🤖 Agents Eat Memory

Good morning to everyone who thinks the agent boom is a GPU boom.

A GPU is the chip that does the math. An agent is software that does a job, not just answers a question.

Goldman just counted how much work those jobs will generate.

[]

Today, why that work lands on memory, not on the GPU.

AI is moving from experiment… to essential.

Every major industry is integrating it.
Every major company is investing in it.

By late 2025, AI was already an $800B market — growing at a pace that could push it well beyond $1 trillion in the years ahead.

Cloud infrastructure is scaling fast.
AI-enabled devices are multiplying.
Automation is becoming standard.

But here’s the real question…

When trillions flow into this transformation — which stocks stand to benefit most?

Our new report reveals 10 AI stocks positioned across the backbone of this shift — from the companies powering the infrastructure… to those embedding intelligence into everyday systems.

If you want exposure to one of the defining growth trends of this decade, start here.

Download the Report Now

Goldman Sachs Research published a chart in May of monthly token use.

A token is the unit an AI model bills by, roughly three-quarters of a word.

The chart says that pile multiplies 24x by 2030, to 120 quadrillion tokens a month.

Almost all of the new pile is agents.

Chatbots, the old "ask a question, get an answer" workload, barely grow.

That gap is the news.

Chat spends a few tokens and stops.

An agent keeps the file open.

It books the travel, checks the ledger, calls a tool, checks again.

Gartner models 5 to 30x more tokens for an equivalent task.

Use a reasoning model and they put the cost at about 150x a simple chatbot.

Micron's Sumit Sadana has used that same 5-to-30 range when he talks about what this does to memory.

Something big is coming to Second Order! 🤖

We’re building a new way to help you go beyond understanding the AI revolution and start acting on it.

Your vote will help decide what we build first.

Thanks for helping us build it right. Now, back to it.

A job that stays open has to remember what it already did.

The GPU does the next calculation + memory holds the file.

That file has a name: the KV cache.

KV is key-value. Think of it as the model's notepad for this shift, so it does not start from zero on every step.

A longer job means a bigger notepad.

The notepad sits in HBM.

HBM is high-bandwidth memory, stacked DRAM glued onto the accelerator so data can move fast enough to keep up.

You can design a faster GPU. You cannot magic more HBM. Three companies make almost all of it: Samsung, $SKYHY ( ▲ 0.56% ) , and $MU ( ▲ 2.56% )

Micron's CEO told analysts in June the company can fill only about half to two-thirds of the HBM requests it gets, and he does not see that catching up through the end of 2026.

SK Hynix CEO Kwak Noh-jung told Reuters in July that 2027 will be the worst year in the industry's history from the supply side, with customer demand still above capacity past 2030.

The price tape already agrees.

TrendForce has conventional DRAM contract prices up about 93% to 98% in the first quarter of 2026. Conventional DRAM is the regular memory in servers and PCs.

So what does this mean for your portfolio?

$NVDA ( ▲ 1.5% ) sells the engine. $MU and $SKHY sell the notepad the engine cannot run without. Samsung is the third name in a three-firm market.

Last week cheaper tokens looked like a warning. They were an invitation. Agents accepted it. Agents pay memory.

The stocks still trade this as a cycle. The Goldman chart trades it as a bottleneck.

Chat answers → the agent works = memory holds the work.

Second Order is for information and entertainment only and is not financial advice.

We are not licensed financial advisors.

Any tickers or companies mentioned are our opinions, not recommendations to buy or sell.

Do your own research and consult a professional before making investment decisions.