Apple announced a Mac this week that does not have an Intel chip in it, and the processor that replaced Intel is built by TSMC. The M1 is the first Apple-designed silicon for the Mac, an ARM chip on TSMC’s 5nm process, and it ends fifteen years of Intel inside the Mac in a single keynote. The A14 in the new iPhones already proved N5 works. The M1 is the node doing something bigger than a phone: taking a customer away from Intel that Intel thought it owned forever.
5nm is the first node built around EUV instead of dabbling in it. Where N7+ used the new light on a handful of layers as a trial run, N5 runs EUV across roughly a dozen of them, and the payoff is the biggest density jump in years, close to twice the logic density of 7nm. The trial worked, the tools held up in volume, and TSMC bet a whole generation on them at the very next chance. A year ago it was learning the technology. Now it is shipping the densest logic on earth inside an Apple laptop.
The M1 is the clearest demonstration yet of what TSMC’s lead actually buys its customers. Apple could design an ARM chip that beats Intel’s laptop parts on performance-per-watt because it had access to a process Intel cannot match and cannot buy, since Intel builds its own. For years the argument for keeping x86 in the Mac was performance. Once the manufacturing lead moved to TSMC and stayed there, that argument died, and the M1 is the funeral. Every Apple Silicon Mac from here is a Mac with a TSMC chip where an Intel one used to be.
The customer that proved EUV first is gone. Huawei, whose Kirin chip carried the first volume EUV silicon last year, was cut off from TSMC entirely this autumn, when the US export rules tightened to the point that a Taiwanese foundry using American tools could no longer legally build its chips. The political variable nobody could price a year ago just printed a number: one of TSMC’s largest advanced customers, erased from the order book by a rule written in Washington. The fab did nothing wrong and lost the business anyway, because it sits at the exact pressure point between the two largest economies on earth.
Intel, meanwhile, is publicly conceding the manufacturing crown it held for thirty years. It announced this summer that its own 7nm is delayed into 2022 at the earliest, and the once-unthinkable question, whether Intel should give up and have TSMC build its chips the way AMD already does, is now an open discussion among its own investors. The company that spent decades as the reason the leading edge lived in America cannot reach the node TSMC is shipping today. Some of the most strategically important silicon in the country is now most cheaply and quickly built on an island a hundred miles off the Chinese coast, and that sentence is keeping people in Washington awake.
Demand is the other story of 2020. The pandemic pushed the entire world onto screens at once, and the orders for the chips behind that, laptops and datacenters and the silicon for working and schooling from home, are running ahead of what even TSMC can build. N5 capacity is tight and getting tighter, and the strain is starting to surface in places nobody expected, like car factories that are about to discover what it means to be a low-priority customer at a sold-out fab. Samsung has its own 5nm running, but it is TSMC that Apple and AMD and Nvidia’s best parts line up for, and TSMC’s share has pushed past half the foundry market this year while Samsung sits in the high teens. The next node, 3nm, is a couple of years out and it is going to be the hardest one yet, because 5nm is close to the end of what the current FinFET transistor can give.
The M1 is the proof TSMC’s lead runs wide, wide enough to peel Apple off Intel after fifteen years. But the line that should stay with you is a few paragraphs up, the customer a government erased with a memo. The same fabs that make TSMC indispensable make it a lever, and Huawei is what it looks like when the lever gets pulled. Intel conceding 7nm this summer means there is no American backstop if the island ever goes quiet. TSMC built the best manufacturing the industry has ever had, and the reward is that its yield reports now get read in capitals that could not care less about transistors. Thirty years of becoming indispensable, and the prize turns out to be a target painted on your back.