ASML Is the Bottleneck Behind HBM. The Stock Already Knows It.

Strip a module of high-bandwidth memory open and you will not find a new kind of chip. You will find a stack of DRAM dies, grown on the same leading-edge memory lines as ordinary DRAM, then wired together through the bottom. And those dies are the part where the monopoly lives. HBM dies run on the most advanced nodes the memory makers operate, nodes where the critical layers are printed with extreme ultraviolet lithography, and where the EUV machine has exactly one vendor in the world: ASML.

So the claim in the headline is true in substance — with one precision worth keeping. ASML does not build the part of HBM that made it famous: the stacking, the through-silicon vias, the bonding. That is SK Hynix, Samsung, and Micron's own engineering, done with packaging tools the lithography companies do not make. ASML's content is the DRAM die underneath, patterned with EUV at a wavelength of just 13.5 nanometers, a technology the company describes as unique to it. You cannot make HBM without those dies, and you cannot make those dies at their required density without ASML's machines. Every HBM ramp is a demand amplifier pointed straight at high-end DRAM lithography.

That is the quiet change in the business model. For years, memory swung between a third and a half of ASML's sales — the cyclical part that expanded when DRAM prices were high and disappeared when they fell. In the first quarter of 2026, the split flipped: memory was 51% of system sales. In the fourth quarter of 2025, memory was 56% of the record €13.2 billion in net bookings. The DRAM makers have moved to the front of the queue, and management now expects memory revenue to climb about 75% this year against roughly 25% for advanced logic. Christophe Fouquet, the CEO, described the demand for HBM and DDR memory as a "perfect storm", with tight supply expected to last through at least 2026.

Why this holds together as an investment map and not just an earnings narrative is the part worth slowing down for: the bottleneck has migrated. In 2024 the constraint on AI was graphics processors. In 2025 it was HBM itself. By 2026, industry watchers had moved the label one more layer up the chain, to EUV capacity — because an EUV shortage is not solved by reallocating wafer starts or packaging lines; it limits the industry's ability to create that capacity in the first place. And EUV capacity is a number ASML controls deliberately: roughly 65 low-NA machines this year, a 30% increase planned for 2027, another 30% under evaluation for 2028.

The customer behavior is the confirming evidence, and it is unusually clean. In March, SK Hynix — supplier to Nvidia and the holder of more than 60% of the HBM market — disclosed a roughly $7.9 billion purchase of ASML EUV tools, which analysts read as about 30 machines scheduled by the end of 2027, spread between its HBM-focused plant in Cheongju and its advanced DRAM cluster at Yongin. It was the largest single EUV order ever publicly disclosed by an ASML customer, and analysts flagged a telling detail: the deal included a pull-in element, SK Hynix accelerating its own factory timeline specifically to secure ASML equipment ahead of competitors. That is what a chokepoint looks like. Customers do not choose the machine. The machine's schedule chooses them.

The moat also has a next step in the same direction. SK Hynix was the first memory maker to install ASML's next-generation high-NA EUV system, the EXE:5200B, at its M16 fab in Icheon — the next node of DRAM scaling where the same sole-source logic applies. Intel is already using high-NA in production. On the other side of the ledger, the alternatives are thin: Canon leads only in the mature i-line segment, Nikon has retreated from advanced lithography, and ASML's own development history — two decades plus a global supply chain of optics and lasers — is the answer to anyone expecting fast substitution.

Now for the hard part of the map, and it is the part the past year has made difficult. The dependency is confirmed. The customer is confirmed. The price is the soft spot. ASML trades near $1,700 per share as of late August, up roughly 137% over the trailing year and about 58% year to date, at a market capitalization near $650 billion — roughly 53 times trailing earnings, about 41 times EBITDA, 16 times sales, with a dividend that yields about half a percent. Even against this year's raised plan — ASML lifted 2026 guidance to €43–45 billion of sales from the €36–40 billion range it gave in April, on a gross margin of 54–56% — the shares sit in the high 40s on expected 2026 earnings. The company reports it is close to holding every order it needs for 2027, with a large number already booked for 2028.

That raises two uncomfortable facts the stock has to digest, and both concern the word "capacity." First, this is still underneath a cyclical memory business. DRAM contract prices rose 93–98% quarter over quarter in early 2026, which is exactly the kind of boom condition that historically does not persist, and memory orders are the lumpiest part of ASML's book — the same customers pulling in machines now have slowed deliveries before when the cycle turned. Second, ASML is not letting the scarcity compound into infinity; it is adding about 30% a year. That 30% is future revenue, but it is also future supply catching up to today's demand.

The valuation argument, meanwhile, is genuinely contested at its core. Morningstar's fair value for ASML sits near €850 with a wide-moat rating — a level the stock has now left far behind — while Morgan Stanley raised its target to €1,400 in January on 2027 earnings expectations. An €850-to-€1,400 spread on the same business is not analysis; it is a bet on whether the memory upcycle and the EUV capacity ramp both run clean for two more years. The market at roughly 50 times earnings and 16 times sales has already paid for the version where they do.

So the honest summary is a split verdict. The structure is real: ASML is the narrowest node in the chain that turns new data centers into usable memory, and the memory boom has made it half of the order book. What the investor owns at $1,700 is that confirmed structure at a multiple that has already absorbed the discovery — with the memory cycle, not the technology, as the swing factor. Watch the quarterly net bookings split between memory and logic, whether ASML commits to another 30% capacity step for 2028, and whether DRAM pricing starts to normalize. Those are the numbers that will tell you whether the chokepoint is still scarce, or whether the market simply paid for it too early.