ASML’s 2028 Capacity Plan Still Leaves a Queue

Despite a previously communicated 30% capacity increase for 2028, ASML says its planned supply of lithography machines will fall short of demand in 2027–28. Customers have also expressed interest in 2029–30 orders, years before those tools could prove their value inside a working fab.

ASML reported €11.66 billion in third-quarter sales, raised its 2026 sales forecast to €44 billion–€45 billion and is investigating further expansion.

AI is turning lithography from a cyclical equipment market into a multi-year capacity-allocation system: ASML’s scarcity gives chipmakers reason to seek access far ahead, while those same signals can synchronize investments in tools, fabs and adjacent inputs that may come online at different times from the demand they were built to serve.

Equipment makers learned how quickly a full pipeline can empty

The semiconductor equipment business was built around customers whose capital spending rose and fell with their chip markets. Toolmakers had to prepare for long manufacturing lead times without mistaking every proposed fab expansion for permanent demand. In 2023, Lam Research reported a 31% year-over-year revenue decline, its third consecutive quarterly decline, amid a soft market for chipmaking tools. The machinery was still valuable; customers’ willingness to buy it had changed.

AI has lengthened the planning horizon because leading-edge foundries cannot readily replace an EUV supplier when a delivery slot disappears. The cost of waiting is asymmetric for a chipmaker: seeking access early can preserve an option to expand, whereas waiting for firmer end-market evidence may leave no tool available on the desired schedule. A chipmaker can seek a slot before it has built its next fab or proved how heavily it will use the tool.

ASML’s 2027–28 supply outlook describes a genuine constraint against its current plan. Its reported interest in 2029–30 orders describes something narrower. The available disclosure does not establish binding High-NA quantities, delivery schedules or common reservation terms for Intel, Samsung and TSMC. Buyers can be prudent in seeking a place in the queue without providing ASML a settled measure of how many wafers they will ultimately need.

ASML’s queue reaches beyond its own factory floor

A lithography delivery slot matters only if a foundry can put the machine to work. TSMC was reportedly asking customers to apply for N2 production allocation as far out as the second quarter of 2027, with large allotments nearly sold out for the following two years. That is a second queue, for wafers rather than tools. A chip designer seeking wafer access and a foundry seeking machine access are making related decisions, but neither queue proves how much finished AI hardware buyers will take years later.

Each participant sees the shortage closest to it. A designer sees scarce foundry capacity; a foundry sees scarce lithography equipment; ASML sees orders that require parts from a supply chain spanning roughly 2,000 companies. If each participant asks its supplier for enough capacity to avoid being the one left out, the upstream request can grow larger than the eventual need. Long lead times make that difference difficult to correct: a supplier cannot wait until the final chip buyer is certain before deciding whether to expand.

ASML is not holding machine output constant. It planned to make at least 60 standard EUV systems in 2026, 36% more than it sold in 2025. That expansion is a serious answer to present scarcity. It also means the relevant question changes with time: how much effective wafer output will those systems provide when they are installed, qualified and running alongside other new capacity?

EUV has several clocks, not one delivery date

Intel began high-volume production using EUV at its Ireland plant in 2023. That milestone required more than receiving a machine: Intel had to turn lithography into a repeatable manufacturing process. High-NA EUV adds another transition. Samsung committed to use High-NA by 2028 and TSMC by 2030, joining Intel; the companies and ASML also agreed to move from 6-inch to 12-inch photomasks. Machine availability, mask changes and process qualification therefore do not share a single timetable.

In 2016, reporters wrote that EUV could enable mass production as soon as 2018. By 2026, Samsung and TSMC had made separate commitments for High-NA adoption late in the decade. Those dates do not describe one machine missing one deadline: standard EUV and High-NA are different generations. They show why a technology roadmap is an unreliable substitute for a fab’s demonstrated output.

Productivity can also change the denominator. ASML researchers raised EUV light-source output from 600 watts to 1,000 watts, an advance projected to yield 50% more chips by 2030. The projection is not a measured industry-wide gain, but it is strong counter-evidence to a simple story in which only more machine shipments relieve the shortage. A fab with more productive tools could add wafer output without adding scanners at the same rate. Conversely, a delivered High-NA system undergoing qualification adds little immediate production.

The other toolmakers are committing capital on their own schedules

ASML prints patterns; fabs need other equipment to turn those patterns into working chips. Lam Research plans to invest more than $3 billion over five years in its global research-and-development network, increasing experiment capacity by more than 50%. Applied Materials opened a $500 million manufacturing campus in Singapore, where roughly half its production capacity will be located. Lam’s experiment capacity and Applied’s manufacturing capacity solve different problems from ASML’s scanner output. They nevertheless become fixed commitments before the industry can observe all the AI demand those investments anticipate.

More suppliers investing can shorten a bottleneck. It can also give several companies the same exposure to a changed customer plan. ASML must coordinate components across thousands of suppliers, while a fab must coordinate tools that perform different steps. A timely scanner cannot compensate for a missing process tool, and a new process-tool campus cannot make a scanner arrive sooner. The broader financing of AI chip capacity reaches upstream because none of these businesses can add industrial output at the moment a finished-chip customer places an order.

A working scanner is not a finished AI system

TSMC’s N3 logic-wafer capacity has been described as one of the AI industry’s largest constraints. AI demand has also strained lasers, optical substrates, fiber and connectors; industry sources have identified printed circuit boards and chip substrates as additional constraints. Foundries need memory and advanced packaging capacity to turn leading-edge logic into usable AI hardware. Data-center builders, in turn, need sites, components and power.

Those shortages complicate both sides of a lithography forecast. If packaging or optics remain scarce, more lithography machines need not produce proportionally more deployed AI systems. If those constraints clear after foundries have expanded against today’s wafer shortage, several layers may deliver capacity together. A foundry can have made a sound decision about its own immediate constraint and still meet an end market different from the one its construction schedule assumed.

China gives ASML a second demand cycle

ASML’s business cannot be read solely as a gauge of frontier-AI spending. The company expected China to account for about 20% of 2026 sales after its third-quarter China sales rose 109% from the previous quarter. Chinese fabs had acquired about 343 immersion DUV tools by early 2026, including about 270 from ASML. These older-generation tools can be adapted for 7-nanometer logic and advanced memory; they are not the same product as a High-NA EUV system.

Export restrictions and Chinese efforts to reduce reliance on Western equipment give that DUV demand a different set of drivers. In 2025, U.S. House lawmakers criticized ASML, Lam, Applied Materials and other toolmakers over sales that aided China’s chip industry. Chinese fabs’ purchases can support ASML’s revenue while restrictions alter what it may sell and domestic substitution alters what those customers may eventually buy. None of those decisions has to move in step with a hyperscaler’s AI-chip budget.

The hard evidence arrives after the reservation

Memory investors were already debating in 2026 whether capacity expansion would bring another glut or whether AI demand would break the sector’s boom-and-bust cycle. ASML’s stated shortage and its planned output increase provide evidence for the first half of that debate: customers need more capacity now, and suppliers are working to provide it. They cannot settle the second half, which depends on when fabs ramp, what their tools yield and whether customers keep taking the resulting chips.

ASML has planned a 30% capacity increase for 2028; customers have yet to show what those future tools will produce. An ASML delivery slot fits neatly on a calendar; a qualified fab, a packaged chip and a buyer do not.