The Cymer acquisition closed last year, so the EUV light source ASML used to buy behind a supplier contract now sits in-house, and with it the hardest engineering problem the company has ever owned. There is no longer anyone else to blame when the power comes up short, and it is still coming up short. Everything around the light is ready, the scanners, the optics, the handling, all of it waiting. The light itself is what will not cooperate, the same as five years ago.
Fifty thousand times a second, a droplet generator fires beads of molten tin into a vacuum chamber, each one a few hundredths of a millimeter across. A carbon-dioxide laser hits each droplet twice: a weak pre-pulse flattens it into a pancake, then a powerful main pulse blows it into plasma that glows at 13.5nm for a few billionths of a second. A giant curved collector mirror gathers that light and funnels it toward the scanner. Do all of it reliably, fifty thousand times a second, for months on end, without the mirror fogging over with tin debris, and you have a production source. It is a small, violent star, and nobody has built one that lasts yet.
A fab running EUV in production needs a source holding around 250 watts, steady, all day, and that single number is the whole game. When I last wrote about this, the demonstrated figures were a small fraction of it. They have climbed since, into the tens of watts in stable operation and higher in short bursts, but the gap between a burst in a lab and 250 steady watts on a factory floor is enormous. Higher power means more tin, hotter optics, a bigger laser, and a collector mirror that degrades faster, so every watt ASML adds makes the next one harder. This is the climb that decides whether EUV ever turns into a business.
Even with the source fixed, the masks would still be a problem. EUV masks are reflective instead of transparent, and keeping them defect-free at these scales is brutal. Protecting them with a pellicle, the thin membrane that shields a mask from particles, is far harder at 13.5nm because almost any material you put in the beam swallows the light you cannot spare. Throughput takes the hit and uptime stays poor, with dose stability drifting in between. Each problem on its own looks solvable. Stacked together, they keep shoving the date when EUV is good enough to trust with real products further out, a quarter at a time.
A single EUV scanner already runs more than a hundred million dollars, and that is before a fab spends more on the infrastructure to feed it. ASML is asking customers to wager enormous sums on tools that do not work reliably yet, on the promise that one day they will. I understand why the three big customers keep writing the checks anyway. The alternative is a future where shrinking the transistor gets so expensive through endless multi-patterning, the way it already bites on the leading FinFET nodes, that the whole economics of chipmaking stalls out. EUV is the way out of that, assuming it works, and the customers are already paying billions to find out whether it does.
I have been writing some version of this for years, and 2014 changes one thing: ASML owns the part that keeps failing. No supplier to coordinate with, no one else to push, no one else to blame if the watts never come. The most expensive bet this industry has ever made now rides on one company’s stamina, and a failure would land in Veldhoven alone. My guess is they get there, late and well over budget, because they have boxed themselves in with no other way out. Ask me when I can buy a phone with a chip printed on one of these, and I still will not give you a year.