The most advanced processor you can buy in a phone this month is made in Taiwan, and the most advanced one Intel promised to ship this year does not exist in any quantity you can find. That inversion is the whole story of TSMC’s 10nm node. The A11 in the new iPhones is built on it in real volume, while Intel’s own 10nm keeps sliding out of every quarter it was promised in. For the first time since TSMC started, the leading edge you can actually order runs through Hsinchu and not Oregon.
10nm is, like 20nm before it, more of a fast bridge than a destination. It is a FinFET refinement of the 16nm generation, tighter pitches and a denser cell, printed with triple patterning on the critical layers because the light is still 193nm and the geometry now needs three masks where 20nm needed two. That third pass is more cost and more delay, and everyone in the building knows the node that actually pays off is 7nm right behind it. So 10nm ramps hard, ships Apple’s volume, and hands the baton along fast.
Intel spent thirty years as the company whose transistors arrived first and best, and its 10nm was supposed to extend that streak. Instead it has slipped again and again, and it is nowhere near the volume TSMC is running right now. There is a real subtlety here: Intel’s 10nm is denser than TSMC’s, packed closer to what TSMC will call 7nm, so on paper it is the more advanced node. On paper. A denser transistor that will not yield loses every comparison that matters to a looser one shipping by the millions in a phone you can buy today, and that is the comparison TSMC just won.
Samsung has its own 10nm shipping, and the high-end mobile market has quietly split along it. Apple builds the A11 at TSMC. Qualcomm builds its flagship Snapdragon at Samsung. The two leading foundries each own a piece of the premium SoC market outright now, TSMC on better than half of all foundry revenue and Samsung a distant but real second, and the contest between them is no longer about whether they can match Intel’s transistor, which they did at FinFET, but about which of the two yields the next node first. Intel has fallen far enough out of the leading-edge foundry conversation that it is barely in this paragraph.
The thing that makes 7nm worth sprinting toward is what finally shows up there. After a decade of being two years away, extreme ultraviolet lithography is close enough to real that TSMC can plan a version of 7nm around it, trading stacks of triple-patterned DUV masks for a single EUV exposure. The 10nm node is the last one built entirely on the old light. If EUV holds together in volume, and that if has burned a lot of money and schedule across the whole industry, the multipatterning cost spiral that has made every node since 20nm more expensive finally gets a release valve.
10nm will be a footnote, a quick node that existed to carry Apple’s volume for a year and tee up 7nm. The thing worth remembering about it is the date the lead changed hands. For its entire history TSMC chased Intel: caught it on materials at 28nm, caught it on structure at 16nm, trailed on the calendar the whole way. At 10nm it stopped chasing, because the company it was chasing tripped over its own next node and has not gotten back up. The lead is TSMC’s now. The only open question is whether it can hold it against Samsung, because Intel just stopped being the one setting the pace.