Was reading through a few things this weekend and found this pretty interesting.
Everyone has obviously been focused on GPUs for the past few years, but CPUs could end up being a much bigger part of the AI trade than people thought.
BofA recently raised its estimate for the server CPU market in 2030 from around $170B to over $210B, compared with only around $35B in 2025.
That’s a massive jump. The reason is AI agents.
For AI training, you needed roughly one CPU for every four GPUs. BofA thinks with AI agents that could eventually get closer to 1:1.
And when you actually think about what these new agents are doing, it makes sense.
Instead of just asking an AI a question and getting an answer back, these agents can actually do things for you. They can browse the internet, execute code, access databases, use different tools and basically operate a computer on your behalf.
All of that requires CPUs.
Meta’s new Muse is a pretty good example of where this could be going.
Muse is basically a personal AI agent that can actually perform tasks for you. But the part I found really interesting is that every Muse runs inside its own dedicated secure virtual machine in the cloud, with its own browser.
Basically, every agent gets its own little computer to work from.
Now imagine where this goes over the next few years.
Instead of someone opening ChatGPT and asking it a question, you could eventually have billions of people with agents constantly working in the background.
And I doubt it stops at one agent per person.
You could have one doing research, another shopping, another handling emails, another working with files, etc., all running at the same time.
That is a ridiculous amount of compute.
And this lines up pretty well with what $AMD has been saying recently.
AMD has talked about seeing more CPU demand from agentic AI because all of these agents need CPUs for things like orchestration, database access and actually using tools.
So I started looking at which names could benefit the most from this.
$AMD is probably the obvious one.
BofA has AMD as its preferred CPU name, and AMD itself is already talking about increased CPU demand coming from AI agents.
Then there’s $ARM, which I think is pretty interesting too.
A lot of the next generation of server CPUs are being built on Arm architecture, including Nvidia’s Vera CPU.
And funny enough, there was some interesting flow in $ARM on Friday too.
804 $ARM 380P 3/19/27 were sold at an average fill of $15.40, around $1.2M in premium.
I like those put sales. They’re pretty far out and a decent amount of premium. Obviously doesn’t mean the stock has to go higher, but I like seeing that kind of trade in a name I’m already looking at.
Then you have $NVDA.
I don’t think CPUs doing well means GPUs suddenly become less important. If anything, I think the better way to look at this is that AI is just going to require a LOT more compute than people originally thought.
Nvidia can still benefit from that. BofA still has it as its top overall semiconductor pick, and Nvidia’s Vera CPU is itself built on Arm architecture.
And then there’s $INTC.
This one is interesting for a couple reasons.
BofA specifically mentioned Intel’s foundry opportunity. If CPU demand actually gets anywhere near these numbers, Intel doesn’t have to suddenly take over the CPU market for it to matter.
And there was some pretty good flow in $INTC on Friday.
3,520 $INTC 95P 2/19/27 were sold at an average fill of $9.95, around $3.5M in premium.
There were also:
500 $INTC 150C 6/17/27 bought at $13.50, around $675K in premium.
I like those 95P sales.
That’s $3.5M in premium on puts going all the way out to February. Someone is clearly comfortable taking the other side down there.
So I don’t really look at this as CPUs replacing GPUs.
I think the easier way to look at it is that the AI compute pie could just get much bigger.
First we had training.
Then inference.
Now we’re moving into agents that can actually use computers and do work.
If we eventually have billions of these things running around in the cloud, potentially multiple agents for every person, that’s going to require an insane amount of infrastructure.
And the timing of all of this is interesting because sentiment around AI has completely changed over the past few months.
Citadel Securities put out a good note this week.
Retail semiconductor activity on their platform is now 46% below the June peak, while opening bullish options premium is down 45%.
Leveraged semiconductor AUM has basically been cut in half from the June peak, and a lot of AI infrastructure names are now down 30-55% from their highs.
Pretty big change from a few months ago when everyone wanted anything with AI attached to it.
Citadel still thinks the rest of September could be messy.
Buybacks are going into blackout, systematic funds still have exposure they could sell, quarter-end rebalancing could create more selling, and roughly $7T of US equity options exposure expired Friday.
But here’s what I found interesting.
The weaker things get into the end of the month, the more constructive they’re becoming.
A lot of the leverage has already come out, positioning has been cleaned up quite a bit, and people clearly aren’t chasing semis like they were a few months ago.
Citadel basically thinks we could still get some weakness through the end of September, but they’re comfortable using more weakness into month-end to add back to core longs.
I tend to think something similar.
There are definitely AI names that got way too expensive, and obviously not every company with an AI story is going to be a winner.
But I definitely don’t think the AI trade is over.
If anything, watching how quickly these agents are improving makes me think we’re still pretty early.
A year ago, most people were just typing questions into a chatbot.
Now we have agents getting their own virtual computers and actually doing things for us.
Imagine what these things look like a few years from now.
And imagine how much compute is needed if everyone has multiple agents running 24/7.
GPUs aren’t going anywhere. CPUs might just end up being a much bigger part of this than people thought
NVDA AMD ARM INTC
QUICK DATABASE UPDATE
Also, quick update on the database.
We’re adding a raw flow tape, which should hopefully be live Tuesday.
So along with everything already in the database, you’ll be able to see the raw options flow coming through too.
Will post when it’s live.