TSMC’s A16 Node: Backside Power and the Angstrom Era

TSMC’s next node is called A16, the A standing for angstrom because the marketing ran out of usable nanometer integers, and its headline feature is the one thing 2nm did not have: backside power. A16 reaches risk production in the second half of this year, with the volume that actually fills customer products landing in 2027, and it is TSMC finally answering the architectural trick Intel beat it to, routing the power rails underneath the transistors instead of crowding them onto the same congested layers as the signal wiring. The transistor stays gate-all-around. What changes is the plumbing, and the plumbing turns out to matter more than another shrink.

On every node up to now, the wires carrying power and the wires carrying signals have shared the same stack of metal above the transistors, fighting each other for room while power threads down through a dozen layers just to reach the devices. Super Power Rail flips the wafer over and builds a separate power-delivery network underneath. The rails that feed the transistors get buried in the silicon early, wired straight to the source and drain. Then the finished wafer is ground down to a few microns of its back, thin enough that a fresh power network can be printed on the underside and tied up to those buried rails through vertical contacts. Power comes up from below, signals stay up top, each with the whole space to itself. The payoff is real, and none of it comes from shrinking anything: voltage droop drops once power stops clawing its way down through the signal layers, and density and efficiency climb as the routing finally breathes. That is the tell. The shrink-the-geometry playbook is done, and structural tricks are what is left, because the geometry has nowhere much left to go.

Intel got backside power to silicon first with 18A, and it has spent the better part of a year telling anyone who will listen that it beat TSMC to the most important architectural change of the decade. On the sequence of events, Intel is not wrong. The question A16 actually answers is whether being first to the feature beats being TSMC, and the early read lands where this contest always has since 16nm. Intel really did get there first on the feature, and on the things that decide who ships, yield and a customer ecosystem, it sits behind where it has always sat. TSMC shows up second to the idea and, as usual, first to the version anyone can buy in volume.

To get at the back of the wafer you grind the finished thing down to a few microns, and silicon that thin bends and cracks, wrecking the lithography on both sides. That is the easy problem. The hard one is alignment: the power network printed on the back has to meet the transistors on the front within a few nanometers, on a wafer that has been flipped and ground down, and every slip is a missed contact or a dead short. Low-resistance connections through the back took new materials and new cleaning chemistry on top of that. Intel walked into the whole gauntlet carrying extra weight, because it moved to gate-all-around and backside power on the very same node, two first-times stacked on one process, and 18A’s long bring-up is the bill for that choice. TSMC is bolting backside power onto a nanosheet transistor it has already been yielding since N2, one hard new thing instead of two. That is the unglamorous mechanism under what keeps looking like destiny. Not magic, just a company that refuses to debug two miracles at once.

Backside power helps most where the current draw is highest. A modern AI accelerator pulls hundreds of amps, and on a frontside design all of that has to fight its way down through the same metal the signals are trying to use, bleeding voltage and dumping heat the whole way. That points it straight at the giant datacenter accelerators that already strain every part of TSMC’s supply chain, not at phones. So TSMC has split its roadmap down the middle: mobile takes the cheaper gate-all-around node, and the AI and HPC silicon that can actually pay for it takes A16 with the power network underneath. The leading edge used to be defined by what went in your pocket. Now it is defined by what trains a model in a building the size of a warehouse, and A16 is the first TSMC node built for that customer first and everyone else after.

Calling it 1.6nm would have been an even more obvious fiction than the last few node numbers, so the industry switched units to angstroms and reset the marketing odometer. That is roughly how much the number still describes anything physical. The wafer cost climbs again and the customer list shrinks to match, and the gap between what the leading edge can do and what all but a dozen companies can afford keeps widening. None of it is new. It is the same trend that has run for twenty years, cost going one way and the buyer list going the other, the simple shrink dying by inches the whole time, only now the integer wears a different letter.

TSMC is still the only place on earth that builds the most advanced chips in volume, the chokepoint every AI ambition routes through whether it likes the dependency or not. The Arizona and Japan fabs ramp, they ease the concentration, they do not end it, because the newest node always lands in Taiwan first and A16 is no exception. What has actually moved across twenty years is not TSMC’s lead, which has mostly just widened. It is the stakes. A foundry that read as a clever business model in 2006 is now infrastructure two superpowers are rearranging their economies to control, and A16 is the node where that has stopped being news and turned into weather.

On the merits A16 is exactly what you would expect. Late to the feature, first to the volume that actually ships. Priced, of course, for customers who train neural networks for a living. The node is not the interesting part. The interesting part is that a company that started out renting fab time to people who could not afford their own has ended up the single most contested piece of physical infrastructure on the planet, and it got there one quietly competent node at a time, beating a rival on the materials or the transistor shape or, this time, the plumbing, then out-yielding them once it caught up. The fin is gone, the easy shrink went with it, and the angstrom era opens on the exact arrangement the foundry has run since 65nm: everything advanced funnels through one company on one island, and the customers have nowhere else to go. The only thing that changed is the size of the bill, and the number of governments now reading it over your shoulder.