Ben Okonkwo: Eleanor, how was the conference — did anyone there actually say the quiet part out loud, or was it all still 'Moore's Law is evolving'?
Eleanor Crane: Mostly 'evolving,' which — I mean, that's the language that's been doing a lot of work for about a decade now. And it's what made me want to do this episode, honestly. Because there's a story underneath the euphemism.
Ben Okonkwo: Okay, name it.
Eleanor Crane: A tape-out at the 3nm node costs on the order of a hundred million dollars now. One design run. And the performance gain that buys you is linear — not the exponential return the whole industry was built around. The bargain inverted, and the industry just... kept using the same vocabulary.
Ben Okonkwo: Wait — linear? Not even a meaningful jump in throughput per dollar?
Eleanor Crane: That's what the cost curve shows at leading-edge nodes. And it matters because Moore's Law was never purely about transistor count — that's the version of the story that makes it sound like a physics puzzle. The actual deal, the one Gordon Moore was describing in that 1965 Electronics magazine article, drawn from data at Fairchild Semiconductor — the deal was that you got more performance and it cost roughly the same. Cheap, universal, automatic.
Ben Okonkwo: Right, and he was generous enough to revise it — 1975, every two years instead of every year — which is the formulation the Semiconductor Industry Association turned into an actual planning roadmap. So it was never just an observation sitting in a journal. It had institutional teeth.
Eleanor Crane: Institutional teeth — that's the phrase I was looking for. Because a roadmap that the whole supply chain schedules around is something closer to a self-fulfilling prophecy than a scientific law. You make it true by acting as if it's true.
Ben Okonkwo: So the question is whether that mechanism is still operating — or whether what we have now is something that borrows Moore's name but runs on completely different logic.
Eleanor Crane: And I genuinely don't know the answer going in, which is why I wanted to start with the hook rather than the thesis. Intel's 4004 in 1971 — first commercial single-chip microprocessor — to where we are now is a real trajectory. But the engine powering that trajectory may have changed under the hood without anyone officially announcing it.
Ben Okonkwo: Hm. And changed in ways that don't distribute the same way the original scaling did — that's the part I want to pull apart.
Eleanor Crane: So that's where we're starting — six decades of a deal that held, and then something quietly broke. What broke it, and what's the cost of not saying so plainly.
Ben Okonkwo: Before we get into what broke it — I want to make sure the mechanism is actually in people's heads, because I think it gets assumed. Imagine a thumbnail. On that thumbnail, you're pressing tiny switches — thousands of them, each one flipping on or off to process information. Every generation of chip, the switches got smaller. Half the size. So you could fit twice as many on the same thumbnail. More switches, more work, same cost. That's the whole trick. That's all Moore's Law is describing.
Eleanor Crane: Oh — that's it? That's genuinely it?
Ben Okonkwo: That's it. And the reason Gordon Moore could see it coming in 1965 is that he had a handful of data points — five, maybe six — from his time at Fairchild Semiconductor. He plotted them, saw a doubling pattern, and called it forward. Not from equations. From a graph with almost nothing on it.
Eleanor Crane: So it wasn't derived from physics. It was pattern recognition.
Ben Okonkwo: Empirical observation. Which is actually — I mean, that distinction matters enormously. A physics law holds whether or not anyone acts on it. An empirical observation about an industry can become true *because* the industry decides to act on it.
Eleanor Crane: And the Semiconductor Industry Association basically made that decision official — turned the observation into a schedule everyone had to show up to.
Ben Okonkwo: Exactly. And once you're on that schedule — the foundries, the equipment makers, the chip designers — you're all investing on the same two-year clock. Which means the doubling happens because you all funded it to happen. The roadmap created the reality it was predicting.
Eleanor Crane: Is there a moment where you can actually see that kick in? Like a starting gun?
Ben Okonkwo: Intel's 4004. 1971. First commercial single-chip CPU — it had around 2,300 transistors. That's the transistor count that Moore's Law would track from there to tens of billions today. And the 4004 existed because Intel committed to making it exist — on a chip small enough to be commercially viable. That's not physics delivering a gift. That's an industry deciding what it was going to build next.
Eleanor Crane: 2,300 transistors to tens of billions. I keep trying to find a frame for that and I can't.
Ben Okonkwo: Right, and the thing worth holding onto is — what made it possible wasn't magic, it was geometric shrinking. Technology scaling. You reduce the size of the switch — the gate length, the oxide thickness — generation after generation, and more switches fit. That's what the industry called a 'node.' Each node, the switches get smaller. It worked for five decades because the switches could keep shrinking.
Eleanor Crane: And what you're setting up — which I think is where this gets genuinely uncomfortable — is that the thing it was always a social contract, and the physics just happened to cooperate for a very long time.
Ben Okonkwo: Cooperated — until the mid-2000s, and then the cooperation stopped, and nobody sent a memo. Think about a chip engineer sitting in a TSMC fab around 2005. Transistor counts are still doubling, right on schedule. But the clock speeds have plateaued. The chip is not getting faster. And that engineer knows why — Dennard scaling is dying underneath them.
Eleanor Crane: Dennard — Robert Dennard, IBM. His principle was that when you shrink a transistor, its power density stays constant. So faster and smaller didn't mean hotter.
Ben Okonkwo: Exactly that. Power per unit area holds steady, which means you can clock the chip faster for free, basically. Once that breaks — and it broke in the mid-2000s — packing transistors more densely just raises the heat. The chip generates it faster than you can pull it out.
Eleanor Crane: Isn't that just a cooling problem, though? Better thermal engineering, better fans, liquid cooling?
Ben Okonkwo: That's the intuitive fix. But — no, wait, here's where it gets uncomfortable — quantum tunneling is happening at the same time. Gate barriers are now a handful of atoms thick. Literally a handful. And at that scale, electrons don't stop at the gate. They pass through it. Probabilistically. You can't engineer around probability.
Eleanor Crane: They just... leak through.
Ben Okonkwo: Leak through, waste power, and make the switching unreliable. That's a hard physical floor approaching around 1nm. And this is the part I want to be precise about — the leakage generates heat. The heat makes leakage worse. So the thermal ceiling and the tunneling problem are not two separate walls. They're compounding.
Eleanor Crane: So the engineer in that TSMC fab in 2005 isn't watching one thing fail. They're watching three failures that are — I mean, each one is making the others harder to survive.
Ben Okonkwo: And the physics papers flagged it. Early 2000s. The signal was there.
Eleanor Crane: But the industry kept building as if Dennard scaling held.
Ben Okonkwo: Because admitting it meant rethinking the entire architecture. The foundries, the tooling, the supply chain — all of it scheduled around shrinking. Stopping shrinking isn't a technical decision, it's an existential one.
Eleanor Crane: And that reckoning — what it costs, who absorbs it, whether architectural innovation actually reaches everyone the way cheap universal scaling did — that's the part we need to get into next.
Ben Okonkwo: And that reckoning has a dollar figure now. A tape-out at 3nm — just the design run, before you've manufactured a single chip — costs on the order of a hundred million dollars. And what you get for that is linear performance improvement. Not the exponential return the industry was built around. The cost curve didn't slow. It inverted.
Eleanor Crane: Wait — actively reversed, not just flattened?
Ben Okonkwo: At 2nm and below, cost per transistor is no longer falling. Which means the entire incentive structure — shrink to save money, save money to reach everyone — is gone. And that's why only three players can even operate at the frontier now. TSMC, Samsung, Intel. That's it. Three fabs for the whole world's leading-edge silicon.
Eleanor Crane: Three. With — I mean, the geopolitical weight sitting on those three locations alone.
Ben Okonkwo: Enormous. But set that aside for a second — the industry's actual response to the physics wall is what I want to name, because it's genuinely clever. You stop making the switch smaller, you just use more switches at once. Multi-core CPUs, GPU accelerators, chiplets. Parallelism and specialization instead of raw shrinking.
Eleanor Crane: And NVIDIA's the sharpest example of that?
Ben Okonkwo: NVIDIA's T4, L4 accelerators — massive throughput for matrix-heavy workloads. AI inference, image processing. But — and this is the thing I want to sit with — that's not a general speedup. If you're writing a general-purpose app, a weather simulation, a hospital scheduling tool, those chips do almost nothing for you. The gain is real but it's narrow.
Eleanor Crane: Which is exactly what I want to press on. Transistor scaling made everything faster for everyone, automatically. You shipped a new node, every developer on earth got a speedup they didn't have to ask for. Does architectural specialization do that?
Ben Okonkwo: Honestly? No. And I think that's the fracture that doesn't get named. Chiplet architecture — disaggregating a chip into separately manufactured dies, mixing process nodes in one package — that's genuinely elegant engineering. But it's engineering for customers who can afford custom silicon. Google designing a TPU. A defense contractor running a specific inference pipeline. The developer writing a general-purpose app is not in that room.
Eleanor Crane: So the memory wall compounds this? Because I'd been thinking of that as a separate problem.
Ben Okonkwo: It's actually — wait, no, it's the same problem wearing different clothes. You add cores, you add throughput, but the data has to move from memory to those cores. And memory bandwidth hasn't kept pace. High-bandwidth memory, cache hierarchy redesign — those help, but only for workloads that are architected to use them. Again, narrow beneficiaries.
Eleanor Crane: And approximate computing fits here too, doesn't it — trading precision for efficiency. Which sounds like a reasonable bargain until you ask what happens when the workload isn't AI inference but, say, a drug dosage calculation.
Ben Okonkwo: That's the unresolved question. Approximate computing works beautifully for AI — the model doesn't care if a weight is off by a rounding error. But safety-critical computing? General-purpose software? The reliability questions are genuinely open. Nobody's closed that debate.
Eleanor Crane: So the new paradigm delivers extraordinary performance — for specific actors, on specific workloads, if they can afford the custom silicon. And the universality that made the original deal democratizing just... quietly left the room.
Ben Okonkwo: And that's — I mean, that's the thing nobody announced. TSMC, Samsung, Intel, operating at nodes only they can afford, pivoting to architectural specialization for specific customers on specific workloads. No press release. The contract just... expired quietly.
Eleanor Crane: Which is what makes it a social contract story more than a physics story. Gordon Moore in 1965, five or six data points from Fairchild Semiconductor, calling sixty years of cheap universal performance gains forward — that promise was always implicit. Computing gets better, and it gets better for everyone, automatically. That's what's been replaced.
Ben Okonkwo: And the post-silicon paths — MoS₂, carbon nanotube transistors, 2D transition metal dichalcogenides — those are real. The physics is interesting. But the gap between a lab demonstration and wafer-scale manufacturing is wide enough that I'd call them a longer-term research bet, not a scheduled delivery. They don't close the allocation question.
Eleanor Crane: The allocation question.
Ben Okonkwo: If leading-edge capacity lives in three fabs — and it does, right now — then the next question isn't physics or even engineering. It's who decides how that capacity gets allocated. And I don't have a clean answer to that.
Eleanor Crane: Neither do I. And I think that's actually where we have to leave it — not because the question is too hard but because it's genuinely open. Someone at that conference said 'Moore's Law is evolving' and I winced a little. But sitting here now, I think the more honest version is: the law held long enough to build the world we have, and what comes next is a choice, not a continuation.
Ben Okonkwo: A choice with enormous stakes and very few people in the room making it.
Eleanor Crane: Well. That's the quiet part out loud, I suppose.
