Weekly Update: The $3 Trillion Agentic AI Trade
Good evening, and welcome to this week’s edition of Stealth Trades!
I don’t need to tell you that artificial intelligence is the big theme of the stock market. It has been since 2022. But that theme is rapidly evolving.
On May 7th, Anthropic’s Claude Cowork was introduced to Microsoft Office. It is now living inside Excel, PowerPoint, and Word. Generally available. Not a demo. Not a waitlist.
It is embedded in the three apps that 1.4 billion office workers open every single morning.
Microsoft’s already there too. Copilot has around 20 million paying users, and they rolled out a new Office bundle on May 1st.
And here’s why that’s a bigger deal than it sounds…
Up until a few months ago, “using AI at work” meant you were a developer writing code in something like Cursor. That’s maybe 25 million people on the whole planet.
Now?
AI is about to operate the actual software that over a billion knowledge workers use to do their jobs.
It is a major inflection point.
And few people understand what agentic AI is about to do to global computing demand.
This is a completely different animal than the chatbots you’ve been using.
And there is a basket of stocks that quietly power this entire build-out. Some of these you have never heard of.
WHY AGENTS BREAK THE OLD MATH
If you haven’t started playing with all the fancy AI tools yet, there’s something you need to understand – an AI agent is not a chatbot.
When you ask ChatGPT a question, that’s quick. It doesn’t use a lot of compute power. But AI agents are different beasts.
Compute is measured in tokens which are just little units of compute.
A developer running a focused 30-minute coding session burns somewhere around 50,000 to 200,000 tokens.
Now compare that to a user like me doing financial research.
I might ask an agent to analyze the nuclear power ecosystem, analyze every publicly traded company in the space, reconcile a 200-row file and compare it to industry comps, write a full email summary, and build a 40-slide deck from the information. All from one command.
That single task burns 800,000 to 2 million tokens.
Now… multiply that by 1.4 billion office workers.
Microsoft has already said over 80% of the Fortune 500 has rolled Copilot out to be part of their workforce. The load this puts on the system makes everything that came before it look like a rounding error.
And the infrastructure to handle it does not exist yet.
A chatbot session is short, GPU-heavy, and low-context. An agentic session is the opposite — it runs for hours, it constantly does things, and it leans on CPUs, memory, and GPUs all at once.
The server CPU-to-GPU ratio has already shifted from 1-to-12 in the old training builds to 1-to-2 in these new agentic builds.
That means a whole pile of components nobody was talking about — CPUs, memory, storage, power chips — just became mission critical. The picks-and-shovels layer just got a massive promotion.
That’s why stocks like Sandisk and Micron are going vertical right now.
SURGING TOKEN DEMAND
According to Goldman Sachs’s models, consumer and enterprise agents will push monthly token consumption to roughly 24 times the entire world’s current capacity by 2030.
And in that same report, Goldman noted that quote “token economics turn positive in the first half of 2026.” That’s the moment an agent’s output is worth more than what it costs to run it.
And once that flips, every company on earth has to deploy agents just to stay competitive.
Right now, agentic AI is still more hobbyist tinkering than full commercial deployment. The second it pays for itself, adoption goes vertical — months, not years.
And get this — we are still ridiculously early. The latest estimates say about 78% of humanity has never used a single AI tool. Not once.
THREE U.S. STOCKS YOU CAN BUY EASILY
Unfortunately, a lot of the best names in this stack trade overseas. But I know most of you want something you can buy in the account you already have, without jumping through hoops. So, here are three US-listed stocks that each own a different layer of the agentic build — CPUs, memory, and storage.
STOCK #1 — ARM HOLDINGS (ARM)
This is the toll booth. When a hyperscaler builds its own custom AI chip — Amazon’s Graviton, Google’s Axion, Microsoft’s Cobalt, Nvidia’s Grace and Vera — almost every one of them is built on Arm’s architecture. And Arm collects a royalty on each and every chip.
They don’t care who wins the design war between Amazon, Google, and Microsoft. They get paid either way.
STOCK #2 — MICRON (MU)
Remember what I told you about agentic AI? Agents run for hours and they’re hungry for memory — that’s where all the context lives. Micron is the only US-listed pure-play in high-bandwidth memory, the stuff that sits right next to the GPU.
And guess what? Their entire 2026 HBM production is already sold out. The whole year. The stock has gone parabolic. It is unquestionably extended. But it’s hard to price a company selling out a year of capacity in advance.
STOCK #3 — SEAGATE (STX)
Agents generate mountains more data than chatbots ever did, and all that data has to live somewhere. Seagate is one half of the hard-drive duopoly, and for the first time since 2017 they’ve got real pricing power — capacity reportedly sold out through year-end.
If you want the storage layer of this trade, this is the easy US way to play it.
And if you want to go deeper into storage, its duopoly partner Western Digital, ticker WDC, and the NAND spin-off SanDisk, ticker SNDK, are other options.
INTERNATIONAL STOCKS
If you have an account somewhere like Interactive Brokers that lets you trade global markets, you might also consider some of these:
- Tokyo Stock Exchange
- Renesas (ticker 6723) – power-management chips
- Nitto Boseki (ticker 3110) – 70% glass-cloth monopoly
- Murata (ticker 6981) – leading capacitor maker
- Kioxia (ticker 285A) – freshly-IPO’d NAND play
- MEC (ticker 4971) – microetching chemicals
- Taiwan Stock Exchange:
- Global Unichip (ticker 3443) — TSMC-backed design house
- Hon Hai, better known as Foxconn (ticker 2317) – server assembly and humanoid robots
- Lotes (ticker 3533) – sole-source CPU socket maker
- Unimicron (ticker 3037) – substrates
MY CONCLUSION
The Nvidia trade is the one everyone already made. The company is worth $5 trillion now. It’s up 1,600%. The move already happened.
But the agentic AI trade — the memory, the storage, the substrates, the sockets, the power chips — is the one investors are finally waking up to.
Some of these stocks are already up big. But so was Nvidia a couple years ago, and it doubled three more times.
Best wishes for your trading,
