ASML closed 2025 with net sales of $39.16 billion and net income of $11.5 billion, propelled by AI demand that shows no sign of cooling. Then it told the market it cannot confirm growth for 2026, the stock fell 11%, and roughly $30 billion in value evaporated. Record year, unconfirmable future. Both things are true at once, and the reason they are both true has less to do with financial guidance than with a physical constraint sitting inside the company’s most expensive product.
High NA EUV systems cost more than $400 million each and use shorter effective light paths to etch finer transistors than the previous generation. They also cannot print the largest chip designs in a single exposure. The imaging field is smaller. To make a big die, you split the pattern and expose it in multiple steps, which means more process complexity, more cost, more alignment risk. ASML has a fix planned, but it is not expected until roughly 2031 to 2033. That is a long time in an industry that reprices itself every quarter.
Why the exposure field matters to anyone building AI systems
I work on agent architecture, not lithography, and I want to explain why this belongs on a blog about agent intelligence rather than a semiconductor trade site.
Modern AI accelerators are enormous by design. The reason is not vanity. Keeping compute and memory physically close reduces the energy and time cost of moving data, and moving data is the dominant cost in transformer inference. A large die means more on-chip SRAM, wider internal buses, fewer trips off-package. When you cannot print a large die in one shot, you get one of two outcomes: you stitch exposures together at higher cost and lower yield, or you break the design into smaller chiplets and connect them.
The industry has mostly chosen the second path, and it works. But the connection is never free. Every boundary you introduce between pieces of silicon is a boundary where bandwidth drops and latency rises. Those boundaries propagate upward through the entire stack until they show up in the thing an agent developer actually feels: how long a tool call takes to come back, how much key-value cache you can hold before you start evicting, how many concurrent agent sessions a single node can serve.
Constraints move up the stack, they do not disappear
This is the part I find genuinely interesting. Physical limits at the bottom of the stack tend to reappear as design patterns at the top, usually several years later and usually unrecognized as such.
- Limited on-chip memory pushed us toward aggressive quantization and attention variants that trade precision for footprint.
- Costly inter-chip communication pushed us toward mixture-of-experts routing, where only part of the model activates per token.
- Expensive long-context inference pushed agent frameworks toward retrieval, summarization, and context compaction instead of simply holding everything in the window.
None of those were purely algorithmic discoveries. They are adaptations to hardware boundaries, and reticle-limited exposure fields are one of the oldest boundaries in the set. If the single-exposure fix genuinely lands in the early 2030s, then the accelerators shipping between now and then will be assembled from parts rather than carved from one piece. Agent architectures built on top of them will keep inheriting a world where locality is precious and communication is the tax you pay for scale.
The customer signal is the loudest data point
TSMC’s Deputy Co-Chief Operating Officer Kevin Zhang told Bloomberg the company has no plans to purchase the newer generation of High NA systems. That is not a small remark. When the most important customer for advanced lithography says the economics do not work yet, it validates the concern behind ASML’s cautious framing. A $400 million tool has to earn its cost through yield and throughput advantages. If splitting exposures eats those advantages on exactly the largest designs, which are the AI accelerators driving demand in the first place, the value case narrows to a subset of layers rather than a whole process.
So the contradiction resolves cleanly. Demand for AI compute is real and reflected in ASML’s results, including $11.62 billion in fourth-quarter revenue. The pathway to selling much more of the most expensive machine runs through a capability that will not exist for another six to eight years. Strong present, uncertain bridge.
What I would take from this
Two things. First, treat hardware timelines as inputs to architecture decisions, not background noise. A fix arriving in 2031 means the design assumptions you make about memory locality and interconnect cost are stable for a while. Build for them rather than around them.
Second, be skeptical of any AI scaling narrative that treats silicon as an infinitely elastic input. The chip that runs your agent was shaped by how large an area a lens can expose at once. That is a mundane physical fact, and it is quietly steering the design of systems we describe in far more abstract language.
đź•’ Published: