Memory’s windfall is someone else’s bill

THE DEEP DIVE · ISSUE 16

INFRASTRUCTURE · THE MEMORY TRADE

Memory’s windfall is someone else’s bill

Micron just posted the best gross margin in tech and the Nasdaq fell anyway. That’s not a contradiction. It’s the whole map — the AI buildout has split into who charges for the shortage and who eats it.

By R. Lauritsen · 26 June 2026 · 8 min read

Here’s the thing that should bother you about this week. Micron walked into Wednesday’s print carrying the highest expectations of any chip name this year — a stock up more than 700% in twelve months, options pricing a 17% move, the kind of setup where a merely-good quarter gets you sold. And then it beat by five and a half billion dollars. Revenue of $41.5bn against a $36bn consensus. Margins guided to 86%, higher than Nvidia’s, higher than Meta’s. The stock popped 17%.

The Nasdaq fell. Fourth day in a row.

Apple dropped 6% the same session. Microsoft dropped 3%. Both because they spent the week raising prices — on MacBooks, iPads, the Xbox — and both blamed the same thing Micron was busy celebrating. Memory got expensive. The chip everyone needs is sold out into 2027, and the company that makes it is now, by margin, the most profitable name in the entire AI complex.

So who’s paying for that margin? That’s the question this piece is about. Because a shortage isn’t a thing that happens to a sector — it’s a transfer. The money Micron just booked came out of someone’s pocket, and if you own AI hardware, there’s a decent chance it came out of yours.

The shortage is a tax, and it has a direction

Follow the money from the chip down to the device, and the winners and losers sort themselves.

Memory isn’t like a logic chip. Nvidia sells you a GPU and that’s mostly the end of the story. Memory — specifically the high-bandwidth memory stacked onto every AI accelerator — is a commodity that suddenly isn’t one, because only three companies on earth make it at scale and demand has run miles ahead of what they can build. Goldman pegged the 2026 DRAM shortage as the worst in fifteen years. When something everyone needs is that scarce, the maker sets the price and everyone downstream takes it.

What makes Micron’s position so unusually clean is how it sells. On its earnings call this week it disclosed sixteen strategic supply agreements worth $22bn in committed customer cash, much of it floor-priced — customers commit and pre-pay before the capacity exists. So Micron charges for scarcity, doesn’t finance it, and gets paid in advance. Now picture the laptop maker at the other end: it buys that memory by the truckload, can’t pre-pay its way out of a shortage, and can’t make its own DRAM. The price rise lands on its cost line whole, and its only defence is to charge its own customers more — if they’ll pay, rather than buy a cheaper box.

That single question — can this company pass the cost on? — is the whole map. It splits the AI hardware chain cleaner than any sector label. So I drew it.

Read it top to bottom and the logic is brutal in its simplicity. The further you sit from the scarce thing, the less you control your own price, and the more of someone else’s windfall ends up on your cost line. Micron sets the price. ASML sells Micron the machines to make more of it — the only player that wins whether the shortage eases or not. Then comes the neutral middle, and the distinction there is the one most people get wrong. These names aren’t immune to the shortage — they pay more for memory too. They’re protected by what they sell. Tier 3 is Nvidia and the custom-silicon crowd: they pay up for memory but bundle it into a system the world can’t do without, so the cost passes straight through. Tier 4 is Apple: protected not by a bundle but by a brand, which it used this week to raise prices and bet customers would swallow it. The tell on how thin that protection really is? Apple locked long-term DRAM agreements early, has the most pricing power of any company on earth — and the stock still fell on the price rise. If the bill is heavy enough to sell Apple, the tiers below it have no defence at all.

And the bottom two tiers? That’s where the bill actually lands. A Dell, an HP, a console maker shipping hardware on single-digit margins has nowhere to hide. HP has said memory now runs about 35% of a PC’s bill of materials, up from 15–18% — the cost line roughly doubled, on a product that already competes on price. Morgan Stanley has already cut Dell to underweight on exactly this exposure. Microsoft can absorb an Xbox margin hit because Xbox is a rounding error next to Azure. A pure-play box maker can’t.

The new entrants are aiming at Tier 3, not the bottom

A footnote to the map, not a second essay: two stories this week that both pressure one tier — Tier 3’s margins.

Qualcomm picked this week to crash the data-centre party. At its investor day it unveiled a full accelerator line plus a Dragonfly CPU, and — the part that matters — named Microsoft and Meta as customers, with Meta signing a multi-generation deal. A smartphone company now has hyperscaler logos on a slide and a credible claim to challenge Nvidia. Same week, OpenAI and Broadcom unveiled “Jalapeño,” OpenAI’s first custom inference chip. The headline was OpenAI going vertical. The investable story was Broadcom, which did the silicon.

Neither of these touches the bottom of the map. They’re an assault on Tier 3 — specifically on Nvidia’s inference margins. The labs don’t want to escape memory; they can’t. What they want is to stop paying Nvidia’s markup on every inference workload, and custom silicon is how you do that. Which is why, quietly, the company you’d actually own to play that shift isn’t a lab at all. It’s Broadcom, the firm that gets paid to design every one of these chips regardless of which lab’s logo ends up on the box.

I keep coming back to that. Every “we’re building our own chip to beat Nvidia” story is, underneath, a pick-and-shovel story for the one company holding the shovels.

The five-minute test

This is the part to keep. Everything above is the argument; this is the tool.

You don’t need a model. You need to know where a holding sits on the map, and the only input is its last earnings call. Here’s the version I actually use — five questions, in order, and you can stop as soon as the answer is obvious.

Five questions, one holding, ten minutes. At the end you’ll know whether you own a company that profits from the shortage or one that quietly pays for it — a different fact from the one the share price is telling you. Run it on one name a week, and inside a month your whole AI book is sorted into chargers and eaters. That’s the habit this newsletter is really trying to build.

What to watch

The map is a snapshot. These are the dates that move things on it.

One last thing the celebration this week missed: a structure at the top of the chain is a cost shock for everything below it. That’s not a reason to be bearish on AI — it’s a reason to stop treating “AI hardware” as one position. Run the test. Classify one holding this weekend. Then do it again next Friday.

Know which end of the chain you’re standing on.

Sources: Micron Q3 FY26 earnings release and call (revenue, EPS, Q4 guidance, the 16 strategic supply agreements); Goldman Sachs research (2026 DRAM shortage); HP commentary (memory as ~35% of PC bill of materials); Morgan Stanley (Dell downgrade to underweight); Qualcomm investor day (data-centre line, named customers); OpenAI and Broadcom joint announcement (custom inference chip); SoftBank disclosures (bridge loan maturity, OpenAI-backed margin loan). Figures as reported for the week ending 25 June 2026.

Disclaimer: This article is for informational and educational purposes only and does not constitute financial advice. iPrompt Signals is not a registered investment advisor. Tickers and tiers are illustrative of the analysis, not recommendations. Always conduct your own research and consult a qualified financial professional before making investment decisions.