ASML Says AI Chip Tool Demand Won't Be Met for at Least Two Years

Full analysis

ASML, the only company that makes the extreme ultraviolet lithography machines that print leading-edge AI chips, told its earnings call that demand is running years ahead of what it can build. CEO Christophe Fouquet said customers ask for more tools daily, and the company is now studying capacity for 2028 and beyond, on top of a 30% expansion already promised and already spoken for. Customer reservations stretch to 2029, and in some cases 2039.

Here's what this means for anyone buying AI, renting GPUs, or paying an inference bill. This is hard to undo, but not for you. For ASML and the chipmakers, the lead times are a decade. For you, the decision is whether to keep planning as if compute gets cheaper and more plentiful every year. Nothing sets a deadline here except your own contract renewals and capacity reservations. The constraint isn't capital. It's a single optics factory in the Netherlands.

The Skeptic. An ASML earnings call is where demand gets talked up. "Customers asking every single day" is a line that props up pricing power and justifies capex to shareholders. Reservations into 2039 are not firm purchase orders, they're options and placeholders. We've seen this movie. The 2023 chip glut followed a 2021 shortage that was going to last forever. If hyperscaler spending discipline returns, or if models keep getting more efficient per answer, a chunk of that 2029 book evaporates. Physical limits make a tidy story. Tidy stories are how cyclical demand gets mistaken for a wall.

The Compute Pragmatist. Everyone counts GPU racks. Wrong meter. The real ceiling is ASML shipment slots, and there is no second source for an EUV machine. Thousands of custom optical and mechanical parts, multi-year build times, one vendor. TSMC, Samsung, and SK Hynix are all bidding for the same pipeline. Elon Musk's "Terrafab" can pour concrete forever and still not make a leading-edge chip without a lithography answer. The 30% expansion is gone before the machines ship. What this means in your bill: expect leading-edge compute to stay expensive and rationed through 2027 at least. Price drops will come from efficiency, not abundance.

The Builder. Plan as if compute is a fixed resource, because it is. The teams that win the next two years treat scarcity as a design constraint. That means smaller models that do the job, quantization (running models at lower numerical precision to cut memory and cost), speculative decoding (using a cheap model to draft and an expensive one to check), and mixing cheaper chips with the scarce ones. If your roadmap assumes you'll just rent more H200s or B200s next year at a better price, there's a hole in it. The procurement answer ran out. The engineering answer is all you have left.

The Safety Lens. A two-year gap hardens the pecking order. Whoever locked in fab access early gets the frontier to themselves, mostly the biggest US hyperscalers and TSMC's anchor customers. That concentrates who can train the largest models, and who can afford to run safety and interpretability work at that scale. The upside, if you want one: a narrow chokepoint makes compute governance tractable. Export controls and know-your-customer rules for fab access are far easier to enforce when the bottleneck is one tool from one vendor. Regulators will notice that.

The Enterprise Buyer. If you sign multi-year commitments with a cloud or model provider, this changes your leverage. Capacity you don't reserve, you may not get at the price you want. But don't swallow scarcity pricing just because ASML put a decade on a slide. Ask your provider what their actual committed silicon pipeline looks like through 2027, and whether your rate card is protected if their costs fall from efficiency gains rather than new supply. The vendors have every reason to pass scarcity through and keep the savings.

Where they split. The Compute Pragmatist and the Skeptic are looking at the same reservation book and seeing different things. One sees a decade-long physical wall. The other sees option bookings padded to flatter a supplier. Both can't be right about 2029. The Builder doesn't care who wins that argument, because the engineering answer, use less compute per job, pays off whether the wall is real or the glut returns. That's the move with no regret. And the Safety Lens and the Enterprise Buyer agree on the fact but want opposite things from it: concentration makes rules easier to write and makes your negotiating position worse.

What this hinges on is one thing you can actually watch: does leading-edge compute get cheaper per unit of work over the next two years, and does that come from more supply or from efficiency? If prices fall while ASML stays sold out, the Skeptic's "cyclical demand" read loses and the "efficiency, not abundance" read wins. That's the fork.

Prediction: The published price per million tokens for the current flagship models from OpenAI, Anthropic, and Google will be lower in April 2027 than it is in October 2026, and the drop will come from model efficiency rather than from any easing of chip supply.

Confidence: Medium. The efficiency trend is consistent, but a demand spike could stall list-price cuts.

Why: ASML being sold out through 2029 means raw chip supply will not loosen in this window, so any price relief has to come from doing more with the same silicon. Every flagship cycle since 2023 has cut the price of a fixed capability through smaller distilled models, sparse designs where only part of the model fires per query, and cheaper serving, and labs keep competing on exactly that number. The opposite outcome, flat or rising list prices, would require the labs to stop competing on cost while a supply wall gives them cover to hold prices, and there's no sign the price war has paused. The interesting part is the cause: prices fall even though chips stay scarce, which means the scaling story and the pricing story have decoupled.

Revisit by 2027-04-20: We're right if the per-million-token list price of the then-current flagship from at least two of OpenAI, Anthropic, and Google is lower than its October 2026 equivalent, with no public report of ASML EUV supply easing. We're wrong if those prices are flat or higher, or if ASML announces a material near-term capacity jump that explains the drop.