Personal AI agents that can browse websites, manage apps and complete multi step tasks are emerging as a new source of data center demand, with some estimates suggesting they could require far more CPUs than GPUs.
The idea is gaining attention as services such as Meta's Muse and Spear Street Technology's Instinct rely on persistent cloud computers that stay active while carrying out tasks on behalf of their owners.
Unlike large AI model training workloads, which lean heavily on GPUs, these agents also depend on ordinary computing work such as browser automation, operating system management, API calls and background processes. Those tasks are much more CPU intensive.
Persistent cloud computers change the hardware mix
Meta's Muse is designed to operate through a dedicated cloud based computer that includes its own browser and isolated environment.
The agent can navigate websites, manage workflows and continue working even after the main app is closed.
That means the cloud environment itself remains active and continues consuming CPU resources.
| Area | Why CPUs matter |
|---|---|
| Virtual machines | Run operating systems and background services |
| Browser automation | Handles websites and page interaction |
| API calls | Moves data between services |
| Email and calendar tasks | Process large volumes of structured data |
| Long running agents | Require persistent compute resources |
| GPUs | Still important for AI model inference |
Some estimates place the eventual CPU to GPU requirement anywhere between 4 to 1 and 40 to 1 for certain agent based workloads.
Those figures remain estimates rather than established industry standards, but they show why chipmakers are paying closer attention to this class of AI software.
Muse combines automation with security controls
Muse is designed to carry out complex tasks through an isolated cloud environment known as a Confidential Virtual Machine.
The system keeps browser activity, credentials and other data inside that environment.
A separate monitoring layer, referred to as a Sentinel, checks what the agent is doing and is intended to prevent transactions from being completed without additional approval.
For sensitive actions, the system requires both internal approval and explicit confirmation from the person using it.
Meta is also expanding connectors that allow Muse to interact with more third party services.
As more connectors are added, the number of tasks that can be automated increases, which could also raise the amount of compute required.
AI agents can generate large amounts of background activity
The CPU argument becomes clearer when looking at the number of actions a single request can create.
One example cited in the report suggests that a complex itinerary search could trigger as many as 146 individual searches.
An agent may also need to compare prices, open several websites, parse results, check calendars, communicate with APIs and keep its cloud session active throughout the process.
These are not purely GPU inference workloads.
The AI model may decide what to do, but CPUs are responsible for much of the software execution surrounding those decisions.
That is very different from earlier AI infrastructure discussions that focused mainly on accelerator demand.
Other personal agents are following a similar model
Muse is not the only service built around persistent cloud automation.
Instinct uses a different interface, allowing people to communicate with the agent through messages and calls rather than a traditional application.
It can connect to services such as email, calendars and messaging accounts while using a cloud computer to perform tasks.
Examples include booking or changing flights, making restaurant reservations, arranging appointments and purchasing everyday items.
The underlying pattern remains similar.
A persistent cloud environment needs processors to manage browser sessions, operating systems, storage, network activity and application logic for every active agent.
CPU makers could benefit if personal agents scale
If this model becomes widely adopted, Intel, AMD and Arm based server platforms could all benefit from higher demand.
The scale of that opportunity remains uncertain.
Claims of CPU to GPU ratios reaching 40 to 1 represent an aggressive upper end rather than a guaranteed outcome.
The actual ratio would depend on how efficiently agents share infrastructure, how much work stays active in the background and how much of each task is handled by CPUs versus accelerators.
Still, the direction is important.
AI workloads are expanding beyond model training and inference into ordinary software automation, where CPUs remain essential.
Supply pressure is already being discussed
Intel CEO Lip Bu Tan recently said the company could currently satisfy only about half of the CPU demand coming from frontier AI companies.
That comment does not prove that personal agents alone are causing a shortage, but it suggests that AI companies are already consuming more conventional processor capacity than many expected.
Capital spending patterns are also beginning to reflect greater interest in CPU infrastructure alongside GPUs.
The broader shift is therefore not that CPUs are replacing GPUs.
Instead, AI systems are becoming more complex and require both.
GPUs handle model execution, while CPUs manage the large amount of surrounding software and cloud infrastructure.
Personal agents could become a major server workload
The most important change is architectural.
Earlier AI growth was dominated by demand for large GPU clusters.
Personal agents introduce a second layer where potentially millions of persistent cloud computers may operate continuously in the background.
Each one needs CPU time, memory and networking even when the AI model itself is not actively generating a response.
If services such as Muse and Instinct continue expanding, server demand could therefore move toward a more balanced mix of accelerators and traditional processors.
That could give Intel, AMD and Arm a larger role in the next phase of AI infrastructure, although predictions such as a 40 to 1 CPU to GPU ratio should still be treated as estimates rather than settled expectations.
Discussion (0)
Be the first to comment.