AI is no longer a brain in a jar
I made a costly mistake.
When I booked our family's summer flights, I forgot to add our frequent-flyer numbers. All those miles were about to vanish.
Fixing that normally means hours on airline websites, filling out forms and waiting on hold.
So I tried something different...
I sent a WhatsApp voice note to Instinct, a personal artificial intelligence (AI) agent you message just like a friend.
Instinct went to work. It created frequent-flyer accounts for my wife and kids. It set up a family account and claimed our missing miles. All I did was answer a few quick questions along the way.
ChatGPT could have explained how. Instinct did it for me.
I've since given Instinct access to my Gmail and computer files. It books hotels and restaurants and handles everyday admin. A friend of mine also used one of my Instinct invites to sort out a complicated medical problem, saving him countless hours.
AI is shifting from telling us what to do to doing it for us.
That shift is very bullish for a small group of stocks.
Demand for computing power is about to explode…
Imagine what happens once people discover how much time agents save.
One errand becomes five. Five become 20. Soon, we'll all have agents working for us all day long, even while we sleep.
Every errand requires computing power.
Instinct's founder, Noah Shinn, says securing computing power now takes up about 40% of his time. Why? Demand is doubling roughly every week. And Instinct is still invite-only.
Big tech was already spending like never before. AI agents pour fuel on that fire.
The big winners of the AI agent boom will be…
CPU makers.
For years, Nvidia (NVDA) has led the AI boom. Its graphics processing units (GPUs) handle the math that helps AI think.
But agents don't just think. They act. That work runs on central processing units (CPUs), the workhorse chips inside computers and servers.
GPUs do the thinking. CPUs do the doing.
A chatbot is a brain in a jar: You ask, it thinks, it answers. Mostly GPU work.
An agent is a brain with hands. It thinks, then opens a browser, logs in, fills out forms, checks results, and goes again.
Every agent needs its own computer…
One GPU can think for many people at once. But your agent's open browser, accounts, and half-finished bookings can't be shared with mine. Every agent needs a small computer working in the background.
When one reviewer asked Meta Platform's (META) new agent Muse to inspect its own computer, it reported two processor cores, about 8 GB of memory, and 100 GB of storage on an Advanced Micro Devices (AMD) server chip. No GPU.
Now picture millions of users with those computers.
The old chatbot servers ran one CPU for every four to eight GPUs. AMD says agents are pushing that toward one for one. Arm Holding PLC's (ARM) CEO says agent-era data centers will need 4X as many CPU cores as they did last year.
Supply can't keep up. Last month, Intel Corp.'s (INTC) CEO admitted it can only meet about half of customer demand. Agents also need memory and storage. Intel says memory prices are already up five- to seven-fold.
My recommendation…
Build a three-stock AI agent infrastructure basket. Each company supplies a different piece of the agent’s computer.
Advanced Micro Devices (AMD): The landlord.
AMD makes EPYC, the chip powering Muse’s computer. Its newest EPYC, Venice, packs up to 256 cores into a single chip. That's room for roughly 250 Muse-style agents. This week, HPE launched Venice servers, pitched at AI agents.
AMD now expects the server CPU market to grow more than 35% annually to over $120 billion by 2030—roughly double its previous forecast.
Arm Holdings PLC (ARM): The tollbooth.
Arm designs chips and licenses those designs to others.
Amazon (AMZN), Google (GOOGL), Microsoft (MSFT), and Nvidia all build server CPUs on Arm designs, paying royalties on each.
Its data center royalties more than doubled over the past year. In March, Arm launched its first chip, with Meta as the first customer. Demand already exceeds $2 billion.
Micron Technology (MU): The memory supplier.
Micron makes the memory and storage every agent needs.
At 8 GB of memory apiece, one Venice chip needs about 2 terabytes of memory.
Business is booming. Micron's sales last quarter came to $54.2 billion, nearly 5X what they were a year earlier. And it expects supply to stay tight through 2028.
We own Micron in our Disruption Investor portfolio (upgrade here), and we're up more than 300% on the position.
But memory is cyclical, and stocks often peak before the shortages end. So size Micron accordingly.
A word on price…
None of these stocks is a secret. So far this year, AMD is up 190%... Arm is up 152%... and Micron is up 263%.
Buy gradually on pullbacks. Arm, for one, already sits 37% below its June high.
Instinct got my miles back with a single voice note. Now multiply that by millions of people running billions of errands.
Each errand runs on chips that do the doing.
Stephen McBride
Chief Analyst, RiskHedge
Share this article
Where Innovation Meets Investing