Cyrus Reed: Okay, week's been weird — I've been reading about transistors at three in the morning, which is either dedication or a cry for help, and I genuinely cannot tell which.
Iris Holm: Probably both. What pulled you in at three AM?
Cyrus Reed: This stat — Intel deployed eighteen times more researchers in 2024 than they had working on this in the seventies. Eighteen times the people. And the cadence is still slipping. How does that happen?
Iris Holm: That's — hold on, eighteen times more people and the output rate is worse?
Cyrus Reed: That's the knot we're pulling on today. Moore's Law — whether the thing that organized sixty years of the tech industry is actually gone, or whether we just... quietly redefined it so nobody had to say the word 'dead.' But first — do you want the thirty-second version of what it actually is? Because I think the plain-language version is genuinely underrated.
Iris Holm: Yes. One sentence.
Cyrus Reed: Transistors are tiny on/off switches. Moore's Law — Gordon Moore, 1965 — said those switches would halve in size every two years, so you'd get twice the computing power at the same or lower price, basically forever. That's it. That's the whole promise.
Iris Holm: And Moore wrote this for a trade publication, not a lab. Intel turned it into a planning document.
Cyrus Reed: A sixty-year planning document — for the whole industry, not just Intel. Every product cycle, every investment round, all of it was scheduled around that doubling. So when you need eighteen times the researchers just to maintain the cadence... something has fundamentally changed about the physics, or the economics, or — yeah. That's what we're here to figure out.
Iris Holm: And here's where the physics stops cooperating. The whole mechanism — CMOS scaling, shrinking the MOSFET — worked because making the switch smaller made it faster and cheaper. That deal is off now.
Cyrus Reed: Wait — the MOSFET specifically? Like, what breaks about it?
Iris Holm: The gate oxide. That's the insulating layer that keeps the transistor's on-off logic clean. Below about one to three nanometers thick, electrons don't wait for permission. They tunnel through.
Cyrus Reed: Tunnel — as in, quantum tunneling? They just... pass through a wall that should stop them?
Iris Holm: Through a wall that classically cannot be crossed, yes. And this is the part that kills the 'better factory' argument — it's not a defect. You cannot manufacture your way out of quantum mechanics.
Cyrus Reed: Okay but — so the transistor is supposed to be off, and electrons are just... leaking through anyway. What does that actually do to the chip?
Iris Holm: Leakage current. The switch is 'off' and current is still flowing. That wastes power. And wasted power is heat. And heat, packed into a smaller area — that's the second wall.
Cyrus Reed: Wait, so these are two separate problems hitting at once? Not the same problem with two names?
Iris Holm: Independent mechanisms. One is quantum, one is thermal. FinFET — the 3D transistor design — actually bought years by getting around the planar limit. But FinFET itself runs out of road below three nanometers. The ~3 nm node is where both constraints become dominant simultaneously.
Cyrus Reed: Huh — so imagine a chip engineer, right, she's staring at a new FinFET layout at, I don't know, two in the afternoon, and she's not hitting a lithography problem — her tools are fine — she's hitting, what, a heat map that looks like the chip is on fire before it even runs?
Iris Holm: Exactly that. The power density per unit area crosses a threshold that conventional cooling — fans, heat spreaders — cannot manage. It's not that the tools failed. The physics of concentrating that much switching into that small a space generates heat faster than any cooling architecture can remove it.
Cyrus Reed: So we've actually hit two independent walls — quantum tunneling kills your logic, thermal density kills your power budget — and neither one cares how precise your fabrication is.
Iris Holm: Neither responds to engineering effort. That's the actual finding.
Cyrus Reed: But that breaks me about it — okay, so the physics is brutal, fine, but the economics used to be the *escape hatch*. Like, Moore's Law wasn't just a physics claim, it was an economic promise. More transistors, lower cost per transistor. That deal — wait, that deal is actually gone now too, right? And that's the part that nobody's saying out loud.
Iris Holm: It's gone. Inverted, actually.
Cyrus Reed: Inverted how?
Iris Holm: The original law coupled transistor count to flat or falling cost per transistor. That coupling is broken. Building a leading-edge fab now costs tens of billions of dollars. And the cost per transistor at the frontier is no longer falling. It's rising.
Cyrus Reed: No way. So — okay, think about what that means for someone actually in the industry. Say it's 2026, a hardware engineer at some mid-size AI startup, she pulls up a TSMC quote for a custom chip at the newest node, and the per-unit cost is *higher* than the previous generation. Not a little higher — higher. And then she scrolls to the minimum order volume and it's... only achievable if you're a hyperscaler. Like, Google or no one.
Iris Holm: That's the real gate now. Not lithography. Capital.
Cyrus Reed: And TSMC sits at the center of all of that — like, geographically, economically, all of it funnels through one foundry.
Iris Holm: Which Pat Gelsinger basically confirmed when he admitted doubling has slowed to every three years. That's Intel's former CEO saying out loud: the two-year cadence Gordon Moore clocked in 1965 is gone. Costs tripled, timescale stretched by fifty percent, and TSMC is now the gatekeeper.
Cyrus Reed: Gelsinger said that publicly?
Iris Holm: Publicly. Which — frankly, that's the tell. CEOs don't walk up to microphones and eulogize their industry's core premise unless they have no other move.
Cyrus Reed: But wait — the law wasn't just describing physics, it was *organizing* the industry. Robert Noyce invents the integrated circuit at Fairchild Semiconductor, Moore makes his observation four years later, and then the whole sector just... locks in. R&D cycles, talent pipelines, supply chains, all of it synchronized to that doubling clock. It became self-fulfilling. The prediction made itself true for sixty years because everyone *built toward it*.
Iris Holm: Moore's Law as coordination mechanism. Yes. And now the mechanism has no floor to coordinate around — which is why I want to push on the architectural pivot everyone's making, because there's a version of this that's rationalization and a version that's genuine, and IBM's hundred-billion-transistor chip is where that argument actually lives. We'll get there.
Cyrus Reed: But before that — the name itself. If the economic promise is broken, the physics is broken, the timescale is broken — at what point is calling this 'Moore's Law' just... a comfort blanket?
Iris Holm: That's actually the right place to start — because IBM's hundred-billion-transistor chip is the answer the industry is giving to that question, and it's genuinely strange.
Cyrus Reed: Wait — hundred billion? On a fingernail-sized area?
Iris Holm: Fingernail-sized. Roughly twice the density of IBM's 2021 chip. And they didn't shrink the transistors to get there — they stacked them. Vertically.
Cyrus Reed: Wait, no, say that more carefully. They added a second floor. To the chip.
Iris Holm: Multiple floors. 3D stacking. You stop fighting the horizontal limit and you go vertical. Transistor count doubles without the gate oxide getting thinner, without quantum tunneling, without the thermal density per layer exploding.
Cyrus Reed: But here's the part I — okay, I did not expect this — that density, hundred billion transistors on a fingernail, that is the same scale as the nucleus of a biological cell. Like, we are packing silicon the way evolution packed organelles. That's... that's not a metaphor, that's actually the comparison.
Iris Holm: That's a real number?
Cyrus Reed: Real number. And it reframes what IBM actually demonstrated — it's not just an engineering milestone, it's... I mean, we stumbled into a packing density that biology arrived at through billions of years of selection pressure. But — okay — does stacking actually continue Moore's Law or does it just continue the transistor count while abandoning everything else the law promised?
Iris Holm: That's the genuine tension. And the node naming collapse makes it worse. TSMC announces a 2nm node. Are the gate lengths actually 2 nanometers?
Cyrus Reed: No. I know they're not.
Iris Holm: Marketing labels. The physical gate lengths stopped matching the node names years ago. So we're debating whether Moore's Law is alive using units that no longer mean what Moore measured. That's — frankly, that's semantic theater. The real pivot is chiplets. Heterogeneous integration. You fabricate a CPU core at one node, a memory block somewhere else, an AI accelerator at a third facility, and you assemble them into one package. Different manufacturers, different geometries, one device.
Cyrus Reed: So — like, picture a chip designer at NVIDIA, late 2028, she's not designing a monolithic die anymore. She's speccing out a package. Picking which foundry cuts which block. TSMC for the compute tiles, maybe someone else for the I/O. It's — wait, that's closer to systems integration than traditional chip design.
Iris Holm: And her accelerator is domain-specific — optimized for AI inference, not general computation. That's the third pillar. You stop trying to do everything faster and you do one narrow workload best-in-class. You get performance gains without needing more transistors overall. But here's what I want to name before we wrap this thread — there's a 2022 arXiv paper that actually proposes a quantitative modification to Moore's Law, reformulating it around performance-per-watt trajectories rather than raw transistor count. Which tells you the conceptual dispute is live. It's not settled. The people closest to the data are still arguing about whether the law needs a new definition or a death certificate.
Cyrus Reed: So the answer to my comfort-blanket question is — maybe the blanket still works, but only if you're willing to admit it's a completely different blanket.
Iris Holm: The blanket problem is actually the real problem. If you redefine Moore's Law as 'continuous performance-per-watt improvement through any means necessary' — stacking, chiplets, specialization, whatever works — you've made it unfalsifiable. And an unfalsifiable law cannot do what the original did for sixty years.
Cyrus Reed: Wait — unfalsifiable meaning what, exactly? Like, you can never prove it wrong, so... what breaks?
Iris Holm: Industry planning. Gordon Moore's original formulation was specific — transistor count, two years, flat or falling cost per transistor. That specificity is what let foundries, investors, and chip designers all coordinate around the same schedule. You lose the specific claim, you lose the coordination mechanism.
Cyrus Reed: Oh — and then the disruption isn't gradual, it's simultaneous. Because every piece of the ecosystem — investment cycles, foundry economics, the design tool assumptions baked into EDA software — all of it was synchronized to the same clock. So when the clock goes vague, they all go wrong at once. That's not a smooth handoff, that's a... I mean, that's a coordinated unraveling.
Iris Holm: And then there's the deeper question. The one that doesn't have a physics answer.
Cyrus Reed: The democratization thing. Yeah. Moore's Law — the real one, the original one — is what made computing cheap enough for everyone. Every two years, the same dollar bought more. That's not just an industry story, that's why a phone in someone's pocket today outperforms a 1990s supercomputer. But the cost floor for being at the frontier now — tens of billions to build a leading-edge fab, minimum orders only hyperscalers can hit — that's reversing. Capability is concentrating in, what, four or five foundries that can absorb the transition cost. That's not a physics question, that's... actually, no, it's a geopolitical question. And it's the one nobody named in 1965.
Iris Holm: Cyrus at three AM reading transistor papers. Eighteen times more researchers. And the thing that actually kept me up isn't the quantum tunneling — it's that the law's slowdown might cost more people access to computing than any physics wall ever could.
Cyrus Reed: That's where it lands, isn't it. Not dead, not alive — just no longer free. Thanks for pulling this apart with me.
