Two things happened to TSMC’s 7nm node within weeks of each other this autumn. Apple shipped the A12 inside the new iPhones, the first 7nm chip anyone can buy, and GlobalFoundries announced it was quitting the leading edge entirely. One is a product launch. The other is the moment the list of companies that can build a state-of-the-art chip shrank to two, and TSMC is the one holding the customers.
7nm is a major density step over 10nm, and it is the first node where TSMC’s lead rests on its own process rather than on Intel’s stumble. It is printed still on 193nm immersion light with heroic multipatterning, no EUV in this version, and it delivers the kind of performance-per-watt jump that makes everyone want to move at once. The A12 is here, Huawei’s Kirin 980 is here, and the part that matters most for the next two years is not a phone chip at all.
AMD bet its entire comeback on this node. Its next-generation server and desktop parts are coming on TSMC 7nm next year, and the logic is brutal and simple: Intel is stuck on 14nm with a 10nm that will not ship, so a smaller, more efficient AMD chip built at TSMC can beat Intel on the thing Intel always won, raw manufacturing. For most of the last decade AMD was a company that could design well and not build well enough to matter. TSMC pulling ahead of Intel is what turns that around, and AMD is the first big customer to weaponize it against Intel directly instead of against another fabless rival.
GlobalFoundries quitting is the structural story. This is the company that showed up in 2009 as the second source TSMC’s customers wanted after the 40nm yield scare, the one that was supposed to keep the leading edge competitive. In August it looked at the cost of getting 7nm to yield, the EUV tools, the capex, the R&D, and decided the bill was not worth chasing a market two players already owned. So it stopped. AMD, its biggest leading-edge customer, walked straight to TSMC. The pure-play bleeding edge is now TSMC and Samsung and nobody else, TSMC booking close to sixty percent of all foundry revenue and Samsung a distant second, while Intel, a foundry for no one but itself, cannot ship the node it would need to even be in the conversation.
This is the cost spiral finishing its work. Every node since 20nm got more expensive to build than the density gain paid back, and the club with the rising cover charge just lost a member who could not afford the door anymore. A leading-edge fab now costs more than most countries’ science budgets, a 7nm mask set runs into the tens of millions, and only a handful of products on earth, the highest-volume phone and datacenter chips, ship in enough quantity to amortize a tape-out here. The leading edge kept getting lonelier as it advanced, and 7nm is where the room emptied out to two.
The version of 7nm shipping right now is the last gasp of the old light, 193nm immersion pushed past every reasonable limit with quadruple patterning on the tightest layers. Next year’s refinement brings EUV into volume for the first time, a single 13.5nm exposure replacing four masks on the layers that need it most. Get it working at volume and the cost curve gets a brief reprieve while the patterning complexity finally drops. Fall short and TSMC is left with a very expensive problem and a roadmap built around a tool the industry spent twenty years and a fortune to make barely functional. The bet is placed, and next year says whether it pays.
Apple ships on 7nm first, AMD bets its survival on it, and the one company that used to set the pace cannot get onto the node it needs. That is what 7nm changed: TSMC is no longer the foundry playing catch-up, it is the one everyone else gets measured against. The uncomfortable part, the part nobody at TSMC says out loud, is that being the only place on earth that can build the most important chips in the world is not just market power, it is a target painted on one set of fabs on one island. The industry just consolidated its most advanced manufacturing into a single company’s hands. That is a triumph and a single point of failure, and right now only one of those words is in the press release.