Ryan Castillo: Jordan, tell me you saw the thing about IBM and 2-nanometer chips before we get into this.
Jordan Hale: I did — and it's the thing I want to come back to because it breaks my brain a little. But first, you know what we have to do. 1965. Gordon Moore.
Ryan Castillo: The founding document.
Jordan Hale: One paper. He's looking at how transistor counts on integrated circuits have been doubling rapidly, and he just... extrapolates forward. And I mean — that extrapolation becomes the compass for the entire semiconductor industry for sixty years. Investors, engineers, chipmakers, all of them navigating by this one observation.
Ryan Castillo: And the observation itself is empirical, not a law. Which is worth saying out loud because it changes what it means when things slow down.
Jordan Hale: Totally — but here's where I want to start, because this is the part that genuinely surprised me. When people say '18-month doubling,' they're not quoting Moore. Moore revised his interval to approximately every two years in 1975. The 18 months comes from David House — an Intel executive — and he was talking about overall chip performance, not transistor count.
Ryan Castillo: So the widely cited version conflates two different claims — transistor density versus performance. Those diverge. And once they diverge, the question of when Moore's Law 'ended' becomes definitionally unstable.
Jordan Hale: Wait, say more about that — you mean like, you can pick your answer based on which version you're measuring?
Ryan Castillo: That's exactly what's happening. MIT's Charles Leiserson says 2016, pointing to Intel's five-year gap between 14nm and 10nm nodes as the evidence. But that's a transistor-count argument. If you're using House's performance framing, the answer shifts.
Jordan Hale: So the industry spent six decades chasing a misquote — and the misquote is also why we can't agree on the ending. There's something almost poetic and deeply frustrating about that at the same time.
Ryan Castillo: Intel, co-founded by Moore, built its entire product roadmap around this. And the version they sold externally was closer to House than to Moore's actual 1975 revision. That's not a small error in attribution — that's sixty years of coordinated industry planning built on a slightly wrong premise.
Jordan Hale: And it still worked. That's the part I can't fully explain — a self-fulfilling prophecy that was already a little bit wrong, and it still drove real gains for generations.
Ryan Castillo: That self-fulfilling part is actually what makes the next step so jarring — because the thing that breaks it isn't a coordination failure or a misquote. Picture a wall. Ball doesn't have enough energy to climb over it — you'd expect it just to stop. But at atomic scales, the ball materializes on the other side anyway. That's quantum tunneling. Electrons do that through silicon barriers, even when a transistor is switched off.
Jordan Hale: Wait — even when it's *off*?
Ryan Castillo: Even when it's off. That's the structural problem. You can't engineer a wall thin enough to do the job and thick enough to stop the leakage, because at these scales physics won't let you have both.
Jordan Hale: Okay, so — and this is where IBM's 2021 announcement hits differently for me now — they demonstrated 2-nanometer chips, which is like... barely wider than a DNA strand, right, 2.5 nanometers? And the instinct is to read that as 'look how far we've gone.' But you're saying the story is actually the opposite?
Ryan Castillo: That's exactly it. IBM proved you can build there. What it actually showed is how close to the edge we already are — not how much runway is left.
Jordan Hale: That's — I mean, that reframes everything. It's like the announcement of the 2-nanometer chip is kind of a warning sign dressed up as a victory lap.
Ryan Castillo: And the consequence isn't just philosophical. Quantum tunneling at that scale causes current leakage, extra heat, chips that misbehave in ways that aren't a defect you can fix on the production line. It's structural. The physics doesn't care about the next process node.
Jordan Hale: With basically every other engineering wall I can think of, there was always a 'but what if we try this material' or 'but what if we redesign the geometry.' Like, FinFETs bought years. High-k dielectrics bought more years. But quantum tunneling... you can't negotiate with it? That's not a temporary problem waiting for a clever fix?
Ryan Castillo: That's the line. Every prior wall was an engineering problem. This is a physics constraint — quantum mechanics doesn't have an exception for clever manufacturing. You can slow the leakage, you can route around it architecturally, but you cannot make it go away by refining the process.
Jordan Hale: So when TSMC is actually running 3-nanometer nodes in production right now, they're not approaching the wall — they are at the wall, just... managing it.
Ryan Castillo: Managing it, yes. And that's a different category of problem than what Moore's original observation was tracking. His observation assumed you keep shrinking, physics cooperates, cost comes down. That assumption is gone.
Jordan Hale: And I think that's what hits hardest — it's not that engineers haven't solved it yet. It's that the answer is no. Which is just... a completely different sentence than anything this industry has had to say for sixty years.
Ryan Castillo: But 'managing it' is where I want to push, because I think we've been framing this as a physics story when there's an economic story running parallel — and honestly, I'm not sure which one hit first.
Jordan Hale: Wait — you mean the $20 billion fab thing?
Ryan Castillo: Exactly. Building a leading-edge fabrication facility now costs upward of $20 billion. EUV lithography machines alone are a massive capital outlay just to get in the door. That number concentrates advanced manufacturing among three players — Intel, TSMC, Samsung — regardless of what quantum tunneling is doing.
Jordan Hale: So like... even in a world where quantum tunneling wasn't a problem, the industry was already collapsing toward three companies just because of cost?
Ryan Castillo: That's the uncomfortable part the research doesn't cleanly resolve. The cost escalation is both a consequence of the physical scaling difficulty and an independent constraint. Those aren't the same thing, and which one is primary actually determines what solutions make sense.
Jordan Hale: I mean — and I keep turning this over — Dennard scaling broke in the mid-2000s, right? Robert Dennard at IBM showed that as transistors shrank, power density stayed flat. That held. And then it just... didn't. Smaller transistors stopped automatically delivering better power efficiency. That's a decade before quantum tunneling becomes the headline, and fabs are already getting more expensive trying to compensate.
Ryan Castillo: So Dennard breaks, costs start climbing to chase efficiency gains that aren't coming automatically anymore, and then quantum tunneling closes the door on the physics side. Two separate walls, arriving in sequence.
Jordan Hale: Which means nobody sent a memo.
Ryan Castillo: Literally nobody. Think about a data center procurement manager in 2018 — she's budgeting server refresh cycles assuming thirty percent performance gains per generation because that's what two decades of experience told her. At some point those gains quietly dropped to eight percent. Her spreadsheet is still running on Moore's Law and the world already changed.
Jordan Hale: That's actually — no, that's the human scale version of everything we've been saying. It's not an announcement. It's a gradual wrongness that nobody flagged.
Ryan Castillo: And the industry's response to that — chiplets, 3D stacking, gate-all-around transistors — we should get into that, because the sheer diversity of those bets tells you something troubling about whether there's actually a successor story here.
Jordan Hale: Yeah — and I don't think it's a hopeful diversity. That's what I want to figure out.
Ryan Castillo: And that's the tell — three different answers. Chiplets, 3D stacking, gate-all-around transistors. When Moore's Law was working, the whole industry pointed the same direction. Now Intel, TSMC, and Samsung are each placing different architectural bets. That's not ingenuity. That's the absence of a shared map.
Jordan Hale: Wait — say that again. Because I think that's the thing I couldn't quite articulate.
Ryan Castillo: If there were a successor to Moore's Law, everyone would be doing it. One thing. Instead you get chiplets — disaggregating a processor into modular tiles assembled via packaging, so you're not forcing every component to the same leading-edge node. You get 3D stacking — layering memory vertically instead of spreading laterally, shorter interconnects, higher density. And then gate-all-around transistors, which wrap the gate on all four sides of the channel instead of three like FinFETs — better electrostatic control, less leakage. Three genuinely different mechanisms.
Jordan Hale: And none of them are, like... the answer. They're each answers to a different piece of the problem.
Ryan Castillo: Right — but the part that doesn't fit is that we can't actually quantify whether all three together deliver equivalent performance-per-dollar gains to what Moore's Law delivered on its own. That data doesn't exist yet. Nobody's shown that chiplets plus stacking plus GAA closes the gap.
Jordan Hale: That's — I mean, that's almost the scariest sentence in this whole conversation. We've replaced one predictable curve with three partial answers and we don't know if they add up.
Ryan Castillo: And Leiserson's point lands harder here because of that. Intel took five years to go from 14nm to 10nm. Five years. Moore's cadence was two. That's not a hiccup — that's a structural stall that happened before any of these architectural pivots were mature enough to compensate.
Jordan Hale: So 2016 is Leiserson's line — that's when it actually ended — and the industry spent the years after it building workarounds instead of, I don't know, acknowledging the map was gone.
Ryan Castillo: Which is where IBM's 2-nanometer announcement in 2021 actually fits. Not as proof we're still climbing. As proof that the distance between a lab demonstration and real market adoption is now where Moore's Law dies. The economic decision to stop climbing — that's the wall, not just the physics.
Jordan Hale: No way — so IBM builds it, proves it works, and... nothing? No one's manufacturing at 2nm at scale?
Ryan Castillo: Not at that node. Because the cost calculus doesn't close. You need a $20 billion fab and EUV lithography just to attempt it in production. The physics cooperated in the lab. The economics didn't follow.
Jordan Hale: So chiplets, you know, look like the smart response to that — because you're not forcing everything to the bleeding edge. You mix nodes. You assemble. And it's clever, it genuinely is, but... it's not a story. It's a workaround dressed as a strategy.
Ryan Castillo: Chiplets and 3D stacking are engineering answers. They're not narrative answers. They don't tell the industry — or investors, or that procurement manager — what to expect next generation. Moore's Law did that. None of these do.
Jordan Hale: And without that shared story, every company is basically guessing — in a different direction — at the same time. That's not a race to the answer. That's just... fragmentation.
Ryan Castillo: And that fragmentation is what Krzanich was actually papering over when he called the deceleration 'a natural part of the history of Moore's Law.' That's a CEO choosing language. TSMC and Samsung were both still claiming they maintained Moore's Law pace into 2019 — but maintained by what definition? They had to redefine the metric to keep the headline.
Jordan Hale: That's — yeah. And that's the thing I've been sitting with. Because if you have to change what you're measuring to claim you're still winning... I mean, that's not a victory lap, that's a quiet admission dressed up in PR language.
Ryan Castillo: Performance-per-watt. System capability. Those aren't bad metrics — they're actually more useful in some ways. But shifting the metric to preserve the Moore's Law headline is its own signal. The original observation has hit its limit.
Jordan Hale: And 'More Than Moore' — you know, that phrase. It sounds expansive, like we grew past the thing. But I don't think it is. I think it's... actually, no, it's closer to an acknowledgment that the one predictable vector is gone, and what's left is everything else, which is a lot of things pointing in a lot of directions.
Ryan Castillo: No replacement story. That's the real problem. Moore's Law wasn't just a technical roadmap — it was a shared contract. Leiserson's 2016 line, the Intel 14nm-to-10nm five-year stall, IBM's 2-nanometer lab result sitting without market adoption — those aren't separate anecdotes. They're evidence the contract already expired.
Jordan Hale: And nobody's issued a new one. That's where I land. Not that the industry stops improving — it won't — but the compass is gone and there isn't another one yet.
Ryan Castillo: Which is an uncomfortable place to stop. But it's the honest one.
Jordan Hale: Good conversation. Genuinely. This one took a while to shake loose.
