EUV lithography still cannot make enough light, and that single shortfall is why the machines I expected in fabs by now are still sitting in research labs. The source meant to throw 13.5nm photons at a wafer hundreds of times a second is managing a fraction of the power a production line needs. So in late 2009 the leading edge is held together by the technology ASML already shipped with 193nm immersion, pushed far past anything anyone thought it could survive. I have spent the year watching the EUV timeline slide to the right, and I have stopped trusting any date attached to it.
Immersion on its own tops out around the 40nm half-pitch, and getting below that without EUV means splitting one dense pattern into two coarser masks, exposing them in sequence, and stitching them back together on the wafer. Double patterning works. It is what keeps 32nm alive. But it is not free: every extra mask doubles part of the cost and hands you another chance to misalign, yields sag, cycle times stretch, and the price per wafer climbs past comfortable. TSMC is living this right now, where 40nm turned into an immersion yield crisis before the process settled down. The whole point of EUV was to dodge exactly this. Instead double patterning has quietly become the load-bearing trick of the leading edge, and the longer EUV slips the more masks everyone is willing to stack.
ASML does not even build the light source itself; it buys that from outside, and the physics gives no quarter. You fire a high-power laser at tiny droplets of molten tin, the tin flashes into plasma, and the plasma radiates a thin sliver of usable 13.5nm light, with most of the energy leaving as heat and as debris that fogs the collector mirror. To run a fab you need well over 100 watts of clean EUV power at the wafer held steady for months, and today’s demonstrated numbers sit at a small fraction of that. ASML has orders for the NXE:3100, a pre-production tool, from a handful of customers, with the first systems due to ship next year so people can start learning the process. Nobody is making chips to sell on it. The source is the wall, and until it clears the power a fab actually needs, EUV stays a spectacularly expensive science project.
Two years ago you could still call the ASML-versus-Nikon race close. Not now. The leading-edge orders go almost entirely to Veldhoven, Nikon’s immersion tools keep losing the design wins at the customers who matter most, and Canon left the high end earlier and has not looked back. I take no pleasure in watching a rival fade, because a field with one dominant supplier tends to drift toward expensive and slow, but that is plainly where this is going. The further ASML pulls ahead on immersion, the more it can bankroll the EUV gamble nobody else can afford to chase.
The 2008 crash gutted chip demand, and 2009 has been brutal for anyone selling fab equipment. ASML cut staff, put people on shortened hours, and watched orders evaporate for a few quarters. Through all of it the EUV budget did not get cut, which is either conviction or plain stubbornness, and I cannot tell which from the outside. Burning research money on an unproven technology during the deepest downturn in a generation is not the safe move. ASML made it anyway.
The roadmap still pencils EUV in for the nodes after 32nm, maybe 2012, maybe 2013, and the customers keep nodding along in public. I have heard that promise for three years running, and every year it slides another year out. Immersion and double patterning will carry the load well past the dates on those slides. ASML’s real problem now is stamina: paying for a machine that does not work yet, quarter after quarter, until it maybe does. Three years in, the source still cannot hold the watts, and I know which way I would bet.