Ben Okonkwo: Eleanor, long week — I spent half of it trying to explain to someone why Moore's Law isn't actually a law, and I think I failed.
Eleanor Crane: Well, you're in good company — Gordon Moore probably couldn't have defended it as a law either, given that he wrote it as an observation in a trade magazine in 1965, not a physics journal.
Ben Okonkwo: Electronics magazine. A chemist making a cost argument.
Eleanor Crane: A chemist — not a physicist — who had just co-founded Fairchild Semiconductor and was, in a fairly direct way, observing his own industry's economics. Transistor counts up, cost per component down. That's the founding text.
Ben Okonkwo: Which reframes the whole question of whether it's 'dead.' Because if it was always an economic trend, then it doesn't die when the physics gets hard — it dies when the cost curve inverts.
Eleanor Crane: And there's a very specific moment that I think names that inversion better than any research paper. TSMC declining to purchase ASML's newest chipmaking machine — not because the machine failed, but because the return on investment went negative. The technology works. The economics don't.
Ben Okonkwo: Hold on — so a working machine, at a working node, gets declined because building at that scale isn't profitable. That's not a physics failure.
Eleanor Crane: That's an accountant's decision. And I think that's the puzzle — how does a trend that shaped sixty years of computing infrastructure end with a line-item on a procurement spreadsheet?
Ben Okonkwo: Interesting — and part of the answer lives in the original formulation. Moore revised his own doubling estimate from every year to every two years back in 1975. So the 'law' we cite as bedrock was already a negotiated figure inside a decade of being named.
Eleanor Crane: Every two years — the figure everyone treats as physical constant, and it was already a concession.
Ben Okonkwo: Which is — I mean, that's the load-bearing thing, right? The law was always negotiable. The question is whether the economics renegotiated it one final time, or whether this is just the latest in a long series of apparently terminal limits.
Eleanor Crane: And Moore co-founded Intel after Fairchild — he watched this from the inside, every year. What it means that the thing he named became the thing the industry finally ran into — that's what I want to understand today.
Ben Okonkwo: But the 'physics got hard' story — that's where most people stop. And it's not wrong exactly, it's just... it's the symptom, not the cause. The actual engine underneath Moore's Law had a name, and most people have never heard it.
Eleanor Crane: Robert Dennard.
Ben Okonkwo: Robert Dennard, IBM researcher, 1970s. He codified what we now call Dennard scaling — the rules for how MOSFETs, the transistors on chips, actually behave as you shrink them. And the core claim is almost — I mean, it sounds too good. Shrink the transistor, and you get three things simultaneously: better performance, lower power per function, and lower cost per circuit. All three. At once.
Eleanor Crane: Which is — all three moving in the same direction, at the same time, every generation.
Ben Okonkwo: Every generation. Think of it like a photocopier — except when it shrinks the document, the copy also costs less to print and comes out faster than the original. That's what Dennard scaling handed the chip industry for thirty years. Not just physics. An automatic reward.
Eleanor Crane: And that automatic reward is what made Moore's Law self-reinforcing. Cheaper transistors funded the next round of shrinkage, which made transistors cheaper still.
Ben Okonkwo: Exactly — and Intel rode that loop for decades. The cost economics of each new process node basically financed the engineering work for the next one. It compounded.
Eleanor Crane: So the question isn't whether transistors can still physically shrink. It's whether shrinking them still pays for itself.
Ben Okonkwo: That's — yeah, that's the precise reframe. And imec, the Belgian semiconductor research institute, has been quite direct about this: 'smaller equals cheaper' broke down. Not eventually. It broke down as the operative economic logic of the industry.
Eleanor Crane: Which makes Dennard scaling the load-bearing wall, doesn't it? Not Moore's observation — Dennard's mechanism. Moore noticed the pattern. Dennard explained why it couldn't stop.
Ben Okonkwo: And when Dennard scaling stopped holding — actually stopped, not theoretically — the self-reinforcing cycle broke with it. You can still physically shrink transistors at some nodes. But you don't automatically get cheaper. You might get more expensive. The photocopier starts charging you more for smaller copies.
Eleanor Crane: And that's what a procurement spreadsheet looks like when a trend ends. Not a press release. A negative ROI on a machine that works.
Ben Okonkwo: But here's what the procurement story doesn't surface cleanly — the ROI going negative isn't just about price. Below roughly five nanometers, you hit quantum tunneling. Electrons passing through gate barriers they classically cannot cross. Not around them. Through them.
Eleanor Crane: Through the barrier itself.
Ben Okonkwo: Which means the transistor leaks current even when it's switched off. And that leakage — I mean, it's not a design flaw you can engineer around. It's quantum mechanics. The gate stops being a gate.
Eleanor Crane: And the heat dissipation problem feeds directly into this, doesn't it? Because that leakage current becomes heat, and heat was already a warning signal — the power wall people were noticing in the early nineties, even before quantum effects were the front-line concern.
Ben Okonkwo: Right — and that's actually where it gets uncomfortable. Because the power wall in the nineties, that was a physical signal. But the industry kept going anyway. Strained silicon, high-k metal gates, copper interconnects. They bought decades. So a listener might reasonably say — okay, they fixed it before, why not now?
Eleanor Crane: Why not now. And what does imec actually say to that?
Ben Okonkwo: imec is specific — and this matters. They don't say the physics is impossible. They say the breakdown of 'smaller equals cheaper' is the central reason conventional scaling has become less economically compelling. That's a different argument. It's not 'the tunnel is blocked.' It's 'the tunnel costs more to dig than what's on the other side.'
Eleanor Crane: Which connects directly to EUV. Because that's the tool that should have solved the lithography problem — and it does, technically. But the equipment is extraordinarily expensive, and then yield loss compounds it.
Ben Okonkwo: Yield loss is — okay, this is the thing that actually surprised me when I sat with it. A chip designer at a mid-sized fabless firm submits a tape-out for a three nanometer design. The yield report comes back. And the per-chip cost is higher than the seven nanometer generation it was supposed to replace. One defect ruins the whole die. Because at that node, the die is so large and complex that a single manufacturing flaw has nowhere to hide.
Eleanor Crane: Wait — higher cost than the previous generation. Not just 'not cheaper.' Actually more expensive.
Ben Okonkwo: That's what TSMC's declining the ASML machine is actually about. Not that EUV doesn't work. It's that you cannot build at scale profitably even when the technology works, because yield collapses the math. The economic limit and the physical limit — they're not cleanly separable. Rising fab costs, yield losses, design complexity, quantum leakage — they're all pulling in the same direction at the same time.
Eleanor Crane: And that mutually reinforcing quality is what makes this different from the nineties. Then you could fix one variable. Now every variable is moving against you simultaneously.
Ben Okonkwo: Which is why what the industry does next — chiplets, GPUs, the hyperscaler buildout — it's worth asking whether any of that actually restores what Moore's Law was really delivering. And I don't think the answer is clean. That's where we're headed.
Eleanor Crane: And that's the question — whether what comes next actually restores what Moore's Law was delivering, or whether it just delivers something that looks similar from far enough away.
Ben Okonkwo: So the industry's answer has three parts. Chiplets. Specialized accelerators — GPUs, TPUs. And massive parallelism. And I want to be honest: they work. Chips keep getting faster. More efficient. If you benchmark against the prior generation, the numbers go up.
Ben Okonkwo: But — okay, chiplets specifically. You take multiple smaller dies, manufacture them separately, integrate them into one package. You sidestep yield losses at the leading edge, you avoid some of the EUV cost. AMD's done this, Intel's done this. The performance scaling continues. Except — wait, that's not the same thing as cost-per-computation falling. Spending more to get more is not the same as getting more for less. Moore's Law was the second one.
Eleanor Crane: That distinction — is that actually measurable? Like, can you point to a number?
Ben Okonkwo: The hyperscaler number is the number. The six largest cloud infrastructure operators are on track to spend over seven hundred billion dollars on computing infrastructure in a single year. Roughly twice the prior year's record. That's from MIT Technology Review's reporting. And what they're doing is compensating — through sheer capital deployment — for the chip-level gains that aren't coming automatically anymore.
Eleanor Crane: Seven hundred billion. In one year.
Ben Okonkwo: Twice the prior year's record. And that's the inversion — Gordon Moore's original observation was fundamentally democratizing. Cheaper transistors meant a three-person team in a garage could build something that competed. The hyperscaler model means you need hundreds of billions in infrastructure just to stay at the frontier. Those are structurally opposite economic logics.
Eleanor Crane: Well, and that's where the nineties counterargument feels most alive to me. Because strained silicon, high-k metal gates, copper interconnects — those were also expensive fixes. And they worked. They bought decades. So why is the seven-hundred-billion answer categorically different from the materials-innovation answer?
Ben Okonkwo: Because — I mean, this is the load-bearing distinction. The nineties workarounds restored the cost-per-transistor trajectory. They were expensive to develop, yes, but once deployed, they handed the cost curve back. Strained silicon made the next node cheaper again. Chiplets and TPUs don't do that. They deliver performance for specific workloads. A GPU is extraordinary at matrix multiplication for AI training. It is not cheaper computation in general. It's better computation for a narrow problem, at higher capital cost.
Eleanor Crane: So the nineties fixes were — I want to make sure I have this — they were fixes that paid forward. The capital you spent to develop them came back through the next node's economics. Whereas chiplets and accelerators are more like... you're spending continuously to maintain a level.
Ben Okonkwo: That's — yeah, that's actually the right frame. The process technology innovation era bought you back onto the original curve. Architectural innovation is running a parallel track. It's not the same curve.
Eleanor Crane: Which is what makes the live debate so genuinely hard to call. Because if you're measuring performance, the story sounds like continuation. If you're measuring who can afford the workaround — that's a concentration story.
Ben Okonkwo: And I don't think we can honestly say which story is right yet. Whether this is Moore's Law transformed or Moore's Law ended — that question is still open. The evidence that it's ended is the capital concentration. The evidence that it's transformed is that engineers have been wrong about terminal limits before. Both are real.
Eleanor Crane: And that's the thing I can't quite set down. Whether we call it 'ended' or 'transformed' — I mean, that framing matters beyond terminology. If the answer is 'transformed,' capital flows toward whoever can afford the transformation. Chiplets, specialized accelerators, seven-hundred-billion-dollar buildouts. If the answer is 'ended,' something else gets funded. A different kind of innovation. And those aren't the same future.
Ben Okonkwo: And the $700 billion isn't just a large number — it's the inversion. Moore's original observation said you get more for less. That number says you now spend more to get more. The direction flipped.
Eleanor Crane: Which brings me back to where we started, actually. A chemist at Fairchild Semiconductor, writing in a trade magazine, observing his own industry's cost curve. That was the promise — not that chips would get faster in some abstract sense, but that each generation would be cheaper, and therefore accessible to more people, more companies, more uses. That diffusion was the point. And I don't think architectural innovation automatically inherits it.
Ben Okonkwo: Right — but the part that doesn't fit neatly is that Gordon Moore couldn't have known he was writing a democratization promise. He was describing a cost trend. The civilization-scale implication came later.
Eleanor Crane: No, fair. But the trend carried the promise whether he named it or not. And now the trend has changed, and the promise — well. That's the unresolved part, isn't it.
Ben Okonkwo: Yeah. And I think that's honestly where the evidence leaves us. I can't close it cleaner than that.
Eleanor Crane: A chemist making a cost observation in 1965, and sixty years later two people can't agree on whether it ended or changed shape. He'd probably find that reasonable.
