Jordan Hale: Alex, I've been sitting with something all week and I need to say it out loud — I think we've been mispronouncing what Moore's Law even is.
Jordan Hale: So it gets called a law of physics, right, like it's in the same category as thermodynamics or something. But Gordon Moore, 1965, writing in Electronics Magazine while he's at Fairchild Semiconductor — what he actually observed was a cost curve. The number of components you could fit on a chip at minimum cost. Minimum cost. That's a business metric dressed up as physics destiny.
Alex Mercer: And to some extent the word 'law' is what made it destiny. Because once you call something a law, the industry starts treating it as a floor — like, you have to meet it or you've failed.
Jordan Hale: Which — wait, that's the self-fulfilling part, isn't it? Like, Moore makes an observation, it gets called a law, Intel gets built around sustaining it, and suddenly the whole semiconductor industry is organized around hitting a doubling rate that one guy measured empirically in 1965 and then, by the way, revised to every two years in 1975.
Alex Mercer: The 1975 revision is underrated. That's not him discovering something new about physics. That's him looking at the market and saying — actually, every two years is the pace that makes economic sense now. It's a market recalibration.
Jordan Hale: So the law was never fixed. It was always... negotiable with reality.
Alex Mercer: Right — but the industry stopped treating it as negotiable. It became a commitment. And now transistors are at three nanometers — barely wider than a strand of DNA — and the negotiation is basically over.
Jordan Hale: Three nanometers. That number just sits there and quietly ends the conversation about whether this is engineering or something else entirely.
Alex Mercer: And what we're actually trying to figure out is what comes after. Whether what's happening now is a continuation of Moore's idea or something genuinely different.
Jordan Hale: But here's what I keep getting stuck on — if it was always economics, why did engineers treat it like a law of nature? Like, they didn't just follow the trend, they built institutions around it, you know? Road maps, ten-year plans, entire careers.
Alex Mercer: Because the economics and the physics were moving in the same direction. That's the thing. Density paid off, so companies invested, so density kept improving, so it kept paying off. It was a self-reinforcing loop. Intel is the clearest example — the whole company was basically a bet that being fastest down that curve was the only competitive position worth having.
Jordan Hale: Wait, so Intel wasn't just following Moore's Law — it was, like, manufacturing it?
Alex Mercer: Basically. And here's the plain-language version of how that worked — think of a bakery. If you double the size of your oven, your cost per loaf drops. That's not a law of thermodynamics. That's just what's profitable to bake. Moore saw chipmakers in the same situation: more transistors per wafer meant your cost-per-transistor fell. The insight was economic, not physical.
Jordan Hale: The bakery thing — yeah, that actually lands. And you're saying the industry kept building bigger ovens because... it kept working?
Alex Mercer: Right — but only until it didn't. And that's the complication nobody wants to say plainly. At a certain point the cost curve inverted. Each new node — going from 5nm down toward 3nm — costs more per transistor to manufacture, not less. The bakery analogy breaks because the oven gets so precise you need a clean room the size of a stadium to run it.
Jordan Hale: Hold on. More expensive per transistor? I thought the whole point was it always got cheaper.
Alex Mercer: That was the deal — that's exactly what broke. The density kept going up, I mean IBM just announced 2-nanometer chips, so physically you can still shrink. But the cost savings that made shrinking worth doing? Those stopped tracking. You're spending more to get incrementally less return.
Jordan Hale: So the physics isn't what killed it. The bakery just... stopped being able to sell enough loaves to justify the oven.
Alex Mercer: That's it. And once you frame it that way, "Moore's Law is ending" sounds less like a physics crisis and more like a business model running out of road. Which changes the question entirely — it's not when does physics say stop, it's when does the return on investment say stop.
Jordan Hale: Okay but — and I want to push on this a little — if it was always a business model question, why did it take until, like, now for anyone to say it out loud? Gordon Moore revised his own observation in 1975 and nobody went, wait, maybe this expires?
Alex Mercer: To some extent, the incentives kept everyone quiet. If you're Intel and the trend is working, you don't call it a trend. You call it a law and you fund the roadmap. The moment you admit it's economic is the moment a competitor asks whether there's a cheaper path. Which, by the way, is exactly what eventually happened.
Jordan Hale: But the cheaper path ran straight into something you can't business-model your way out of, you know? Like — okay, so I've been trying to wrap my head around what actually happens when you go below 5nm, and it's... it's genuinely weird. The electrons don't wait to be switched. They just go through the barrier.
Alex Mercer: Quantum tunneling.
Jordan Hale: Quantum tunneling. Which — I mean, the name sounds like science fiction, but the consequence is just: your transistor leaks. You're trying to switch it off and current flows anyway. The gate barrier is thin enough that electrons ghost through it.
Alex Mercer: And the brutal part is it gets worse as you shrink. It's not a bug you fix at 3nm. Shrinking is what causes it.
Jordan Hale: So IBM announces 2nm chips in 2021 and everyone reports it like a victory lap — and it kind of is? Like, they did the thing. But wait, does that mean they built something that leaks by design?
Alex Mercer: 'We got to 2nm' and 'it was worth it' are genuinely different claims. IBM proved density is still physically possible. Whether the leakage, the yield costs, the power overhead — whether any of that math closes is a separate question.
Jordan Hale: And then there's the heat. Like, even if you tolerate the leakage current — the transistors are packed so densely that the heat per unit area just... conventional cooling can't touch it, right? It's not that the chip melts, it's that reliability starts failing.
Alex Mercer: Thermal wall on top of the tunneling problem. Two separate mechanisms hitting at roughly the same scale. But the one I think is actually more damaging — economically — is Dennard scaling collapsing. That one's been quiet for over a decade and it basically nobody talks about it.
Jordan Hale: Wait — explain Dennard. Because this is the part I didn't understand until recently.
Alex Mercer: So Dennard scaling was the rule that said: shrink the transistor, and its voltage and power draw shrink proportionally. Which meant every new node was basically a free lunch — more transistors, same total power. That's why chips kept getting faster without your laptop catching fire.
Jordan Hale: The power bill stayed flat even as you crammed more in. That's — no, that's a hidden subsidy. That's what made every generation feel like magic.
Alex Mercer: And when Dennard scaling broke down — more than a decade ago, not recently — that subsidy vanished. You add transistors now, you get heat. Not proportionally better performance per watt. Just heat. So tunneling makes control unreliable, thermal limits make density dangerous, and Dennard's collapse means the power efficiency you'd expect to compensate for both of those isn't there. That's not one wall — that's two walls and a missing floor.
Jordan Hale: Two walls and a missing floor. Which — and this is the part I think we're going to need to get into — is exactly why what Google did with TPUs and what NVIDIA did with inference hardware starts to look less like innovation and more like, I don't know, survival.
Alex Mercer: Survival is the right word, actually. Because if you can't get more from density, the only move left is to get more from shape — you redesign what the hardware is *for*. And Google's TPU is the cleanest example of that logic. They didn't build a faster general processor. They built one that is extremely fast at exactly one kind of math: the matrix multiplication that underlies every AI inference query. That's it.
Jordan Hale: Wait — one kind of math?
Alex Mercer: Basically. The performance-per-watt gains you get from a TPU over a general-purpose CPU on that specific workload are enormous. It's not a marginal improvement. And no amount of density scaling on a CPU gets you there, because a CPU is designed to do *everything* adequately. A TPU is designed to do one thing at a level a CPU can't touch.
Jordan Hale: And NVIDIA landed in the same place from a different direction, right? Like, the T4 and the L4 — those started as graphics chips and then the architecture just turned out to be, I don't know, accidentally perfect for AI inference?
Alex Mercer: I'd push back slightly on 'accidentally.' NVIDIA made very deliberate choices to target inference workloads. The T4, the L4 — those are purpose-shaped for that kind of parallelism. It's the same logic as the TPU: match the hardware architecture to the computational pattern, stop pretending one chip should handle everything.
Jordan Hale: Picture the researcher — like, someone in 2025 running a large language model on a server rack, and they look at what's actually doing the heavy lifting, and it's... not a CPU. It's TPUs, it's inference GPUs. The general-purpose processor is basically a traffic cop at that point.
Alex Mercer: That's exactly it. The CPU is orchestrating. The actual work is happening on domain-specific silicon.
Jordan Hale: Which is — I mean, that's a genuinely different world than what Moore was describing. He was imagining one chip that just keeps getting better. Not a whole ecosystem of chips that each do one thing.
Alex Mercer: And then chiplets take that further. Instead of one monolithic die, you build from multiple smaller specialized dies, package them together — sometimes in 3D stacks. Better yield, you can reuse modules, you're extracting more performance from your existing transistor budget without needing a new density node. It's a completely different cost logic.
Jordan Hale: And Intel — wait, Intel is doing this now? The company that was basically the physical embodiment of Moore's Law for fifty years?
Alex Mercer: That's the tell. Intel has reorganized around chiplet architectures and domain-specific designs. If Intel is doing it — Intel, the company Gordon Moore co-founded — that's not a hedge. That's a concession that the old model is structurally done.
Jordan Hale: So the question I can't shake is — is this durable? Like, is specialization actually the new engine, or is it... a very sophisticated way of buying time until something else shows up?
Alex Mercer: To some extent it's both, and I think that tension is real. It's durable as long as quantum effects keep constraining classical scaling — which, given where we are at 3nm, is probably a long time. But 'durable' doesn't mean 'free.' Every new specialized chip solves one problem and creates a different one: you now have a fragmented hardware landscape where software has to know what silicon it's running on. That cost gets paid somewhere.
Jordan Hale: And that cost — I can't get past it, you know? Like, the automatic part is what died. Not compute. Not chips. The *automatic* part. Software you wrote for general-purpose hardware just got faster, for free, every two years, because a new process node arrived. That deal is gone. MIT CSAIL looked at what fills that gap and basically... there's no consensus answer. Nothing replicates the universality of density scaling.
Alex Mercer: And the memory wall makes that worse. Because even where you do get compute gains — TPUs, inference GPUs — the bottleneck has quietly shifted. It's not transistor count anymore. For large language models especially, the binding constraint is memory access speed, not raw processing density. You can have the fastest silicon on earth and still be waiting on data.
Jordan Hale: Wait — so the problem moved while we were all still arguing about transistors?
Alex Mercer: Basically. The compute got fast enough that memory became the choke point. Which is its own kind of irony — Moore's Law succeeding hard enough that it exposed a different limitation entirely.
Jordan Hale: I mean — okay, I want to sit with the MIT CSAIL thing for a second, because it's not like they proposed a bad answer, it's that... no answer has stuck. Like, everyone agrees the gap is real. Nobody agrees what fills it. That's a genuinely uncomfortable place for an industry that ran on certainty for sixty years.
Alex Mercer: I'm not totally convinced the industry will find one thing that fills it. Gordon Moore had one lever — density — and it worked universally. What we have now is a drawer full of partial tools. That's not nothing. But it's not the same.
Jordan Hale: And if specialization really is the endgame — not a bridge, the actual endgame — then compute stops being this shared rising tide, right? The assumption underneath the last sixty years was: it gets cheaper for everyone. If the gains only land where there's a profitable enough use case to justify a custom chip... I don't know what that does to everything else.
Alex Mercer: That's the part I keep not being able to close. If Moore's Law was always economics, what does the industry optimize for when density stops being the variable? I don't have a clean answer to that.
Jordan Hale: Neither do I. Which might just be where this one lands — not with a verdict, but with that question still turning. Good one to sit with, though.
