Somebody is putting a memory price into a 2027 bill of materials this week, and a slide with 2033 on it does nothing for them. Here is what buyers actually do with a dated fab roadmap: they check whether the date lands before the thing they are quoting ships. Here is what suppliers tell themselves buyers do: they read the whole decade and feel covered.
The reason the masks matter at all is optical. High NA systems use anamorphic optics to resolve finer features, and the trade is that they expose roughly half the area of the older generation in one pass [9]. Processors designed by Nvidia and Google already run to about 800 square millimetres, which sits at the ceiling of what current tools print at once, so on a High NA machine a die that size has to be stitched from two exposures, costing throughput and introducing a seam where yield problems like to live [10]. TSMC's own description of the payoff is that bigger masks increase fab productivity, lower chipmaking costs and remove stitching constraints [17].
The dates do not arrive in that order. TSMC intends to run High NA in high-volume manufacturing at advanced nodes from 2030 on the existing 6-inch masks, and ASML chief executive Christophe Fouquet framed adoption as progressive, 6-inch first and 12-inch later [11]. Pilot masks come in 2031 and the systems are not expected ready for advanced-node production until 2033 [4]. That is at least three years of TSMC High NA output with the seam still in it for any die above the single-exposure limit [20], and two years between the pilot line and production readiness [21]. Samsung's DRAM target sits five years ahead of the mask date, so the memory plan is not waiting on masks at all [22].
The throughput number is conditional in its own wording. ASML chief technology officer Marco Pieters told Reuters that "if we're going to pull it off as an industry, then you'll actually see that the productivity of those systems will go up by 40 percent" [8]. Pulling it off means blank suppliers, pellicle makers, mask writers, inspection and repair tools, and every handling system in a fab, none of which ASML owns [13]. TNW describes the result as a decade of coordinated capital spending across companies that compete with each other, a reasonable bet that is still a bet [19]. Intel bought into the machines early, when they cost roughly $400mn each, and each High NA tool is worth several times a conventional EUV system to ASML [15]. Engadget's caution is worth keeping next to the calendar: ramping the technology is not easy, and both companies may hit bottlenecks that thwart their planned production dates [16].
Two questions sort a roadmap like this, and neither is about the timeline. First, whether a single company controls the date. Second, whether the date lands inside the horizon of the purchase in front of you. Samsung's 2028 DRAM clears both for anyone specifying memory for 2028 and later, which makes it the one line here worth raising with an account team. TSMC's 2030 is single-owner and outside most hardware horizons, so watch it rather than plan on it. The 2031 and 2033 mask dates are owned by no one in particular and sit past almost everybody's horizon, which puts them closer to a standards charter than to bookable capacity. The earliest milestone in the whole announcement is 2028, so for anything shipping before then, this contains no dates [25].
