ASML Just Posted a Record Quarter — and Warned It Still Can't Build Enough Machines for the AI Boom


A record quarter was not the headline

ASML reported its third-quarter results on Wednesday, October 14, and by any normal measure they were exceptional. Total net sales reached €11.7 billion, a record for the company and up sharply from €9.3 billion in the second quarter. Net income came in at €5.0 billion, gross margin was 57.5%, and the company sold 113 new lithography systems in the quarter, alongside €3.3 billion in service and upgrade sales from its installed base.

The outlook rose with the results. ASML guided fourth-quarter sales of €14.2 billion to €15.2 billion at a gross margin of around 57%, and lifted its full-year 2026 sales expectation to €44 billion to €45 billion.

None of that was the most important part of the announcement. The important part was what chief executive Christophe Fouquet said about the next two years: customers keep raising their equipment demand forecasts, driven mainly by how large and how long they expect AI investment to run, and ASML’s supply as currently planned for both 2027 and 2028 will not fully meet that demand.

Read that slowly. The company that sits at the very start of the advanced chip supply chain — the one supplier every leading chipmaker depends on — is saying, in an earnings release, that even its planned expansion is not enough. And some customers, Fouquet added, are already expressing interest in placing orders for 2029 and even 2030.

Why ASML gets to make that call

To understand why this statement matters more than a typical earnings beat, you have to understand ASML’s position. The Dutch company, headquartered in Veldhoven, makes lithography machines — the systems that use light to print the tiny circuit patterns that become microchips. Its extreme ultraviolet (EUV) machines are, in practice, irreplaceable for making the most advanced processors, including the AI accelerators that fill modern data centers. There is no second source a chipmaker can call if ASML’s queue gets longer.

That gives ASML an unusual vantage point. It does not see one company’s roadmap or one government’s industrial policy. It sees the combined, forward-looking demand of essentially the entire leading-edge chip industry, expressed as orders and forecasts years in advance — because its machines take a long time to build, ship, and install, and fabs are planned around them.

So when ASML says demand will outrun planned supply in 2027 and 2028, it is not offering an opinion about the AI boom. It is reporting the arithmetic of its own order book and its customers’ stated plans. The company also said it is investigating how to increase capacity for 2028 beyond the 30% increase it had previously communicated, and pointed to its productivity-upgrade business — making customers’ existing machines produce more wafers — as another way to add capacity without waiting for entirely new tools.

One honest caveat belongs in the numbers: the quarter’s net income included a €500 million gain from revaluing ASML’s stake in Mistral AI. That does not change the demand story, but it does mean the headline profit figure is not purely a lithography figure. The record sales, the raised guidance, and the supply warning stand on their own without it.

Two days, two different bottlenecks

This warning lands in an interesting sequence. On Monday, at a conference in Singapore, Broadcom’s Hock Tan and Intel’s Lip-Bu Tan argued that capital — not chips or power — is now the biggest constraint on the AI buildout, because the money has to earn a return within a few years or the spending stops.

ASML’s results do not contradict that argument. They complete it. Put the two together and the AI buildout now has at least two hard constraints operating at once:

  • Money has a clock. The capital funding data centers and fabs needs a credible return, on a timeline investors can see.
  • Machines have a queue. Even with unlimited money, the tools needed to make the most advanced chips cannot be produced faster than ASML and its own supply chain can build them.

That second point is the one money cannot simply fix. A hyperscaler can raise more capital in a quarter. It cannot conjure an EUV machine that was never built, and it cannot move itself up a queue that its competitors are also trying to lengthen — to 2029 and 2030, if Wednesday’s statement is any guide. In a supply-constrained world, early commitments become a durable advantage, and latecomers pay in time rather than just price.

What this means if you buy, build, or plan around AI compute

Most businesses will never order anything from ASML. Almost all of them will feel this announcement anyway, because it sets expectations for the price and availability of advanced compute through the end of the decade.

For enterprise buyers: Treat relief from scarce, expensive frontier compute as unlikely in 2027 or 2028. Multi-year cloud and GPU commitments will continue to favor the largest buyers, who can commit earliest and biggest. If your AI roadmap quietly assumes that top-tier capacity becomes abundant and cheap in the next two years, Wednesday’s results are a reason to revisit that assumption — and to put more weight on efficiency: smaller models where they suffice, better utilization of capacity you already have, and workloads designed to tolerate being scheduled rather than run instantly.

For chipmakers and governments: The queue is the strategy. Nations pursuing domestic semiconductor capacity and chip firms expanding advanced fabs are all, in the end, waiting on the same supplier’s output. ASML’s search for extra 2028 capacity will be watched as closely as any product launch this year, because whatever it finds — or fails to find — effectively caps how much leading-edge capacity the world can add.

For investors: Note what kind of signal this is. This is not a startup projecting demand for its own product. It is a monopoly-position supplier with a multi-year order book raising guidance for the third time and warning that planned supply falls short two years out. The risk in that story is not that demand is imaginary — ASML’s customers are committing real money years ahead. The risk is the mirror image of Monday’s warning: that the investment cycle turns before those 2029 and 2030 expressions of interest become delivered, revenue-generating machines. Interest, ASML itself notes in the fine print of such forecasts, is not the same as a firm order.

For AI builders: Scarcity at the top end quietly strengthens the case for everything happening below it — local and on-device AI, distilled and open-weight models, and inference optimization. When frontier capacity is rationed by a physical queue, getting more useful work out of each available chip stops being a cost nicety and becomes a competitive strategy.

The honest version of the story

It would be easy to read ASML’s warning as simple boom confirmation: demand is huge, buy everything. That misses the more useful reading.

What ASML actually described is a buildout now pressing against a physical limit at its foundation, at the same moment its financial limit — the return on the capital funding it — is being openly questioned by the chip industry’s own chief executives. Both things can be true at once: the demand is real enough that customers are asking about 2030, and the economics still have to prove themselves before those years arrive.

That combination is what maturity looks like in an infrastructure cycle. The question is no longer whether AI investment is large — a record €11.7 billion quarter from the industry’s sole critical supplier settles that. The questions are who secured their place in the queue, who is paying for capacity they may not fill, and whether ASML can find capacity in 2028 that it does not currently have planned.

For anyone planning around AI, the practical takeaway is plain: plan for scarcity, not abundance, at the frontier through at least 2028 — and build the efficiency and flexibility to thrive if the constraint, not the hype, is what lasts.