. At large scale, that model effectively means operating huge numbers of virtual computers alongside the accelerators serving Meta's AI models. This is why agentic AI is beginning to change the infrastructure conversation. GPUs remain essential for running increasingly capable models, but an agent that performs dozens of actions across web services can generate far more CPU activity around every inference request than a straightforward text conversation. Goldman Sachs Research similarly describes consumer agents as the beginning of a shift from conversational AI toward systems that actually execute increasingly complex tasks, while expecting AI infrastructure spending to remain elevated through 2027 as compute demand continues exceeding supply.
Some market commentary now places potential CPU to GPU requirements for agentic workloads anywhere from around 4 CPUs per GPU to as high as 40 CPUs per GPU. That upper figure has attracted significant attention, but it needs important context. Meta has not disclosed that Muse operates at a 40 to 1 ratio, and the figure should not be interpreted as an official architecture target for Meta data centers or as a universal requirement for AI agents. Different agents can have radically different compute profiles depending on how many virtual machines they use, how long those environments remain active, which models handle inference, and how much browser automation, data processing, and tool execution takes place around each task.
One of the things I shared 😱
— Damnang (@damnang2) September 19, 2026
CPU LFG!!! https://t.co/yyrnmbN1ty pic.twitter.com/zFAyM40YCa
The underlying CPU demand trend is nevertheless becoming increasingly difficult to ignore. Intel CEO Lip Bu Tan recently said that current demand has reached a point where the company can fulfill only around half of customer requests, although his comments did not specify that all of this demand comes from agentic AI or identify the exact processor segments involved. The pressure also aligns with what AMD and Arm have been communicating throughout 2026. AMD CEO Lisa Su described agentic AI CPU growth as additive to accelerator demand rather than replacing GPUs, while Arm reported more than $2 billion in customer demand for its AGI CPU across fiscal 2027 and 2028.
"CPU demand is so high that we can only supply 50% of customers."
— Quote by: Lip Bu Tan
This shift could have implications well beyond Intel and AMD. Arm based server processors are already deeply established across hyperscale cloud platforms, while NVIDIA is expanding from accelerator attached CPUs into standalone server products with Vera. Qualcomm has also confirmed work on data center CPUs, creating an increasingly crowded market around a component that only a few years ago appeared likely to become secondary to enormous GPU clusters. We previously examined this transition through NVIDIA Vera's growing challenge to Xeon and EPYC and the broader push toward CPU architectures designed specifically around agentic infrastructure.
Muse does not suddenly mean AI data centers need 40 CPUs for every GPU. That number is an aggressive external estimate, not a Meta specification, and treating it as a fixed infrastructure ratio would oversimplify how agentic systems actually work. What Muse does demonstrate much more convincingly is why the CPU is becoming important again. Every persistent browser, virtual machine, operating system instance, API request, tool call, memory operation, and background process needs general purpose compute even when a GPU or another accelerator handles the underlying AI model.
The emerging infrastructure model therefore looks less like CPUs replacing GPUs and more like AI systems requiring much more of both. Training created extraordinary accelerator demand. Inference expanded it further. Persistent autonomous agents now add another compute layer around those models, and that could become a substantial opportunity for Intel, AMD, Arm, NVIDIA, and other server CPU suppliers if consumer and enterprise agents reach the scale companies such as Meta are targeting.
Do you think agentic AI will make CPUs as strategically important to future data centers as GPUs, or will accelerators continue dominating AI infrastructure spending?
