Micron’s Run Looks Unstoppable. Here’s the One Number We’re Watching.
Why one of the market’s most volatile trades might still have room to run… and how to know when it doesn’t
Listen to the audio version of this article (generated by AI).
If you’ve ever read George Washington’s journal entries, they’re about as dry as the wheat Washington farmed at Mount Vernon. But they provide an interesting look into Washington as an early adopter (pivoting from tobacco to wheat) and a patient farmer who obsessed over timing.
In the Washington Papers, we’re transported back to July 1770, where a month of harsh June rain had beaten the straw flat, yielding but a few grains per head. Most of Washington’s crop had either perished or was too mildewed to harvest. It was, in his own words, “exeeding[ly] bad.”
But George knew where it had gone wrong, and he waited… tallying the month’s losses in his journal and reaffirming his thesis. That is, a three-week harvest should begin before the wheat is ripe, or one might risk the entire loss of one’s crop.
More than two centuries later, that is the question we’re asking in our collaboration with Stansberry’s Director of Research, Matt Weinschenk… are you too late to the harvest?
Micron (MU) has been one of the great winners of the AI Boom. Over the past year, the MU stock chart has run up and to the right with almost no resistance.
Micron stock’s run is exactly what makes serious people nervous, because they know the memory cycle. They have watched it come and go. It has burned investors and minted fortunes, and the oldest saying on the desk is that you own memory only as long as the music is playing (and whoever is still standing when it stops loses their shirt).
We’ve heard that warning for two years running, and we’re still in the bull camp. Our case is that this cycle is built differently. Every memory boom that came before it was capped by the number of folks buying phones and laptops. An agentic AI world has no such ceiling … every new agent wants its own memory, and the only real gauge of the cycle becomes how much the hyperscalers are willing to spend.
That number is still climbing toward a trillion dollars a year, and we do not see the peak inside the next one to two years.
So the field is still filling. The music is still playing loud. And we are watching the one gauge that will tell us when it’s time to cut. Check out the episode below:
Why AI Needs More Memory Than Ever
Here’s the case in its simplest form: compute needs context.
The GPUs that Nvidia (NVDA), Advanced Micro Devices (AMD), and increasingly Intel (INTC) and Qualcomm (QCOM) are selling are the brains of this buildout. But a brain without context can’t do much. Memory is what gives these systems the context to actually answer a question or finish a task… and as AI has shifted from chatbots that simply respond to agents that remember, plan, and execute multi-step work, that context window has become the bottleneck.
I look at the chips themselves as the clearest evidence.
Nvidia’s A100 carried 80 gigabytes of memory. The H200 jumped to 151. The Blackwell chips now ship with nearly 300. Each generation of GPU needs exponentially more memory than the one before it… and every new data center needs exponentially more GPUs than the one before that.
The demand curve is compounding.
The Memory Cycle, Explained (And Why It’s Dangerous)
Memory is also the most commoditized link in the entire semiconductor chain. DRAM is DRAM, NAND is NAND, whether Micron makes it or SanDisk (SNDK) or Seagate (STX). That commoditization is exactly what makes the cycle so violent.
The pattern repeats: a demand boom arrives (PCs, mobile, cloud, work-from-home) and suppliers race to build capacity. By the time that new supply comes online, 12 to 24 months later, the demand that justified it has already cooled. Supply glut meets falling demand, margins collapse, and the stocks that looked unstoppable get cut in half. I count seven or eight separate 50%-plus drawdowns in Micron since 2008.
That history is where the old trading rule comes from: own memory stocks only as long as the music is playing, because when it stops, whoever hasn’t found a seat also loses their shirt.
Why I Believe This AI Memory Cycle Is Different
Every memory boom before this one was capped by a number Wall Street could count: how many phones, how many laptops, how many people were buying them. Human demand has a ceiling.
Agentic AI doesn’t.
There’s no point at which all the agents have enough memory and the industry stops needing more… the answer is simply to build another agent. The cap on past cycles was the size of the human population.
The cap on this one is hyperscaler spending, and that number keeps moving in one direction: this year’s roughly $700 billion to $800 billion in AI infrastructure spend is on pace to approach $900 billion next year, with outside estimates from Goldman Sachs and Bloomberg projecting close to $1 trillion annually by 2029 or 2030. Those forecasts keep getting revised up, not down… and that’s the trend I watch most closely.
Google’s (GOOGL) recent decision to raise $80 billion (upsized to $85 billion, with $10 billion coming from Berkshire Hathaway) is the tell.
Every hyperscaler is racing the others, and none of them believe they can afford to fall behind. When one raises its spending, the rest match it… and the shockwave multiplies across Amazon (AMZN), Microsoft (MSFT), Meta (META), and the Chinese cloud giants doing the same thing in parallel.
The Rolling Bottleneck
This buildout hasn’t moved in a straight line. Rather, it has moved from constraint to constraint. GPUs came first, which is why Nvidia led. Then the industry needed to assemble those chips into servers and racks, which is why Super Micro (SMCI) ran hard before legal trouble handed that business to Dell (DELL).
Then it needed data centers to house them, then networking and optics to connect them.
Memory is simply the latest link in that chain to come into focus. And from where I sit, Wall Street is only now catching up to what should have been obvious from the start: a data center needs all of it.
How High Can Micron Stock Go
Micron has run from roughly $800 toward $1,000 in recent weeks. My framework: the stock tends to advance in sharp bursts of 70% to 80%, then give back roughly 20% before resuming the climb. I expect that pattern to hold… a pullback toward the $800 level on any catalyst, geopolitical or otherwise, followed by a resumption of the climb toward $1,200 and eventually $1,400.
My case for staying long isn’t built on multiple expansion. Despite the stock’s run, Micron trades near nine times forward earnings, against a five-year average closer to 6.7 times — only modestly above its historical range, even after a tenfold move since 2024.
What’s actually driven Micron stock is the earnings power underneath it: EBIT near $9 billion in 2024 is on pace to approach $40 billion over the trailing twelve months, with estimates near $150 billion by 2027. The dollars have moved first. The multiple has barely followed.
What Could End the AI Memory Boom
I want to be direct about the risk here: this can’t go on forever, and no buildout in history has gone in a straight line.
The vulnerability I watch is the consumer. The hyperscalers’ AI budgets are ultimately funded by ad sales and product purchases — Amazon, Meta, and Google’s spending all trace back to discretionary spending from ordinary households.
The personal savings rate has fallen to 2.6%, a level I’ve only seen matched twice before: briefly in 2022, and in the two years leading into the 2008 financial crisis.
Real wages are negative…
Consumer sentiment sits near record lows…
If oil holds above $100 or the 10-year Treasury yield pushes past 5% for a sustained stretch, discretionary spending slows, ad revenue softens, and the hyperscalers’ capacity to keep funding this race shrinks with it.
There’s a political risk layer too: proposed legislation to redirect AI profits to households, and a handful of states moving to restrict new data center construction. None of it has teeth yet.
All of it is on my watchlist.
The Bottom Line
My read: we’re in the third mile of a marathon on AI model development, but the spending race itself could end far sooner if a shock — economic or geopolitical — forces the hyperscalers to pull back. Until that signal shows up, the music is still playing, and I remain firmly in the bull camp on memory stocks.
P.S. For the full conversation, including more on the memory cycle’s history and Luke’s case for why this one breaks the pattern, watch this week’s full episode of Top Stocks with Matt Weinschenk, featuring Luke Lango. And be sure to subscribe to Top Stocks on YouTube for more exclusive content.
Also, Join Luke at this year’s Stansberry Conference & Alliance Meeting – where ideas move fast, conviction gets sharper, and the next big opportunities come into focus.
You’ll get live market updates, learn about top ideas and stock picks from Jonathan Rose and Luke Lango, and have the chance to meet some of your favorite editors – like Marc Chaikin, Whitney Tilson, Dr. David Eifrig, and Keith Kaplan.
Attendees will hear from bestselling authors and experts in economics, technology (including AI), and more. This year’s featured speaker lineup also includes famed actor Henry Winkler (aka “The Fonz” from Happy Days).
Expect two days packed with intriguing presentations and fun social events – all in luxurious Las Vegas. It pays to be in the room where it all happens.
