Finn Brooks: Clara, hey — did you sleep okay? Because I did not, I went down this IBM rabbit hole at like midnight and now I have feelings.
Clara Bennett: The transistor thing, or — wait, which part got you?
Finn Brooks: Okay the number. A hundred billion transistors. On something the size of a fingernail. IBM just — they built that. A 0.7 nanometer prototype, two stacked layers, doubles the density of their 2021 design. And my brain just kept going: that's more than ten times the neurons in a human brain, sitting on a chip. That is absurd.
Clara Bennett: It is. And yet — the cost per transistor is going up, not down. That's what gets me.
Finn Brooks: Wait, up?
Clara Bennett: Up. And that matters because Gordon Moore's original 1965 observation — it wasn't just about density. The whole point was economic: more computing for less money, every year. Density climbing while cost per unit rises, that's... technically impressive and actually a broken promise at the same time. Which is sort of the strange shape of what we're looking at today.
Clara Bennett: Think of it like a parking garage. For sixty years, every time they built a new one, they fit twice as many cars — and each parking space cost less to construct. That's the deal. Density *and* cheaper. Now they're stacking floors to hit the density numbers, but each floor costs more to build. You've got the cars. You've lost the deal.
Finn Brooks: Wait — so the IBM prototype, the stacked layers thing, that's literally the garage adding a floor?
Clara Bennett: Exactly that. Gordon Moore in 1965, he wasn't describing a law. He was noticing a pattern in cost curves. An experience-curve relationship. Then in 1975 he revised it to every two years, and the industry just... treated it like a target. For six decades Intel, TSMC, everyone built roadmaps around it. It became self-fulfilling because they decided it would be.
Finn Brooks: Hold on. It was never actually a law?
Clara Bennett: Never. Named like one, treated like one — but it was always a prediction people chose to make come true. And that's, I mean, that's actually what makes this moment so disorienting. The density metric still climbs, so technically you can say Moore's Law is alive. But the economic promise — cheaper per transistor — that's already gone. Intel and TSMC are both pouring dramatically more capital into each new node. So it's simultaneously alive and dead depending on which promise you're measuring.
Finn Brooks: Okay that framing — alive *and* dead at the same time — that is actually the most useful thing I've heard because every article I read just picks one side and argues it.
Finn Brooks: Okay but wait — what actually stops you from just making transistors smaller forever? Like at some point is it just... hard manufacturing, or is there a real wall?
Clara Bennett: Actual physics. That's the wall. At sub-5 nanometers, electrons don't stay where you put them. Quantum tunneling — they probabilistically pass right through the energy barrier the gate is supposed to enforce. The transistor is switched off and current is still leaking through.
Finn Brooks: Ghost through the barrier.
Clara Bennett: Exactly. And that leakage generates heat — and now you're not just dealing with one transistor, you've got billions of them leaking simultaneously. The heat density becomes, I mean — there's no cooling solution that just absorbs it. You've hit a thermal wall that the shrinking itself created.
Finn Brooks: So the roadmaps going to 0.2 nanometers by 2046, those aren't — wait, are those even real? Like is that theoretically possible or is that just a spreadsheet someone made?
Clara Bennett: Theoretically possible, economically brutal. They'd need 2D materials — molybdenum disulfide instead of silicon, because silicon just stops functioning reliably at those dimensions. The physics isn't forbidden, but the cost and thermal penalty at every successive node get sharper. No company's roadmap overrides quantum mechanics.
Finn Brooks: So the industry response is — what, work around it? Samsung adopted Gate-All-Around at 3 nanometers, Intel's got 18A coming, wrapping the gate entirely around the channel to suppress the leakage — but even that's buying time, right? Not solving tunneling.
Clara Bennett: Buying time, yes. GAA improves electrostatic control, chiplets disaggregate the die so you're not cramming everything onto one shrinking piece of silicon — those are real architectural gains. But none of them eliminate the underlying constraint. That's the part that's genuinely unsettling.
Clara Bennett: The thing I'm still sitting with — and I don't have a clean answer — is that Moore's Law was one number. One metric, and the whole industry knew where to walk. Chiplets, silicon photonics, neuromorphic computing for AI workloads — those are all real. TSMC's COUPE platform is in mass production, Ayar Labs just closed a five-hundred-million-dollar Series E for their TeraPHY optical engine chiplets, Lightmatter's shipping the Passage L200 and L200X now. That's not vaporware. But none of it is one number. There's no shared north star anymore, and I mean — imec is literally still publicly debating whether Moore's Law is dead because the answer changes depending on which definition you're measuring.
Finn Brooks: And whoever gets to *define* the replacement metric — TSMC, the photonics startups, whoever cracks the software stack for chiplet architectures — that's not a technical question, that's a power question. Like, that's who builds the next sixty years.
Clara Bennett: Yeah.
Finn Brooks: Finn started this at midnight because of a hundred billion transistors on a fingernail. Now I'm losing sleep for completely different reasons.
