To The Moon

AI may be moving from “answering questions” to actually doing work — and that could make CPUs important again.

The latest AI trade is no longer just about GPUs.

Meta’s new AI agent, Muse, has pushed investors to rethink what the next phase of AI infrastructure may actually require.

Why?Because an AI agent does much more than generate an answer.

It may need to:open a browser,search the web,fill out forms,call APIs,run tools,manage files,and keep working in the background for minutes or even hours.

That changes the compute equation.

For a chatbot, the workflow is relatively simple:

Prompt → GPU inference → Response

For an AI agent, it looks more like:

Think → Browse → Execute → Check → Think Again → Continue

GPUs still handle the heavy model inference.

But many of the surrounding workloads — browser environments, operating systems, task scheduling, API calls, databases and background processes — still depend heavily on CPUs.

That is why the market is starting to ask a new question:

If every user eventually has an AI agent working for them 24/7, how much CPU capacity will that require?

CPUs may be coming back into the AI infrastructure story

Over the past two years, the AI hardware narrative has been dominated by GPUs.

Then came HBM.

Then power.

Then optical networking.

Now AI agents may be expanding the bottleneck again.

As AI moves from simply “thinking” to actually executing tasks, the infrastructure stack becomes much broader.

It is no longer enough to ask how many GPUs a data center has.

You also have to ask:

How many tasks can the system actually run at the same time?

That makes CPUs, networking, storage and software orchestration increasingly important.

In simple terms:

GPU = thinking
CPU = execution

And the more autonomous agents become, the more valuable that execution layer may become.

That helps explain why names such as:

$INTC
$AMD
$ARM
$QCOM

are suddenly back in the AI conversation.

And then there is Tencent

The China-side read-through is also interesting.

If Meta has Muse, one obvious question is:

Who has the strongest ecosystem to build a similar agent in China?

Tencent immediately comes to mind.

Not necessarily because it already has an identical product, but because an AI agent becomes much more powerful when it can access real-world services.

Tencent already sits on top of a huge ecosystem:

WeChat
Mini Programs
WeChat Pay
Enterprise WeChat
Tencent Docs
Cloud infrastructure
and a massive network of merchants and services.

That matters because the real value of an AI agent may not be answering:

“What restaurant should I go to?”

The more important capability is:

Find a restaurant → compare options → ask friends → book it → pay → add it to the calendar

That is where ecosystem matters.

The model provides intelligence.

The ecosystem provides action.

And that could become one of the biggest competitive advantages in the agent era.

Tiger View

The biggest takeaway here is not that CPUs suddenly replace GPUs.

They do not.

The more important shift is that the definition of “AI compute” is expanding.

The last phase of the AI trade was about:

How many GPUs does it take to train and run models?

The next phase may be about:

How much infrastructure does it take to keep billions of AI agents working continuously?

That requires an entire stack:

GPUs for inference
CPUs for execution
HBM for data
Optical networking for connectivity
Storage for memory
Power for everything

That is why the CPU trade and the optical-interconnect trade are not separate stories.

They are part of the same shift:

AI is evolving from a model into an operating layer for the digital world.

And if AI agents really do become mainstream, the next winners may not just be the companies building the smartest models.

They may also be the companies providing the infrastructure — and the ecosystems — that let those agents actually get things done.

$Meta Platforms, Inc.(META)$ $Intel(INTC)$ $Advanced Micro Devices(AMD)$ $ARM Holdings(ARM)$ $Qualcomm(QCOM)$ $Tencent Holding Ltd.(TCEHY)$