Cyrus Reed: Iris, I've been sitting with something uncomfortable since yesterday and I think you're going to find it either obvious or infuriating.
Iris Holm: Those are my two favorite outcomes. Go.
Cyrus Reed: Gordon Moore writes this paper — April 1965, he's at Fairchild Semiconductor, it runs in Electronics magazine — and he projects that transistors on a chip will double every year. Every year. That's the original call. And then ten years later, he revises it to every two years. Halves his own prediction.
Iris Holm: Right. And the 1975 version becomes the canonical benchmark. The one everyone memorized.
Cyrus Reed: So here's what's been bothering me — Intel's former CEO, when the current slowdown gets press attention, cites that 1975 revision as precedent. As reassurance. The message being: 'we've already done this, Moore adjusted, the industry adapted.' But wait — that revision didn't prove a pattern of successful adaptation. It proved the original 1965 prediction was wrong. Those are completely different things.
Iris Holm: No way he framed it as precedent.
Cyrus Reed: Framed it as precedent. And I keep thinking — if invoking your founder's error becomes the justification for the current slowdown, then what even is the law anymore? Because Moore's Law was never a law of physics, it was an observation about what engineers could accomplish economically. A self-fulfilling roadmap. It can slow down without nature having anything to say about it.
Iris Holm: So the question is whether this slowdown is inside the same logic — a correction the industry absorbs and routes around — or whether it's a different kind of limit entirely.
Cyrus Reed: And whether — huh — we'd even be able to tell the difference from the outside, because the data might look identical either way.
Iris Holm: That's the uncomfortable part. When does a revised prophecy become a different prophecy altogether.
Cyrus Reed: That's exactly where I want to start.
Iris Holm: But that's not actually the whole story — because calling it a prophecy still undersells the mechanism. It's more like a shared deadline. Think about it: every company in the industry — Intel, TSMC, Samsung — was reading the same benchmark, and every one of them funded toward hitting it. So the prediction came true partly because everyone agreed it would, and then spent money to make sure it did.
Cyrus Reed: Wait — so it's less 'we discovered a law' and more 'we all committed to the same deadline and collectively made it happen'?
Iris Holm: Exactly that. It's a coordinated bet. Not a discovery.
Cyrus Reed: Okay but — wait, no — that's actually kind of wild when you make it concrete. Because there's a chip architect sitting down on a Tuesday morning, coffee still cooling, and she's reviewing her team's five-year roadmap. And those targets — the node sizes, the transistor counts — they don't come from a physics equation. They trace back to a magazine article Gordon Moore wrote in 1965. That's the — huh — that's the actual floor under her spreadsheet.
Iris Holm: And she probably doesn't think about that. She thinks she's engineering toward capability. She's actually engineering toward a calendar.
Cyrus Reed: The calendar that Intel institutionalized, that TSMC and Samsung organized their entire node progression around — so the feedback loop just compounds.
Iris Holm: Right. Moore revised the article himself in 1975. Annual doubling to every two years. That revision is the one the industry canonized. But the mechanism — capital organized around a published target — that didn't change. The prophecy got edited and then re-believed.
Cyrus Reed: So we've already lived through one revision and treated it as continuity. Which means — wait, is that the same thing as what's happening now? Or is it different because the thing doing the constraining has changed?
Iris Holm: That's the fracture. The original constraint was economic — cost of fabs, tooling, R&D. You throw capital at it, you clear it. But the trend Moore was describing was never a law of nature. It can slow without nature caring at all.
Cyrus Reed: So the whole time, it was the industry's collective will wearing a physicist's clothing.
Iris Holm: That's the plain-language version, yes.
Cyrus Reed: But then — okay, if the constraint was always economic and the roadmap was always a negotiated bet, why does the current slowdown feel categorically different? Because the data on cost-per-transistor and performance-per-watt diverging — that could just be the next renegotiation, right? Or is something actually new in the room?
Iris Holm: Something new is in the room. And it's not a cost problem.
Cyrus Reed: Okay — meaning what exactly? Because that's the part I want to pull on. Like, what changed around 2004, 2006, in that window?
Iris Holm: Dennard Scaling. That's what changed. The companion principle — the one that said shrinking transistors also shrink their power draw proportionally. That broke down. Mid-2000s. Gone.
Cyrus Reed: Wait — so before that, you get more transistors and lower power basically for free?
Iris Holm: Free. That's the word. Moore's Law rode Dennard Scaling. They were coupled. And then Dennard collapsed, and suddenly more transistors means more heat. Which means you can't just keep pushing clock speed. Which is why clock speeds basically plateaued after 2005 while transistor counts kept climbing.
Cyrus Reed: So we're counting a metric — transistors — that got decoupled from the thing we actually wanted, which was faster and cheaper compute. But wait, no — that's not even the deepest layer, right? Because the thermal wall is still sort of an engineering problem. Like, better cooling, different materials—
Iris Holm: Until you hit quantum tunneling. That's the wall capital can't buy through.
Cyrus Reed: Okay, say more on that — because I think people hear 'quantum' and assume it just means really small and really hard. But it's actually — it's weirder than that.
Iris Holm: It's a lot weirder. At near-atomic gate dimensions, electrons don't stay where you put them. They probabilistically pass through the insulating barrier. The transistor is 'off' and current is still leaking. Not because the fabrication is sloppy. Because the electron's wavefunction says so.
Cyrus Reed: And TSMC can't — I mean, no amount of capital from TSMC or Intel fixes that. The physics doesn't have a budget line.
Iris Holm: That's the regime change. Every prior barrier — cost of fabs, lithography tools, R&D — those were engineering cost problems. You organize capital, you clear them. Quantum mechanics is not a cost problem. And the heat it generates compounds the thermal wall you can't cool your way out of either. Which — frankly — is what makes the current moment different from 1975.
Cyrus Reed: And that's before we even get into how chiplets and the transistor-count-versus-performance divergence are — wait, that's actually the next thing we need to open up, because the workaround story is stranger than it looks.
Iris Holm: The workaround story is stranger than it looks — and this is where the metrics question gets genuinely uncomfortable. Because transistor counts at TSMC, Samsung, Intel nodes keep climbing. That part is real. But cost-per-transistor has stabilized or inverted. Performance-per-watt has flatlined for three generations. Those aren't the same trend. They're different trends wearing the same name.
Cyrus Reed: Wait — so 'is Moore's Law alive' is actually — huh — that's not even a yes-or-no question anymore. It depends entirely on which number you're reading.
Iris Holm: Nick Zhang's 2022 paper makes this explicit. He argues the law 'started to falter' and proposes a quantitative modification. Not just 'it's slower.' The original formulation breaks as a predictor.
Cyrus Reed: Okay but — wait, there's a 2025 paper that goes even further on this, right? Alfonso Farina et al. — they fit the 1971-to-2025 transistor data to a golden-ratio model instead of an exponential. φ, 1.618. And it fits better than the original curve.
Iris Holm: Which means the exponential model systematically overestimates recent growth. Moore's Law wasn't just slowing — it was overclaiming the whole time and the data only revealed it now.
Cyrus Reed: That's — no way. So the law that the entire industry organized its capital around was the wrong shape of curve?
Iris Holm: Apparently. And chiplets are the industry's answer to that. But the mechanism — look, chiplet architecture partitions functionality across multiple smaller dies, assembled into one package. The system-level performance gains are real. The engineering lever is packaging and integration, not gate-length shrinkage. Those are categorically different.
Cyrus Reed: It's Moore's Law in the branding but not — wait, actually not in the physics at all.
Iris Holm: The chiplet review literature literally frames heterogeneous integration as a strategy to 'extend Moore's Law.' That phrase is doing a lot of work. It's extending the name. The mechanism underneath has already changed.
Cyrus Reed: So 'Moore's Law continues' and 'Moore's Law is dead' are both defensible at the same moment — because the law was never defined precisely enough to falsify. You just pick your metric and get your answer.
Iris Holm: That's the definition problem. And then AI pressure lands on top of all of it.
Cyrus Reed: Vishakha Agrawal's 2025 paper — this is the part that actually unsettled me — she documents AI compute demand doubling on timescales far shorter than Moore's two-year cycle. So even if classical scaling had continued perfectly, historically, it wouldn't have kept pace with what the models need.
Iris Holm: The demand curve outran the supply curve before the supply curve even bent.
Cyrus Reed: Which means — okay, so the pressure isn't just physics killing the old model from below. It's AI pulling compute demand from above faster than any version of Moore's Law ever moved. The ceiling and the floor are both closing, and chiplets are — I mean, they're buying time, not solving it. Right?
Iris Holm: Buying time is the right frame. And here's what I can't settle — if the industry pivots to quantum computing, neuromorphic, optical switching — none of those has replaced general-purpose CMOS silicon. Not one. They're candidate successors. That's a very different thing from successors.
Cyrus Reed: Right — but the part that doesn't fit is — wait, quantum computing isn't even going around the wall. The physics that breaks classical scaling, the tunneling, the quantum behavior at atomic scales — that's literally the physics quantum computing is built on. So are we routing around the barrier or are we just... deciding to build with it instead?
Iris Holm: That's a genuinely different question from anything we've asked today.
Cyrus Reed: And I don't — I mean, I don't have the answer. Because the thing is: if the 1965 prediction was wrong, and the 1975 revision became gospel anyway, and now the 1975 version is straining — each time, the industry called it continuity. Each time. So the question for the next decade isn't whether Moore's Law is alive. It's whether the semiconductor industry can converge on a new self-fulfilling prophecy that actually achieves the same roadmap consensus. Or whether chiplets, quantum approaches, AI acceleration — they just fragment into specialized architectures with no single law left to coordinate around.
Iris Holm: And without the coordinating function — without everyone reading the same benchmark and funding toward it — you lose the prophecy mechanism entirely. The self-fulfilling part requires a shared belief. Fragmentation breaks that.
Cyrus Reed: Which might mean the sixty-year run wasn't the law. It was the consensus. And the consensus is what's ending.
Iris Holm: I can sit with that. I can't tell you if it's right.
Cyrus Reed: Neither can I. Which is — honestly, that's a strange place to end up after all the specifics. Farina's golden-ratio curve, Zhang's modification, Agrawal's demand numbers — all that precision, and the thing we can't answer is whether there'll be a next story or just a lot of smaller ones. Anyway. Thanks for pulling on this with me.
