Who Gets Paid When AI Needs More Memory? A Field Guide to the Toolmakers

Article Brief

Key Takeaways

5 Points30s Read

1. 01The setup–Four companies — ASML, Applied Materials, Lam Research and KLA — sell nearly all the machines that turn memory capex into memory capacity.
2. 02The logic–Toolmakers get paid when fabs are built, not when chip prices hold. Order books follow capital-spending commitments, which run on next year’s forecast.
3. 03The AI kicker–HBM is the most equipment-intensive memory ever made — thousands of etched through-silicon vias per stack plus a full advanced-packaging flow.
4. 04The buyers–SK Hynix, Samsung and Micron fund the arms race; whoever wins HBM share, all three pay the same four vendors on the way in.
5. 05The catch–Equipment is the first line memory makers cut in a downturn — the group’s history includes 50% drawdowns in cycles that only bruised its customers.

Every gigabyte of AI memory starts life as a purchase order for a machine. Before SK Hynix can stack a single layer of HBM4, before Micron can ship a wafer of high-bandwidth DRAM, somebody has to buy the etchers, the deposition chambers, the lithography scanners and the inspection rigs that do the physical work. Four companies sell almost all of those machines — and in an AI cycle where memory prices whipsaw week to week, their order books have quietly become the steadiest way to read what the industry actually believes.

That gap between what chips cost and what chipmakers spend is the whole story. Memory equities fell into a bear market at the start of July while record-scale memory capex was being financed the same week — SK Hynix raised about $26.5 billion in a Nasdaq listing earmarked, per Reuters, for new factories and equipment. Prices said fear. Spending said build. The toolmakers get paid on the second signal.

The four companies that build the machines

Wafer-fab equipment is an oligopoly with unusually clean lanes. Each of the four leaders owns a different step of the process, and they mostly do not cross into each other’s territory. That is what makes the group readable: when memory capex rises, each one gets paid in a specific, predictable way.

ASML (ASML)

Everything begins with lithography — projecting the circuit pattern onto the wafer — and ASML is the only company on Earth that ships the extreme-ultraviolet machines advanced chips require. Memory has historically leaned on EUV less than logic, but that changes with every DRAM generation: the tighter the pitch, the fewer tricks remain besides better optics. ASML’s backlog is the industry’s longest-dated commitment to the future, which is why its earnings reports get read like macro data for the whole chip complex.

Applied Materials (AMAT)

Applied Materials has the broadest catalog in the business — deposition, ion implant, polishing, and an expanding advanced-packaging line. Breadth makes it the diversified bet: no single process step can sink it, and its management commentary tends to describe the whole spending environment rather than one niche. When Applied’s chief executive talks up a multi-year boom, as he did during Lam’s July repricing week, he is describing the tide, not one boat.

Lam Research (LRCX)

Lam Research is the memory pure-play of the group. Etch and deposition — carving structures and layering films — are the two steps that dominate a memory fab’s tool budget, and Lam leads both. High-bandwidth memory adds a third: the thousands of through-silicon vias drilled through each HBM stack are an etch problem, squarely in Lam’s toolbox. That concentration is why Lam gained 154% in the first half of 2026 without selling a single chip — TECHi’s breakdown of the repricing argument now running in the stock covers the $348 consensus versus the $475 fresh marks — and why it falls hardest when a memory maker so much as hints at deferring a fab.

KLA (KLAC)

KLA gets paid to find mistakes. Its metrology and inspection systems measure whether the other three companies’ machines did their jobs, and its economics improve as processes get harder — more layers and tighter tolerances mean more measurement per wafer, not less. Yield is the difference between a profitable memory fab and an expensive science project, which makes KLA the closest thing the group has to a recurring-revenue story.

Why memory is the tool-hungry end of the business

Memory scaling stopped being about shrinking a flat design years ago. NAND now goes up instead of out — flagship parts stack well over 200 layers of storage cells, and every added layer is another deposition pass and another high-aspect-ratio etch, like drilling an oil well the width of a virus. DRAM keeps shrinking, but each node needs more patterning steps to get there. And HBM, the stacked DRAM feeding every AI accelerator, bolts an entire packaging flow on top: drill the vias, fill them with copper, thin the wafer, stack the dies, test the tower.

The consequence is simple and structural: every generation of memory costs more machine-hours per bit than the one before it. The bit gets cheaper; the factory gets dearer. That treadmill is the toolmakers’ business model, and AI just turned its speed up — HBM4 pricing arrived roughly 60% above HBM3E, and the capacity race it kicked off is measured in new fabs, not incremental upgrades.

The buyers: three companies, one arms race

The customer list is as concentrated as the supplier list. SK Hynix has spent 2026 racing to feed AI memory demand and just financed the next leg of that race at record scale. Samsung is fighting to reclaim HBM share it ceded. And Micron passed its $41 billion test — a blowout quarter — even as its stock round-tripped on glut fears. On the NAND side, SanDisk was the S&P 500’s best stock of the first half before the same fears bit it too.

Notice what all three DRAM players have in common: whoever wins the HBM race, they buy from the same four vendors on the way in. That is the toll-booth logic — equipment is the one lane every competitor has to drive through. A share fight between Seoul, Suwon and Boise is, from the toolmakers’ side of the counter, just more traffic.

How the cycle pays — and how it ends

The honest version of this story includes the ending. Equipment orders lead memory prices in both directions: fabs are financed on next year’s demand forecast, so the order book turns before the spot market does — up and down. When a memory maker facing softer prices defers a fab, the toolmaker loses that revenue instantly and completely, which is why equipment names have a long history of 50% drawdowns in downturns that only bruised their customers.

The glut question is therefore the equipment question. TECHi’s breakdown of the July glut scare walked through the math of whether AI demand can absorb the capacity being financed; if the answer ever turns no, the toolmakers will price it first. The mid-July tape showed the mechanism live: when SK Hynix’s cautious outlook rattled the memory complex, the equipment names fell alongside — and in some sessions harder than — the memory makers themselves. The market was repricing future purchase orders, not current chips. That asymmetry — first paid, first cut — is the price of the toll booth.

What to watch

• WFE guidance, not EPS. At equipment earnings, the number that moves the group is the industry-wide wafer-fab-equipment spending forecast, and each management’s memory-versus-logic mix commentary.
• HBM supply contracts. Multi-year supply deals between memory makers and AI chip designers are the demand signal that de-risks fab construction — and triggers tool orders.
• Capex announcements from the big three. SK Hynix, Samsung and Micron publish capital-spending plans; revisions in either direction hit the toolmakers before they hit memory prices.
• China and export rules. A meaningful slice of every toolmaker’s revenue depends on what Washington allows to ship east; rule changes reprice the group overnight.
• The packaging line item. Advanced packaging is the fastest-growing tool category of the AI era — it is where HBM demand shows up first in equipment revenue.