Hugo Vance: You look like you've been reading chip industry news, which is never a good sign for your weekend.
Lila Soto: Ha — yeah, I have, and I have a number for you. TSMC, 2026: they halted purchases of the most advanced chipmaking equipment available. The world's leading foundry just said no to the best tools money can buy. What does that tell you?
Hugo Vance: It tells me the cost-per-transistor link has genuinely fractured. That's the tell.
Lila Soto: Say more — because I think a lot of people hear that and go, sure, supply chain thing, geopolitics, whatever.
Hugo Vance: Well, it's neither of those. You have to go back to 1965 — Gordon Moore at Fairchild Semiconductor observing that transistor counts on a chip doubled annually. He revised that to every two years in 1975. And from that point, the entire semiconductor industry built its investment logic around that rhythm. Not because nature forced it to. Because the observation became a coordination mechanism — almost self-fulfilling.
Lila Soto: The roadmap becomes the reality.
Hugo Vance: Precisely. And the reason it worked for sixty years wasn't the density per se — it was that more transistors per wafer meant the cost of each individual transistor kept falling. Computing became cheap enough to be everywhere. Now that link has snapped: you can still get denser chips, technically, but the wafer costs more. The two things that always moved together have come apart. TSMC stopping equipment purchases is them naming that, quietly.
Lila Soto: Hm — so it's not that chips stop getting better, it's that they stop getting cheaper. Which is kind of the whole point of chips getting better.
Hugo Vance: That's exactly the bargain. And it's worth being concrete about what that bargain actually looked like. The Computer History Museum documents transistors costing several dollars each in the early 1960s. By the time Gordon Moore wrote his 1965 paper, the curve was already bending toward negligible. Think of it this way — imagine a grocery store that, every two years, figured out how to fit twice as many items on the same shelf. Not because food got cheaper to grow, but purely from the stacking. The price per item gets cut in half each cycle. That's Moore's Law for transistors.
Lila Soto: And the shelf is the wafer.
Hugo Vance: The wafer is the shelf, yes. Fixed cost, more units on it, unit cost falls. Simple arithmetic with profound consequences.
Lila Soto: So who was actually running this? Like, who was enforcing the two-year cadence? Because it sounds almost — I mean, was there an actual schedule somewhere?
Hugo Vance: Intel, largely. They internalized Moore's observation — Moore co-founded Intel, you see — and built their entire R&D calendar around it. When Intel committed to a new process node every twenty-four months, their suppliers had to commit too. Their competitors had to match or fall behind. The prediction created the pressure that made the prediction come true.
Lila Soto: Wait — so Intel was kind of setting the tempo for the whole industry? Like a metronome nobody voted on?
Hugo Vance: Precisely that. And once capital markets priced in that cadence — once venture funding, factory construction cycles, chip design timelines all assumed the doubling — you had a self-reinforcing loop. Miss the beat and your cost structure falls apart relative to competitors who didn't.
Lila Soto: So it's less a law of physics and more a — kind of industry-wide contract with itself.
Hugo Vance: A coordination mechanism. That's the precise word. Moore himself understood this — which is why his 1975 revision to biennial doubling wasn't a retreat. It was the mechanism self-correcting to stay credible. A law of nature doesn't revise itself. A roadmap does.
Lila Soto: But that contract held — and then, around 2020, it kind of just... stopped. I keep trying to picture what that actually feels like from the inside. Like, a CFO at some mid-sized software company, building a three-year hardware budget in late 2025, and the whole assumption is: compute costs will halve on schedule, same as always. And they won't.
Hugo Vance: And nobody sent a memo.
Lila Soto: Right — but the part that doesn't fit is, density is technically still going up. IBM's fingernail-sized die, a hundred billion transistors in two stacked layers — that's real. So how is the law dead if the transistors are still doubling?
Hugo Vance: Because the doubling and the cost collapse were always two different things that happened to travel together. Classical planar scaling — shrinking transistors flat, in the same plane — hit physical walls around 2020. Quantum tunneling, short-channel effects, thermal management. You can't keep shrinking the geometry and expect electrons to behave. The physics stopped cooperating.
Lila Soto: Wait — quantum tunneling meaning electrons just... leak through?
Hugo Vance: Essentially. The gate that's supposed to switch the transistor on and off becomes thin enough that current passes whether you want it to or not. You get heat, noise, waste. And heat is the other wall — you can only dissipate so much from something that small. So the industry moved to 3D stacking, chiplets, novel materials. IBM's prototype is brilliant engineering. But here's what that CFO needs to understand: a more expensive wafer that's denser is not the same bargain. Advanced node costs per wafer have risen. The density went up; the price didn't come down.
Lila Soto: So Moore's Law is technically alive on one number and effectively dead on the number that actually drove competitive behavior.
Hugo Vance: That's the precise fracture. And TSMC halting purchases of the most advanced equipment in 2026 — that's not a supply chain hiccup. That's the world's dominant foundry deciding the cost curve no longer justifies the next step.
Lila Soto: Which — I mean, that changes everything about who survives this. And the way some companies have already repositioned around it, Apple and Nvidia especially, is pretty wild.
Hugo Vance: Yes — and the repositioning is not accidental. When the cost-decline competition stops rewarding you, you stop competing on it. Apple saw that clearly. The M5 Pro and M5 Max, announced March 2026 — chiplet architecture, not a single monolithic die. They're assembling chips from modular pieces, mix-and-match, and the yield improvements alone change the economics of what you can build.
Lila Soto: Wait — chiplet meaning, what, like Lego bricks instead of one solid piece?
Hugo Vance: Essentially. You fabricate smaller dies, each optimized for its function, then connect them. A flawed monolithic die is a dead chip. A flawed chiplet is a discarded module. The whole unit survives. But the deeper move — Apple is competing on software-hardware integration now. Not transistor count. The chip is designed around the operating system, around their own workloads. That's vertical integration as strategy.
Lila Soto: And they can do that because of the fabless model — TSMC makes the chip, Apple designs it.
Hugo Vance: Exactly. TSMC's foundry scale is what made that separation possible — design differentiation without owning a fab. Nvidia runs the same playbook, though differently. They're not chasing density cycles at all. They compete through architectural specialization — GPU workloads, AI inference, the specific math that large language models need. They own the architecture for that problem.
Lila Soto: Hm. So both of them opted out of the commodity race before it broke, or — I mean, did they see it coming, or did they just happen to be building something different?
Hugo Vance: Well, I think — actually, the honest answer is: both. Nvidia's GPU specialization wasn't a hedge against Moore's Law slowing. It was a bet on a different workload. The Moore's Law fracture just happened to reward that bet enormously. Apple is the more deliberate case — vertical integration was their explicit strategy going back to the M1.
Lila Soto: But then IBM's hundred-billion-transistor die — that's still someone believing density is the game.
Hugo Vance: Yes, and I won't dismiss it. Three-D stacking, dual-sided wafer integration, atomic-level deposition — those extend the density trajectory genuinely. Willy Shih at Harvard Business School frames it as extending Moore's Law by other means. I think the boundary between extending and replacing is honestly contested. What I'd insist on: IBM can build it. The question of whether a customer can buy it at a price that resets the commodity race — that remains unanswered.
Lila Soto: So density as a fact, and cost-decline as the actual power — those two are just... uncoupled now, and nobody's quite naming it.
Hugo Vance: They're naming it, quietly, in the only language capital markets actually hear — which is where you put the money next. Apple and Nvidia aren't winning because they found a better way to chase Moore's Law. They own the stack. Software, silicon, the workload it was designed for. That's the automotive model, actually. Or pharmaceuticals. Fragmented by specialization, dominated by whoever controls the full chain. Not a universal commodity race anymore.
Lila Soto: The law isn't dead, but the bargain it created — density up, cost down, on schedule, for everyone — that's the thing that's gone. I think that's what I actually buy.
Hugo Vance: Yes. That's the precise landing. Gordon Moore revised a roadmap in 1975 to keep it credible. Nobody gets to revise this one. The cost half simply isn't there.
Lila Soto: Good place to stop. Thanks for sitting through my chip anxiety with me.
