ASML raises the number twice, and the SOX sells off anyway

ASML raised its 2026 revenue guidance to €43–45bn, up from €36–40bn — the second lift this year. Q2 net sales came in at €9.33bn against a €8.8bn consensus, net income at €2.92bn against €2.62bn. The Philadelphia Semiconductor Index sold off on the same day. Both facts are worth holding in mind at once.

What was actually filed. On the Q2 2026 earnings call, CEO Christophe Fouquet attributed the raise to "accelerating AI-related capacity investment and increased visibility into longer-term demand".¹ The order book was described as "extremely strong".¹ Gross-margin guidance moved to 54–56%. And ASML said it would add roughly 30% to EUV and DUV production capacity for delivery in 2027.¹²

That last point is the one to sit with. EUV machines (extreme ultraviolet lithography, the only tools that can print sub-5nm chips, and a product ASML is the sole global supplier of) have lead times measured in quarters, not weeks. A 30% capacity expansion for 2027 delivery means customers are contracting now for equipment that will not ship for twelve to eighteen months. That is not inquiry-level demand. That is committed capex, booked against balance sheets, showing up in an order book ASML is willing to describe in a superlative on a public call.

Why this is the cleanest AI-capex read-through in public markets. Most AI infrastructure signals are one step removed from the money — a hyperscaler capex guide, a data-centre lease, a bond issuance earmarked for "AI infrastructure". ASML is closer to the metal. If you want to build leading-edge fab capacity anywhere on earth, you buy from ASML. There is no substitute supplier. Which means the order book is, near enough, a direct index of committed advanced-node capacity investment globally.

Read that way, a €6bn midpoint guidance raise is not a beat-and-raise quarter. It is the equipment monopolist telling you that TSMC, Samsung and Intel, the only three customers who can absorb leading-edge tools at scale, have collectively signed for more capacity than ASML thought they would three months ago. The AI-performativity frame (the idea that AI spend has become large enough to be self-fulfilling regardless of near-term unit economics) fits the shape of this cleanly. The capex is real, contracted, and looking through the current inference-margin debate.

Then the SOX sold off. Which is the part of the day that deserves the second paragraph. If ASML is genuinely the read-through, why did the Philadelphia Semiconductor Index, the basket of downstream chipmakers, designers and fabless players, fall on the same tape?²

The most parsimonious reading is a capex/consumption timing mismatch. Equipment suppliers get paid at the beginning of the cycle, when foundries and hyperscalers commit to capacity. Chip designers and fabless players get paid at the end of it, when the capacity is filled with wafers that turn into product that gets consumed. Right now, capacity is being contracted faster than end-demand for chips (as opposed to end-demand for AI infrastructure, which is a slightly different question) is expanding. That is a familiar shape in semiconductors. It is not a new pattern. The AI overlay is what is new.

The polite version is that ASML sees the money the customers have already committed, and the SOX is trading the money the customers are trying to earn back. Both can be right at once, for a while. They do not usually stay right at once forever.

The China absorption. Fouquet's phrasing that the raise came "despite uncertainty over sales to China" is the tell worth pulling out.¹³ ASML's China revenue has been constrained by US export controls on all EUV systems and on advanced DUV. China was a large share of ASML's revenue in 2023 before controls tightened. If that share has compressed materially, and the company can still raise the full-year midpoint by roughly €6bn, the Western AI-capex increment is large enough to be doing more than offsetting the China drag — it is doing so with room to spare.

That is worth naming plainly. The guidance raise is not neutral to the China question. It absorbs a real, ongoing revenue loss and still lands where it lands. Which tells you something about the size of the Western capex wave that a clean beat against flat China would not.

High-NA as a mix-shift signal. The other detail on the call was that a major foundry has now taken high-NA EUV, the next-generation lithography step, to production-level deployment.¹ The unit economics of high-NA are different: fewer machines, higher price per machine, higher gross margin per machine. If the mix is shifting toward high-NA earlier than planned, that helps explain the 54–56% gross-margin guide. It is not just volume. It is a richer per-unit product mix landing on top of the volume.

That said, the customer concentration at high-NA is severe. TSMC, Samsung, and Intel are the credible buyers. Intel is restructuring. A retrenchment at any one of the three would hit the ASP-mix story harder than it would hit the headline volume story.

What could break this reading. ASML's 2022–23 order book was also described in strong terms, and it corrected. Foundries over-order during capacity expansions; they defer and cancel when demand softens. A 30% capacity build for 2027 is a bet by ASML's customers that AI-driven wafer demand is durable through the decade. If inference economics compress hard, or if the frontier labs' revenue does not scale to justify the compute, some of those commitments will be renegotiated. This has happened before in this industry. It will probably happen again.

The right frame, I think, is that ASML's print is the strongest primary-document evidence available that the AI-capex commitment is real and multi-year — and that the SOX sell-off is the market pricing the uncomfortable question of who, downstream, actually earns the return on all that committed capacity. The first question has an answer today. The second one does not, yet.

Footnotes