The AI Memory Gold Rush: The Real Money Is in the Picks and Shovels — A 5-Layer Supply Map

In a gold rush, the steadiest money goes to whoever sells the gear

Three years ago, $10,000 put into Vertiv — a company that does nothing but build cooling systems for data centers — would be worth more than $170,000 today. It never built a single chip. All it did was sell the picks and shovels that the entire AI buildout runs on.

That, to me, is how this game tends to work. In a gold rush, the steadiest money has always gone to the ones quietly selling the gear, not the ones swinging for the gold. And right now that same thing is happening in memory, where AI is eating it up faster than the world can make it.

The memory AI craves is HBM — high bandwidth memory. It's the chips stacked right next to every AI processor, where prices have roughly doubled in a single quarter. The makers are completely sold out into next year, which is exactly why Micron just crossed $1 trillion.

But Micron is the obvious name. The real money, in my view, sits in the picks and shovels underneath it. The memory suppliers get paid no matter which maker wins. Those are the names I'm hunting down here.

The memory stack framework, in five layers

I look at this as five layers, and every single one is a choke point — a spot where the whole thing grinds to a halt without it.

Layer	Job	Key names
1. Interface	Chips that let memory run at AI speeds	Rambus
2. Tools	Machines that physically build the memory	Applied Materials, Lam Research
3. Inspectors	Catch a single defect before it ruins a chip	KLA, Onto Innovation
4. Testers	Prove every layer actually works	Advantest, Teradyne
5. Bonders & materials	Fuse the chips, feed the fab its raw inputs	Besi, Kulicke & Soffa, Entegris

The point is that these 10 names win whether SK Hynix, Samsung, or Micron comes out ahead. Most of them you've never heard of — which is exactly why they belong on a watch list. This is a 5-year build, and we're still in the first few innings.

Layer 1 — Interface: Rambus and the model that gets paid twice

The interface layer belongs almost entirely to Rambus. Three companies control 97% of a market most investors don't know exists.

Rambus doesn't make memory. It makes the registering clock driver, the chip that sits on a DDR5 module and keeps data clean at the speeds AI demands. The module simply doesn't run without it. On top of that, Rambus owns the underlying memory patents, so the big manufacturers pay it a royalty on the memory they sell. Put those together and Rambus gets paid twice on the same chip — once to build it, once to license it.

The numbers prove how clean the model is. Over four years revenue roughly doubled, but operating income exploded almost 800%, and the operating margin went from 9% to 37% because those royalties are nearly pure profit. The one caveat: there's an open antitrust inquiry worth keeping on your radar.

Layer 2 — Tools: Applied Materials and Lam Research

HBM burns through roughly three times the equipment per gigabyte of standard memory. Applied Materials supplies the widest slice of it — deposition, etching, layering, the steps that turn sand into a working chip. It even runs a joint research center with SK Hynix, so its machines get designed into the next generation first. Its HBM revenue went from almost nothing to roughly $1.5 billion in a single year, with management targeting $3 billion next.

Its opposite number is Lam Research, and I see Lam as the purer memory bet because the physics are simple: every time a memory stack gets taller, you can't build it without more of Lam's tools. Lam owns the etch and deposition steps that carve deep vertical channels and fill them with copper — literally how you stack memory into a tall HBM chip. Its revenue fell in the last memory downturn, then snapped to an all-time record the moment spending returned, climbing 24% in a year with net income up 40%.

Layer 3 — Inspectors: KLA's near-monopoly and Onto's hidden-defect hunt

Bonding memory into a hybrid stack creates up to 1,000 times more connections than the old method, and a single buried crack scraps the whole part. Onto Innovation built the high-speed imaging that sees defects underneath the surface, and its new Dragonfly G5 was just picked by a leading maker for inspecting next-gen HBM4. The whole company does only about $1 billion a year, yet one customer locked it into a purchase agreement worth more than $240 million running through 2027 — nearly a quarter of a full year's revenue committed years in advance.

KLA's edge is sheer scale. It's the closest thing this industry has to a monopoly in finding nanoscale defects, running many times the size of its nearest rival. The more advanced chips get, the more places a defect can hide, and the whole industry leans on KLA to catch them. Over four years revenue grew 76%, operating income climbed 110%, and the margin pushed from 36% to 43%. With no real number two, almost every inspection dollar funnels into a single company.

Layer 4 — Testers: Advantest and Teradyne

A stacked memory chip takes up to 10 times the testing of an ordinary one, which is why Teradyne just launched the Magnum 7H for HBM. But I want to be precise: Teradyne's total revenue has actually slipped about 14% over four years as its oldest test lines cool. Underneath that, the memory engine is taking over — DRAM test revenue went from $80 million to $350 million, more than 330%, and AI work has already passed 60% of the business heading toward 70%.

The global giant on this rung is Advantest, which by most estimates holds around two-thirds of the entire chip-testing market. The clever part: the same family of testers checks both the AI processor and the HBM stacked around it. It's effectively the toll booth for the whole testing layer. In its most recent fiscal year, sales grew about 45% but operating profit jumped almost 120%. The practical note: in the States it trades over the counter, so it's a bit thinner to buy.

Layer 5 — The foundation: bonders (Besi, Kulicke) and materials (Entegris)

The bottom layer does two jobs: fusing chips together, and feeding the fab raw materials by the ton.

The leader in fusing is Besi. HBM4 increasingly can't be built without hybrid bonding, which welds memory dies directly copper-to-copper instead of using tiny solder bumps. Besi is at the front of it, building bonding lines jointly with Applied Materials. Its revenue grew 28% last year, but new orders did something far more dramatic — up 104% year over year. In this business orders come a year or two before revenue, so that gap is the real signal.

Kulicke & Soffa is, honestly, the speculative pick. Revenue has been cut more than half over four years, down 57%, and the company sits near break-even today. The entire case is forward-looking — its advanced packaging line is just starting to ramp, and real HBM volume is a 2027 story. You hold it for where memory packaging is heading, not for what it earns today.

The materials side is Entegris, which supplies ultra-pure chemicals, filters, and polishing materials. Here's the number that reframes the company: the most advanced AI and HBM chips are only about 5% of all wafers made in the world, yet they already drive close to 30% of Entegris' revenue, because each one soaks up far more material. Entegris wins on complexity alone — and complexity is the one thing memory is guaranteed to keep adding.

How I'd put $100 to work

I wouldn't split it evenly. I'd weight toward the layers that keep winning even if data-center spending cools — the ones every advanced chip has to pass through.

• Tools (Applied Materials, Lam Research): $25
• Inspectors (KLA, Onto): $25
• Testers (Advantest, Teradyne): $22
• Foundation / bonders & materials (Entegris, Besi, Kulicke): $20
• Interface (Rambus): $8

Rambus is the name I opened with and the royalty model is beautiful, but it's the narrowest name on the board with an open legal question, so I size it down. Best story, smallest bet — that's just discipline.

The core of this framework is simple: whoever wins the memory war, these 10 layers collect the toll. That's reason enough to build the watch list now, in the early innings of a 5-year build.