Intel 20A was supposed to be the node where Intel got back in front, and this month Intel cancelled it. Not delayed, cancelled, scrapped as a product node, the resources redirected to the node after it. 20A was carrying two genuine firsts, RibbonFET, Intel’s version of the gate-all-around transistor, and PowerVia, a way of delivering power from the back of the wafer that nobody has shipped, and it was meant to be the moment the comeback stopped being about catching up. Instead Intel killed it, and the chip it was supposed to build, Arrow Lake, is being built at TSMC.
RibbonFET is Intel’s gate-all-around transistor, the same generational jump Samsung reached first in 2022 and TSMC is still racing toward, the gate wrapping the channel on all four sides instead of three. PowerVia is the bolder half of the bet: it moves the power wiring to the back of the wafer, off the front where it fights the signal lines for room, which clears the congestion and lets the transistors run harder. Backside power is a real Intel-first, something even TSMC has not put into production. Both were supposed to debut on 20A. Neither is gone, they just move up to 18A, which is the only reason the cancellation is survivable at all.
Officially, 20A had served its purpose as a learning vehicle, and skipping it frees the engineers and the cash to get 18A out the door faster. That is not nonsense. 20A and 18A are close cousins, and proving the hard new physics once instead of twice is defensible on the merits. What is harder to wave off is the timing. Intel is bleeding cash this year, posting a multibillion-dollar quarterly loss over the summer and announcing layoffs of around fifteen percent of its workforce, the whole turnaround under brutal financial pressure, and cancelling a node you have already largely paid to develop is also a quiet way to stop the bills on it. Calling that focus is generous. It is a company trimming a plan it can no longer fund, and dressing the cut as strategy.
Arrow Lake is what makes this humiliating. Intel’s flagship desktop part for the year, the exact chip 20A was meant to manufacture, is instead taping out on TSMC’s process, because TSMC’s node yields reliably right now and Intel’s own leading edge does not. Sit with that for a second. The company going door to door asking the industry to build its chips in Intel fabs is quietly building its own halo processor in a competitor’s fab, because the competitor’s is simply better this year. IDM 2.0 always left room to use TSMC when Intel’s process was behind, fine. Using it for the flagship, in the same twelve months you are pitching customers on Intel Foundry, is not a line that survives contact with a sales meeting.
Three of the five nodes landed, and the foundry-grade Intel 3 arrived on time this summer, which is the part of Gelsinger’s scorecard that actually held. But 20A was one of the two leadership nodes the entire sprint pointed at, and it is gone, folded into 18A. So 18A now carries the whole burden of proving Intel can still lead, alone, with nothing held in reserve to catch a miss. Best case, this consolidation looks disciplined in hindsight and nobody remembers 20A existed. The case that should worry Intel is the other one, where 18A slips the way 20A just did, the five-in-four story is exposed as more roadmap than the company could afford, and there is no third leadership node left to pretend otherwise.
There is a defensible version of this: a disciplined company refusing to pay twice to prove the same physics, putting every dollar behind the node that actually matters. Perso, I half-believe it, and the work in RibbonFET and PowerVia is the most interesting thing Intel has done in a decade. What does not fit that tidy story is Arrow Lake rolling off a fab in Taiwan, the plain tell that Intel’s own leading edge still is not ready when its own products need it. 20A was meant to be the node where Intel stopped needing TSMC. It became the node Intel cancelled on its way to building its flagship at TSMC anyway. Everything rides on 18A now. There is nothing staged behind it.