High-NA is waiting on a bigger photomask

High-NA EUV was sold as a way to print smaller features without stitching two exposures together. On the scanner, that is true. On the mask, it is not. The industry still writes patterns on a 6-inch photomask that has barely changed in forty years. On a High-NA tool, that format throws away about 30 percent of the throughput. You buy a machine that can print. You then spend a third of the shift reprinting.

On 7 September, ahead of the SPIE BACUS conference in Monterey, ASML and TSMC formed a joint initiative to move the industry to a 12-inch mask. They announced it the next day from Hsinchu and Veldhoven. Samsung said it would join. Intel said it had been running a parallel large-format effort for more than three years. The dates they put on paper are not heroic. A 12-inch mask pilot line by 2031. High-NA lithography systems on the new format ready for advanced-node production by 2033.

Seven years. A scanner ordered this winter will still be depreciating when the blank size changes.

30%

Throughput lost to stitching on High-NA with a 6-inch mask

2031

12-inch mask pilot line, per ASML and TSMC

2033

Full High-NA readiness on the new format

2028

TSMC’s public date for A14 volume

Why the plate never grew

The 6-inch mask survived because the rest of the line was built around it. Blank makers, writers, inspection, pellicles, the cleanrooms that hold them, the software that fractures a design onto that field — all of it is amortised against a format that every foundry already knows. Changing the plate is not a drawing change. It is a second industry, running in parallel with the first, until the first can be wound down on mature nodes.

Low-NA EUV could live with the old plate. High-NA cannot, not if anyone intends to use the numerical aperture they paid for. Christophe Fouquet, ASML’s chief executive, was plain: the industry will keep the 6-inch mask first, then move. TSMC’s C.C. Wei said the usual thing about problems no single company can solve. That is true, and it is also why a consortium that takes until 2033 is a production fact, not a talking point.

Taiwanese trade reporting the day after the announcement put more meat on TSMC’s own roadmap than the press release did. A14 is still aimed at 2028. A13, and A12 with backside power, at 2029. High-NA is talked about from 2030. The reading in the supply chain is that early High-NA layers will still ride 6-inch masks, with the new format arriving as the ecosystem catches up. If that is right, the first years of High-NA in Hsinchu will carry the 30 percent tax. The tool will be ahead of the plate. That is the expensive way to run a shift.

The last time a format was supposed to change

The industry has tried this before, one step up the line. 450-millimetre wafers were going to cut die cost by growing the plate the chips sit on. Intel, TSMC, Samsung, the tool companies, the governments. Billions. The wafer size did not change. The tools got more expensive, the node kept shrinking on 300 millimetres, and 450mm became a seminar topic. The lesson was not that formats never change. It was that a format changes only when the companies that already make money on the old one will lose more money by keeping it.

The 12-inch mask has a better case than 450mm did, because High-NA is already in the building and already wasting a third of a shift. The scanner is bought. The waste is measured. That is a different political economy from a wafer size that required every tool in the line to grow at once. Mask is a path, not the whole fab. It is still a path with blanks, writers, inspection and pellicles that have to move together. A consortium with a 2033 date is how you admit that.

The bill of materials

A mask shop that stays on 6-inch will still have customers. Mature nodes, and the long tail of layers that never see High-NA, will pay for that line for a decade. A shop that wants the leading edge will capex twice: once for the new format, and once for the people who can write it. Blanks, writers, inspection, pellicles. The pellicle in particular is not a commodity. High-NA already argues with heat and transmission on the current size. A larger plate makes that worse before it makes it better.

Stitching is not an abstract loss. Two exposures have to land on each other. Overlay error on a High-NA layer is yield, not a metrology curiosity. The 30 percent is time. The rest is the chance that the two fields disagree. Foundries live with that on some layers today. They will live with more of it until the plate grows. Die-cost models that use the scanner’s brochure wafers-per-hour are already lying, politely, in the customer presentation.

Intel’s claim to a three-year head start is worth taking seriously and not taking as a schedule. Intel has processed more than a million High-NA wafers, including production layers on 18A. That is real tool time. It is not a qualified 12-inch mask line. The consortium’s 2031 pilot is the first date that binds more than one company. Until then, every High-NA layer is a stitching problem with a different name.

For a foundry customer the relevant question is not when 14A or A14 appears on a slide. It is which critical layers will still be stitched in 2028 and 2029, and what that does to die cost and to cycle time. For an equipment buyer it is which mask-path tools become stranded when the blank grows, and which can be rebuilt. For a materials house it is who is qualifying the 12-inch blank, and whether that qualification is a TSMC list, an Intel list, or both. Those lists will not be the same.

Who has to move

The names on the mask path are not household names outside the industry, and they are the ones who will spend the money. Dainippon Printing, Toppan, Photronics, the blank makers behind them, the inspection houses. A leading-edge mask set already costs what a programme can feel. High-NA layers cost more. A 12-inch blank is a second qualification, a second writer, a second inspection recipe. Those companies will not tool up on a press release. They will tool up on a TSMC volume forecast they believe, and on Intel’s if they believe that too. The consortium is the signal. The purchase order is the act.

A foundry customer lives with this as die cost and as schedule. The NDA that starts this winter for a 2030 product will be written against a lithography assumption. If that assumption is High-NA brochure throughput, the costed unit is wrong. If it is High-NA with stitching through 2029 and a 12-inch mask sometime after 2031, the unit is uglier and more honest. The companies that will ship are the ones whose cost models already have the 30 percent in them.

What 2031 actually means

A pilot line in 2031 is not a second source. It is a few tools, a few blanks, and a process of record that does not yet have to survive a 90 percent yield gate. Production systems in 2033 is the date that matters for anyone signing a foundry NDA this winter for a 2030 product. The silicon may be designed against High-NA. The mask that prints it may still be the old one. Cost models that assume the scanner’s brochure throughput are already wrong.

There is a pattern here that is not unique to lithography. The machine arrived first. The thing the machine eats did not. Capex committees sign the scanner because the scanner has a name and a delivery slot. The plate is a supplier problem until the supplier problem is the shift. We have seen the same shape on CoWoS, on high-NA’s cousin in advanced packaging, on dry rooms built before offtake, on mills built before a grade has a buyer. The bottleneck moves one step off the famous tool and onto the consumable.

Dual-path is the only sane industrial answer, and it is also the expensive one. Keep 6-inch. Staff 12-inch. Name an owner for each. Do not rip out the old line because a conference said the new format is coming. Mature nodes will still want that plate in 2035. The mistake is to treat the consortium as a reason to wait, or as a reason to scrap. It is a reason to put mask format on the same Gantt as the scanner — not as a footnote under ‘ecosystem.’

The industry has just said the next lithography problem is not the light. It is the plate. A scanner ordered this winter will still be running when the format changes. That is now a line-item, not a paper.