Hope Sterling: Hey, I have to warn you — I came in today a little bit haunted.
Spuds Oxley: Haunted by what exactly?
Hope Sterling: By a contradiction that I keep turning over — like, TSMC is actively shipping 2-nanometer chips right now, this year, Intel is ramping its 18A process which is roughly 1.8 nanometers, imec has a roadmap projecting all the way down to 0.3 nanometers by 2038, and IBM literally doubled its own transistor density record with about a hundred billion transistors on a fingernail-sized chip in 2024 — and somehow the official story is that Moore's Law is dead?
Spuds Oxley: That is a genuinely strange thing to hold in your head at once.
Hope Sterling: Right — but the part that doesn't fit is, like, if it's dead why is everyone still racing down the same road? That's what today is actually about — because Moore's Law, Gordon Moore's original 1965 observation in Electronics Magazine, was that transistor count would double roughly every year, and he revised it to every two years in 1975, and that rhythm is what made chips cheaper for everybody for sixty years.
Spuds Oxley: Now Pat Gelsinger — Intel's CEO — said the doubling period has stretched to three to four years. Not two. That number is, I think, the hinge of the whole argument. Because a slowing train and a stopped train feel identical until you check whether you're still moving.
Hope Sterling: Okay the slowing-versus-stopped thing — that's exactly the thing I couldn't name, yeah.
Spuds Oxley: And the question underneath it — whether the death announcement is really about physics or about something else entirely — that's what I want to pull apart.
Hope Sterling: But okay, physics versus money — that's actually the thing I can't unsnarl, like, are we hitting a wall we literally cannot cross, or did we just decide it costs too much to cross it?
Spuds Oxley: Well, there's a city in there somewhere. Picture a city that keeps getting denser — more people, more buildings, same pipes, same power lines. For a long time that's fine, actually great, more people same energy bill. Then at some point the pipes can't carry the water and the heat from all those bodies can't escape the buildings. That's the transistor situation. No metaphor, just — that's it.
Hope Sterling: Wait, the heat thing is literal? Like actual heat?
Spuds Oxley: Completely literal. Now, for fifty years there was a companion rule to Moore's Law that almost nobody talks about — Dennard scaling, 1974. The principle was that as transistors shrank, you could also reduce their voltage and current proportionally, which kept power density roughly constant. Smaller, faster, same heat. That's what let clock speeds climb and climb. It's what made Moore's Law affordable, not just physically real.
Hope Sterling: Dennard scaling was like — Moore's Law needed a wingman this whole time and we just didn't know it?
Spuds Oxley: The wingman disappeared in the mid-2000s. Below a certain size, voltage could not be reduced proportionally anymore. Power density started climbing instead of holding flat. The industry couldn't just keep racing the clock speed up — so they split the core, went multicore, which is a workaround, not a solution to the underlying heat problem.
Hope Sterling: That's why my laptop has eight cores instead of one insanely fast core.
Spuds Oxley: Exactly that.
Hope Sterling: Okay but — wait, the quantum tunneling thing, I need that explained because I've read the words and they make no sense to me. An electron just goes through a wall it's not supposed to cross?
Spuds Oxley: Truth is, it's almost more unsettling than the heat problem. Below roughly 5 nanometers, electrons don't read the off switch. There's an insulating barrier — the wall that's supposed to say 'not today' — and the electron just... appears on the other side. Probabilistically. It hasn't broken through, it's tunneled. And at that scale it's not a curiosity, it's constant leakage. The transistor is switched off and current is moving anyway.
Hope Sterling: So the transistor is just — lying to you. It says off, it means sort of off.
Spuds Oxley: And that leakage generates more heat, which feeds the first problem, which limits the clock speed, which is why crossing below 5nm stopped yielding the cost-per-transistor improvements that made Moore's Law the promise it was. The two walls aren't separate — they're the same wall from two directions.
Hope Sterling: And that's — okay, that's the thing that nobody led with when I was reading about this. Because I kept hearing 'physics problem, physics problem,' and yes, fine, quantum tunneling, I get it now — but the cost-per-transistor thing? That's a completely different story and it's the one I'm kind of furious about.
Spuds Oxley: You see, that's the decoupling nobody names cleanly.
Hope Sterling: TSMC deferred — like, literally said no thank you — to a High-NA EUV machine from ASML. Not because it doesn't work. Because each one costs four hundred million dollars.
Spuds Oxley: Four hundred million. Per tool.
Hope Sterling: Per tool! And ASML is the only company on earth that makes them — like, there is no alternative vendor, no competitor, it's just ASML — so if the price is prohibitive, you just... don't get the machine. That's a money wall, not a physics wall, and it's sitting right on top of the physics wall and everyone's conflating them.
Spuds Oxley: A single chokepoint in the entire global supply chain for leading-edge chips. One manufacturer. That's — well, that's a fragile thing to build a civilization's computing future on.
Hope Sterling: And it used to be that every new process node made each transistor cheaper to manufacture — that was the whole engine, right? Like, the reason Moore's Law mattered to the startup in a garage wasn't transistor counts, it was that the transistors got cheaper. Below 5nm that curve just... bent the wrong way.
Spuds Oxley: Now only a handful of companies can even play at the bleeding edge. TSMC, Intel, Samsung. That's roughly the list. Everyone else is buying yesterday's process node.
Hope Sterling: Which means — wait, is this actually a concentration problem? Like, the thing that made Moore's Law democratic is exactly what's gone?
Spuds Oxley: That's the loss. For sixty years the promise was: if you wait, it gets cheaper for everyone. The garage gets what the giant had four years ago. That promise required the cost curve to keep falling — and it has stopped falling at the frontier. The physical and economic dimensions of scaling have come apart. They used to be the same thing.
Hope Sterling: Okay and — I mean, there's an answer coming, like the industry does have architectural moves it's making, and honestly that part makes this even more complicated because I'm not sure it's actually the same game anymore.
Spuds Oxley: Funny enough, that's exactly where the story tilts next — because the workarounds are real, but they come with a different set of winners.
Hope Sterling: Different winners — okay, but like, different HOW? Because the industry keeps saying 'we have answers, chiplets, stacking, new transistor designs,' and I want to believe that, I do, but I keep thinking... is that actually the same game or did we just quietly change the rules?
Spuds Oxley: Well, take gate-all-around transistors. That's the actual engineering answer to the quantum tunneling wall — instead of the gate touching the channel on three sides like FinFET, it wraps all the way around. Better electrostatic control, you can keep shrinking below 3nm without the electron just deciding the off switch is optional.
Hope Sterling: Wait — Samsung moved to that first, before TSMC?
Spuds Oxley: Samsung adopted it ahead of TSMC. TSMC brought it to 2nm production in late 2024. So that's real progress, that's not marketing. But the transistor architecture is only one piece now.
Hope Sterling: Because chiplets are a completely separate move, right? Like, that's not shrinking — that's just... breaking the chip into Lego bricks and hoping they talk to each other?
Spuds Oxley: That's — actually that's closer than it sounds. You take a monolithic chip, a single giant die, and you break it into smaller modular dies made separately, then integrate them into one package. The yield benefit alone is significant — a defect in a small chiplet kills a small piece, not the whole thing. But the deeper point is you can mix process nodes. Your compute die at 2nm, your memory interface at an older node, all talking together.
Hope Sterling: Okay that's actually kind of genius — wait, but someone has to design all those handshakes between the pieces, and that's not cheap.
Spuds Oxley: That's the hardest fact in this whole story. Chiplets, custom silicon, 3D stacking — and HBM, High Bandwidth Memory, which stacks dies vertically to feed AI accelerators enough data fast enough — all of it requires massive upfront design investment. Teams of hundreds of engineers. Apple has that. Google has that. Amazon, NVIDIA — they have that.
Hope Sterling: And the startup in a garage —
Spuds Oxley: Does not have that. Picture a solo developer in 2005 — the hardware wave was coming for everyone on the same schedule. Wait two years, your phone is twice as capable, same price. In 2025, the best silicon is bespoke. It's designed for a specific workload by a team that costs more than most companies' entire valuation.
Hope Sterling: So 'More than Moore' — like, heterogeneous integration, specialization instead of shrinking — that's not a rising tide. It's a rising tide for like five companies.
Spuds Oxley: That's the paradigm shift nobody names. More Moore meant: physics does the work, everyone benefits. More than Moore means: design does the work — and design is expensive, and it concentrates.
Hope Sterling: And like — specialized accelerators, GPUs, TPUs, the custom ASICs, those are incredible for the workloads they're built for, but they're not general purpose, which means... I don't know, is that progress or is that just a very expensive kind of lock-in?
Spuds Oxley: Truth is, I think it's both. The efficiency gains are real — an ASIC built for one task will bury a general-purpose chip on that task. But the old promise was that the gains were universal, automatic, almost... democratic. This one isn't. The paradigm hasn't just shifted, it's narrowed. And I'm not sure the industry has fully reckoned with what it means to build a future where the best compute is only reachable by the people who could already afford it.
Hope Sterling: And that's — I mean, for sixty years a garage startup and IBM were riding the same wave. Like, literally the same physics was doing the lifting for both of them. The startup just got IBM's 2005 capability by 2010. That inheritance mechanism — that's what's breaking. Not the transistors, not really. The inheritance.
Spuds Oxley: And we don't yet know — I mean, genuinely don't know — whether chiplets and specialized accelerators can rebuild that rising tide, or whether they're just a better deal for the people who were already winning.
Hope Sterling: Which is — yeah. That's the unresolved part. The cost structure for AI compute right now, HBM, the specialized accelerators, the billion-dollar fabs — all of that only works if architectural scaling can somehow replicate the cost curves that density scaling used to just... automatically generate. And nobody knows if it can.
Spuds Oxley: No one does. And if it can't — if no rising tide replaces the old one — then who defines 'good enough' computing stops being a temporary question. It becomes a structural feature of the next era.
Hope Sterling: That's the part that I think I'll be sitting with for a while, honestly.
Spuds Oxley: Same. Gordon Moore drew a line on graph paper and it turned into a promise. Sixty years is a long time to keep a promise. The fact that it's straining now doesn't mean it failed — it means it ran as far as physics would let it go.
Hope Sterling: That's — I actually needed that framing. Thank you for going down this one with me.
Spuds Oxley: Worth every nanometer.
