Lila Soto: Iris, good to be back — I have to tell you, someone in my building asked me this week what '2nm' means on the new iPhone chip, and I genuinely started explaining it and then just... stopped.
Iris Holm: Because you realized you couldn't.
Lila Soto: Because '2nm' doesn't mean two nanometers. It's a label. A marketing generation name. The actual gate dimensions on that chip are nowhere close to two nanometers. So what was I supposed to say?
Iris Holm: That's the episode. The node names — '2nm,' '1.4nm' from TSMC, Samsung — they've decoupled completely from physical measurement. Which raises the question: if the label is fiction, what's actually still true about Moore's Law?
Lila Soto: And I think we have to start with what Moore's Law actually was, because — okay, Gordon Moore published this in 1965, in Electronics Magazine. Not Nature, not a physics journal. A trade publication. He looked at the data on transistor counts and said, roughly, density is doubling about every year.
Iris Holm: He revised that to every two years in 1975. And then the whole industry — Intel, TSMC, Samsung — organized their multi-decade capital cycles around that revised number like it was a physical constant.
Lila Soto: Which it wasn't. It was a trend. An observation. And somewhere between then and now it became — a promise, almost. And I think the moment the node names stopped being measurements is the moment that promise quietly broke.
Iris Holm: The question is whether it broke because the physics stopped cooperating — or whether the physics just became the convenient explanation after the economics already stopped working.
Lila Soto: Oh, that's a good way to frame it. Because those are actually different stories.
Iris Holm: Very different. And the node-naming thing — '3nm' as a brand, not a dimension — is the first piece of evidence I'd put on the table.
Lila Soto: So if the label is theater, what does that tell us about what's actually continuing underneath it? That's what I want to figure out.
Iris Holm: The theater part is downstream of something real, though. Because there's an actual physical reason the labels had to become fiction — and it's not complicated. A transistor gate is basically a wall. It tells electrons: stop here, go no further. That's it. That's the whole job.
Lila Soto: A wall that controls traffic.
Iris Holm: Right. Now make the wall thinner and thinner — because that's what scaling meant for fifty years, shrinking the MOSFET, the Metal-Oxide-Semiconductor Field-Effect Transistor, every dimension including that wall. At some point the wall is only a few atoms thick. And electrons stop caring about it. They just — pass through.
Lila Soto: Like, literally through? Not around?
Iris Holm: Through. That's quantum tunneling. It's not a defect, it's not a manufacturing problem you can fix — it's what electrons do at that scale because quantum mechanics says they have a probability of being on the other side of a thin barrier whether or not classical physics permits it.
Lila Soto: The ghost-through-a-door thing is actually real physics, not just an analogy.
Iris Holm: It's exactly right. And the consequence — the thing the industry hit head-on in the late 1990s and into the early 2000s — is leakage current. The transistor is switched off. No signal, no command. And current is still flowing, because electrons are tunneling through the gate oxide anyway. You're burning power doing nothing.
Lila Soto: Hm. So it's not that the transistor breaks — it just... can't actually be off.
Iris Holm: Can't be fully off. And that becomes catastrophic at scale — I mean, a modern chip has billions of MOSFETs. If each one is leaking even a tiny current when it should be idle, the aggregate heat and wasted power is — yeah, that's why laptops throttle, that's why data centers became furnaces. The threshold where this gets serious is roughly three to five nanometers of gate oxide thickness.
Lila Soto: And we hit that threshold — when, exactly?
Iris Holm: Late nineties, early 2000s. That's when quantum tunneling became an industry-wide engineering crisis, not a theoretical curiosity. Which is the part that I think gets buried — this wasn't a surprise. The physics was known. The industry kept shrinking anyway and just... managed the symptoms.
Lila Soto: So the node labels drifted into fiction not because anyone decided to lie — but because the thing they were originally measuring had already become unmeasurable in the old terms. The wall got so thin it stopped being a wall.
Iris Holm: But here's what that clean tunneling story skips — the wall wasn't actually the first thing that broke.
Lila Soto: Oh — wait, what do you mean?
Iris Holm: Dennard Scaling. The companion principle. The reason shrinking worked so well for so long wasn't just density — it was that as transistors shrank, power density stayed flat. Smaller transistor, same heat per unit area. Speed went up, heat didn't. That was the actual bargain.
Lila Soto: So you got faster and cooler — or, not cooler, but not hotter — simultaneously.
Iris Holm: Right. And that collapsed in the mid-2000s. Before quantum tunneling was the headline crisis. The power density wall hit first.
Lila Soto: Hm. So — okay, I want to sit with that timeline for a second, because that's actually kind of wild. The thing we talk about as the end of Moore's Law, the tunneling story, the 3nm threshold — that came after? The industry was already in trouble on power grounds before the quantum stuff became the public narrative?
Iris Holm: That's exactly it. And the tell is multi-core. Why did Intel and everyone else suddenly ship dual-core, quad-core processors in the mid-2000s? Not because parallel computing was newly fashionable — because clock speeds couldn't keep rising. Dennard was gone. You couldn't just crank the frequency anymore without the chip becoming a furnace.
Lila Soto: So multi-core was — I mean, it was sold as a feature, but it was actually kind of a workaround.
Iris Holm: Structural workaround. And then the FinFET — the three-dimensional transistor architecture — that got invented specifically to keep planar scaling alive on power grounds at all. You go from a flat gate to a fin sticking up vertically, better control, less leakage. But the point is: the industry had to invent a fundamentally new transistor geometry just to keep going. That's not incremental. And then dielectric breakdown starts compounding everything — that's not tunneling, that's the gate oxide film catastrophically failing under electrical stress. Sudden, irreversible.
Lila Soto: Wait — catastrophically as in the transistor just dies?
Iris Holm: As in the insulating film stops insulating. Instantly. So you've got tunneling as the slow leak and dielectric breakdown as the potential blowout. Different mechanisms, both compounding at nanometer dimensions.
Lila Soto: And this is — okay, this is what makes the fintech scenario so concrete to me. A team in 2024, legacy single-threaded trading simulator, and they're used to the rhythm of just... waiting. New chip generation comes, they get a free twenty percent speedup, no rewrite needed. That free lunch is gone because single-threaded clock speed gains stopped when Dennard collapsed. Now they either rewrite for GPU compute — which is a capital expense, it's hiring people who know parallel architecture — or they just accept that their performance ceiling is done. That's the actual cost landing on an actual team.
Iris Holm: And the chiplet pivot, the 3D stacking — that's the industry's answer, but it trades one set of physical problems for another. We'll get to why that answer doesn't actually restore the broad economic bargain Moore's Law promised.
Lila Soto: And that answer — chiplets, stacking — I mean, it's real, but I keep wondering if we're calling it a solution when it's more like... a different problem.
Iris Holm: That's the exact framing. Chiplet architecture — you take a processor, instead of one monolithic die, you assemble it from multiple smaller dies connected by high-bandwidth interconnects. Intel does this, TSMC does this, Samsung does this. The yield argument is real: smaller dies have fewer defects.
Lila Soto: Hm. So it's cheaper to make several small things than one big perfect thing.
Iris Holm: And you can mix process nodes — put your high-performance compute die on the expensive cutting-edge node, your I/O on something older and cheaper. That's genuinely clever. 3D stacking goes further — you stack chips vertically, connect them through the die itself, so you're getting density gains without shrinking transistors at all.
Lila Soto: Which sounds like — okay, actually no, wait — that sounds like it just moves the heat problem somewhere worse. Everything's touching.
Iris Holm: Right. And the deeper issue is dielectric breakdown in the stacked structures. The insulating films between layers — under the electrical stress of dense vertical interconnects — fail. Catastrophically. So you traded quantum tunneling through a gate oxide for dielectric breakdown in your stack. Different physics, same class of problem.
Lila Soto: Oh wow. So it's not solved — it's just... relocated.
Iris Holm: Relocated. And here's where the GPU and TPU fit — because those aren't workarounds for general compute. They're purpose-built exits. A GPU is a massively parallel processor; it was never trying to be faster at single-threaded work. A TPU — Google's Tensor Processing Unit — is designed entirely for matrix math in AI models. Brilliant for that. Useless for your fintech trading simulator.
Lila Soto: Yeah, and the cultural thing is Moore's Law was a universal bargain. Cheaper, faster compute for everyone, every two years, automatically. GPUs and TPUs are not that. They're... narrow wins.
Iris Holm: Market segmentation. That's the actual pivot. The industry didn't find a new universal mechanism — it found several specialized ones and called it continuation.
Lila Soto: And the language game — '2nm,' '1.4nm' — is that the industry's way of not saying out loud that the universal promise ended?
Iris Holm: I think it's exactly that. The node names used to track a real physical quantity. Now they're generational brand labels — TSMC's '3nm' process, the gate dimensions aren't three nanometers. It's a way of keeping the forty-year narrative alive without having to announce a discontinuity.
Lila Soto: So the engineers who built careers on shrinking now have to think in chiplets, in stacking, in specialization — that's not just a technical retool, that's — I mean, the mental model of the whole industry was 'shrink it and the performance follows.' That assumption is just gone.
Iris Holm: Gone for general-purpose. Still alive — explosively — for AI accelerators. But that's the fracture. The question closes with what would have to be true for this to be Moore's Law continuing: you'd need cheaper performance broadly, for everyone. Chiplets and TPUs don't deliver that. They deliver cheaper performance for a specific class of workload.
Lila Soto: Which means software developers writing general-purpose code in 2024 are essentially on their own in a way they haven't been for forty years.
Iris Holm: On their own in a way that's actually hard to name cleanly. Because for forty years the economic coordination — Gordon Moore publishes one observation in a trade magazine, and Intel, TSMC, Samsung, the whole supply chain organizes around it — that coordination was doing enormous invisible work. Software teams didn't have to think about hardware. The hardware would just get better.
Lila Soto: And now — I mean, it's not that compute isn't getting better. It's getting wildly better for AI. GPUs, TPUs, the parallelism story is real. The demand for computation to train these models is exponential, and the accelerator world is racing to meet it. But that's a narrow track. One kind of work on one kind of hardware.
Iris Holm: Two scaling regimes now. That's the bifurcation. Specialized accelerators — density, power efficiency, everything — still moving fast. General-purpose compute? Plateauing. And the cost and accessibility curve that made Moore's Law matter broadly, that's the part that isn't being rebuilt.
Lila Soto: Yeah. And I keep sitting with whether anything actually can rebuild it. Like — the original coordination worked because shrinking a transistor benefited everyone on the same curve. Chiplets, 3D stacking, GPUs, TPUs — those are each somebody's answer to a specific problem. There's no single curve anymore that pulls everyone forward.
Iris Holm: That's where I land. Uneasy.
Lila Soto: Same. It's a genuinely open question — not the physics part, the physics is what it is. The question is whether the social and economic scaffolding that made semiconductor progress universal for fifty years has a second act. And I don't think we know.
Iris Holm: Worth spending a morning on. Thanks for doing that with me.
