Ben Okonkwo: Marcus — I'll be honest, I came into this one thinking I had a handle on it and then I hit a fact that just sat wrong with me.
Ben Okonkwo: IBM, 2021, 2-nanometer chip. Built it. Announced it. Demonstrated that transistors about the width of a DNA strand can actually function. And it has never — not once — shipped at profitable manufacturing scale.
Marcus Vale: And the industry just... moved on?
Ben Okonkwo: Moved on. Which, now, I think is the tell. Because we've been trained for sixty years — since Gordon Moore's 1965 observation — to treat 'transistor density doubles every two years, cost falls' as a single unified fact. Physics and economics in lockstep. IBM 2021 is the first time you can hold up one example and say: look, they moved independently. The physics is still going. The economics stopped following.
Marcus Vale: Two variables that used to track together — and the industry reported on them like they were one.
Ben Okonkwo: And here's the part that actually puzzles me more the longer I sit with it: Moore's Law wasn't a physics law. Moore noticed a pattern in 1965, Intel organized their entire R&D roadmap around hitting it, and the industry made it come true for over fifty years through sheer coordinated effort. It was a self-fulfilling prophecy, not a constraint of nature. So when it breaks, what exactly broke?
Marcus Vale: The economics of the prophecy. Not the physics.
Ben Okonkwo: Which is — okay, let me be precise about the mechanism, because this matters. At sub-3nm feature sizes you hit quantum tunneling: electrons pass through transistor gates that are physically too thin to contain them, generating leakage current and heat you can't remove. And heat dissipation at that density scale becomes exponentially more costly. So the physics didn't stop being possible. It became unprofitable to contain.
Marcus Vale: Two walls. Tunneling and heat. IBM built a chip past both of them — and couldn't sell it.
Ben Okonkwo: Which is the question the whole episode is really about. Not 'is Moore's Law dead' — that framing collapses the two problems back into one. The real question is: the physics wall and the economics wall arrived at the same address in 2021, and the industry has been treating them as one thing ever since Moore made his observation. What happens now that they've split?
Marcus Vale: But that split — physics and economics diverging — it didn't happen overnight. And I think most people miss when it actually started cracking. Because the apartment analogy works here. Imagine every time you renovated your apartment, it got bigger and cheaper and used less electricity. Simultaneously. Every two years. For fifty years. That's what transistor shrinking actually delivered.
Ben Okonkwo: That's — yeah. All three at once.
Marcus Vale: Smaller transistors switched faster, consumed less energy per operation, more of them fit into the same area. You got speed, efficiency, and density in one move. That's the free lunch.
Ben Okonkwo: And there was a named principle underneath that. Robert Dennard, 1974 — Dennard Scaling. The prediction was that as transistors shrank, power density stayed constant. Smaller transistors used proportionally less power, so the efficiency gains were basically automatic. That's what made the free lunch structurally guaranteed, not just lucky.
Marcus Vale: Until it wasn't.
Ben Okonkwo: Dennard Scaling broke down around 2005, 2006. Leakage current — electrons stop obeying the gate because the barrier is too thin — and suddenly the power density isn't staying constant anymore. The chips are running hot. The efficiency isn't automatic. That's the first crack.
Marcus Vale: So that's 2005. Which means the free lunch was over almost twenty years ago, and the industry kept reporting like it wasn't.
Ben Okonkwo: Kept hitting the transistor count target — the density number Moore described in 1965 — while the efficiency half of the bargain quietly decoupled. That's the thing that's hard to hold in your head. Moore's Law was always two promises bundled together. Density doubles. Cost falls. Dennard Scaling was what made both promises true simultaneously. Once Dennard broke, the bundle started coming apart.
Marcus Vale: And now below 3nm — this is the cost curve inversion — the bill for continued miniaturization exceeds what you get back. You're spending more to shrink than the shrink is worth. That's not a physics failure. That's an economics failure.
Ben Okonkwo: Which is actually — hm, wait — that reframes the IBM 2021 story. They proved you can still shrink physically. What they couldn't prove is that anyone would pay for it.
Marcus Vale: Exactly the proof point. Moore's Law was an economic promise dressed as a physics observation. Gordon Moore noticed a cost-and-density pattern in 1965. Intel organized decades of R&D around hitting that number. The moment the economics of hitting it inverted, the promise broke — not the physics.
Ben Okonkwo: The question is — below 3nm, what does 3nm even mean? Because if the label has already drifted from the physical reality, that's... the cost curve inversion is already hiding inside the marketing.
Marcus Vale: That's the next layer. Because the number and the node are not the same thing anymore.
Ben Okonkwo: Right — and the label drifting is actually where the physics gets hidden. TSMC's '3nm' node doesn't contain 3-nanometer features. That number describes a process generation, not a gate dimension. The actual gate lengths are closer to, I don't know, 12, 15 nanometers depending on what you're measuring. So when someone says 'we're pushing below 3nm,' the physical threshold you're worried about — the one where quantum tunneling kicks in — that wall may already be behind us.
Marcus Vale: Wait. So the label already crossed the wall.
Ben Okonkwo: The label crossed it. Whether the actual transistor gate did is — hm, it's genuinely murky. But here's what we do know about what happens when you get into that regime. Quantum tunneling isn't a leakage problem you can engineer around with better materials. At a few atoms thick, the gate barrier is physically too thin to contain an electron. The electron doesn't climb over it. It just... passes through. Probabilistically. And you get current you never asked for.
Marcus Vale: Current you can't bill for and can't remove.
Ben Okonkwo: Which generates heat. And at sub-3nm densities the heat dissipation cost — actually, let me restate that, because 'cost grows' undersells it. It grows exponentially in engineering effort. You're packing more transistors per square millimeter, each one leaking, and the thermal solution has to get correspondingly more extreme. The gains from miniaturization get eaten by the thermal management bill.
Marcus Vale: So the physics is actively working against you now. Not just slowing down — reversing.
Ben Okonkwo: That's the specific thing industry analysts mean when they say progress at the sub-3nm frontier is 'prohibitively expensive and slow.' It's not a speed problem. It's that the miniaturization is now producing its own headwinds. And there's a really concrete way to see it — imagine a chip designer at a fabless startup in 2024, speccing an inference chip. She runs the numbers on moving from 5nm to 3nm. Thirty percent more in manufacturing cost. Maybe twelve percent more performance. Because half the theoretical gain from the smaller node is eaten by leakage current and thermal management.
Marcus Vale: Thirty to twelve. That's — that's a negative ROI on the node shrink.
Ben Okonkwo: And she's not making a physics mistake. She's doing the accounting correctly. The physics has turned against the economics.
Marcus Vale: Which is the cost curve inversion landing in one real decision. One person's spreadsheet. And that's not a research finding — that's a purchase order that doesn't get placed.
Ben Okonkwo: And what she does instead — that's actually the whole next chapter. GPUs, chiplets, domain-specific accelerators — and the $700 billion capex number sitting underneath all of it, which I think is going to complicate whether any of that counts as a real solution or just a very expensive delay.
Marcus Vale: Whether it's exponential or a one-time step. Yeah. That's the question.
Ben Okonkwo: But the step to architectural solutions — that's where I keep wanting to slow down, because the framing I hear is 'GPUs and chiplets extended Moore's Law.' And I'm not sure that's right. What MIT CSAIL and the IEEE have both said is that architectural innovation is the structural *successor* — not a continuation. That distinction matters enormously.
Marcus Vale: Okay, break that down. What's the actual mechanism?
Ben Okonkwo: GPUs — instead of making one transistor switch faster, you run thousands of operations simultaneously. Massive parallelism. You get more computation without needing a smaller feature. It's not shrinking the transistor, it's changing the question entirely.
Marcus Vale: Right. And chiplets push that further — you disaggregate the monolithic die. Memory at one process node, logic at another, interconnected at high bandwidth. You're mixing and matching instead of hitting one perfect node.
Ben Okonkwo: And domain-specific accelerators go even further from general-purpose. A chip optimized purely for matrix multiplication — AI inference — doesn't need to be good at anything else. You trade generality for performance-per-watt within that narrow class. Which is actually — hm, that's the metric shift, isn't it. Performance per watt has basically replaced raw transistor count as the primary benchmark.
Marcus Vale: The metric changed because the improvement strategy changed. That's not nothing.
Ben Okonkwo: No, but — is any of this exponential? Or is it a step function?
Marcus Vale: That's — frankly, that's what the $700 billion answers. And not in a flattering way.
Ben Okonkwo: Walk me through that.
Marcus Vale: Hyperscalers are tracking toward more than $700 billion in capex in 2026. Roughly double the prior year's record. If chiplets and GPUs were delivering compounding exponential efficiency, you'd spend less over time — same output, cheaper. Instead they're spending double. That's not compounding efficiency. That's brute-force capacity — buying more of what already exists and stacking it.
Ben Okonkwo: So the spend is actually evidence of the plateau.
Marcus Vale: It's the tell. Jensen Huang calls Moore's Law dead. Intel and Samsung say it lives. And actually — both are defensible, which is the uncomfortable part. Depends entirely on which metric you pick and which workload class. Transistor density? Slowing. Performance per watt on a matrix multiplication accelerator? Still moving. Those are different curves.
Ben Okonkwo: Which means chiplets are — I mean, disaggregation is genuinely clever. But it's a one-time architectural rearrangement. Once you've exhausted the mix-and-match combinations, quantum tunneling and heat are still waiting at the bottom. You haven't moved that wall.
Marcus Vale: Step function. Not a new exponential. And the $700B is what a step function costs when you need to scale anyway.
Ben Okonkwo: So the real open question is whether this is a new engine — architectural innovation compounding forward — or a very expensive gear shift that eventually stalls at the same wall Gordon Moore's original observation was always pointing toward.
Marcus Vale: There's a proposed modification to Moore's Law — an actual quantitative one, covering derivative laws, trying to predict chip performance and energy consumption going forward. Someone sat down and said: the framework is updatable. Not dead. Updatable.
Ben Okonkwo: And I find that genuinely interesting as a move — because it's saying the 1965 observation is still the right container, we just need to... re-pour it. But the thing I can't settle is whether that's intellectually honest or whether it's the industry's way of not saying the word plateau out loud.
Marcus Vale: Could be both.
Ben Okonkwo: Right — but here's the consequence that keeps pulling me back. The distinction between 'the chip got smarter' and 'we bought twice as many chips' — it sounds like an engineering footnote. It's not. If you need a 2x performance gain and you get it from a smarter chip, your software runs faster on the same hardware, your energy bill stays flat, your competitive moat is durable. If you get that same 2x from buying twice the chips... you've written software that only runs at scale, your energy costs double, and whoever controls the orchestration layer above the chip — not Intel, not even necessarily NVIDIA, but whoever owns that software stack — that's where the moat moves. The downstream consequences are completely different.
Marcus Vale: And that's actually — wait, that loops back to where we started. IBM 2021. Built the 2nm chip. Proved the physics. Couldn't make it pay. What nobody asked at the time was: if it had shipped at scale, who would've captured the value? Probably not IBM.
Ben Okonkwo: No. And I think that's genuinely the unresolved thing. Gordon Moore noticed a pattern in 1965 and the whole industry organized around hitting it — and for fifty years the chip maker captured the efficiency gain. The question now, with $700 billion going into raw capacity instead of node-level efficiency, is whether that value capture ever comes back to the chip, or whether it already migrated upstream permanently. I don't think we know. I'm not sure the industry knows.
