Jensen Huang – TPU competition, why we should sell chips to China, & Nvidia’s supply chain moat

Dwarkesh opens with the bear case in its strongest form. Software valuations have been collapsing on the expectation that AI commoditizes software. Nvidia, on that reading, is a software company whose software other people manufacture: it sends a GDS2 file to TSMC, TSMC builds the logic dies and the switches, the parts are packaged with HBM from SK Hynix, Micron and Samsung, and an ODM in Taiwan assembles the racks. If software is about to be commoditized, the question writes itself.

Huang answers from the physical end of the chain. Something has to turn electrons into tokens, and making a token more valuable over time is not the kind of thing that commoditizes cleanly. The analogy he reaches for is molecular: making one token worth more than another is like making one molecule worth more than another — artistry, engineering, science and invention, and none of it deeply understood yet.

He then endorses Dwarkesh's framing as his own mental model of the company: input electrons, output tokens, Nvidia in the middle. The operating rule is to do as much as necessary and as little as possible — anything he does not need to do, he partners with someone and folds into the ecosystem.

The result is what he calls the largest partner ecosystem in the industry, upstream and downstream, across what he describes as AI's five-layer cake. The strategy is minimalism, but the residue is not small: the part Nvidia has to do itself turns out to be insanely hard, and that is precisely the part he does not expect to commoditize.

He then inverts the premise. Most software companies, he argues, are tool makers — Excel, PowerPoint, Cadence, Synopsys — with a smaller set of workflow codification systems alongside. On his read, the number of agents grows exponentially and the number of tool users grows exponentially with it, so the number of instances of these tools should skyrocket rather than collapse.

The concrete image is a chip design flow. Instances of Synopsys Design Compiler, of floor planners, layout tools and design rule checkers, all multiply — because today the constraint is the number of engineers, and tomorrow those engineers are supported by fleets of agents exploring a design space nobody has searched before, using the same tools.

So tool use is what makes software companies valuable, not what destroys them. The reason it has not shown up in the numbers yet is narrower: the agents are not good enough at using the tools. That gap closes from both ends — vendors build their own agents, and general agents get good enough to drive existing tools.

Dwarkesh turns to the balance sheet. The latest filings show close to $100 billion of purchase commitments across foundry, memory and packaging, and SemiAnalysis has reported figures nearer $250 billion. One interpretation is that this is the moat: years of scarce components locked up, so that a rival may have an accelerator design but cannot get the memory or the logic to build it.

Huang accepts the framing and widens it. The commitments are partly explicit and partly implicit: much of the upstream investment is made by the supply chain itself, after he has sat with the CEOs of each layer and walked them through how large the industry is going to be, and why.

The interesting question is why they make those investments for him rather than for someone else, and his answer is not about charisma. They invest because they know he can buy their supply and sell it through his downstream. The downstream demand is large enough to de-risk their capex.

Which is why he describes the moat as the ability to build for a future rather than a stock of components. If the next several years are trillion-dollar scale, the chain has to exist first; and nobody builds a supply chain for an architecture whose business churns are low. Sustained scale is downstream demand seen clearly enough, early enough, that other people will bet capex on it.

GTC is the mechanism, and he describes it functionally rather than as a trade show: the whole AI universe in one place because the pieces need to see each other. Downstream sees upstream, upstream sees downstream, everyone sees the advances, and everyone meets the AI natives whose startups are the evidence that what he has been saying is true.

That explains the part of the keynote that critics find tedious. The education segment is deliberate: the supply chain and the ecosystem have to understand what is coming, why, when, and how large, and be able to reason about it systematically — the way he reasons about it. Informing is the job that makes the capex decisions possible.

Dwarkesh pushes on whether the upstream can physically follow. Nvidia has been doubling revenue year over year, and more than tripling the FLOPS it delivers to the world.

Huang's aside is that doubling at this scale is itself remarkable — which sets up the harder version of the question.

The harder version is arithmetic. Nvidia is the largest customer on TSMC's N3 and one of the largest on N2; AI as a whole is reportedly 60% of N3 this year and 86% next. How do you double when you are already the majority of a node, year after year? Does AI compute growth have to slow because of the upstream, and is there any way around it?

His answer starts from a claim about the general condition: at some level instantaneous demand exceeds supply, upstream and downstream, and at any instant the binding constraint could be something absurd — the number of plumbers, which has actually happened.

Dwarkesh's reply is that the plumbers should be invited to next year's GTC. Huang's is that this is a good condition to be in: you want an industry where instantaneous demand exceeds total supply. The bad version is the reverse.

And the mechanism that resolves a specific shortage is swarming. When one component falls too far behind, the industry piles onto it. CoWoS is his example — nobody talks about it much anymore, because for two years they swarmed the living daylights out of it, doubling several times over.

The consequence is a status change. TSMC now knows CoWoS has to keep pace with logic and memory demand, and is scaling packaging at the same rate as logic. CoWoS and HBM stopped being specialties and became mainstream computing technology, which is exactly the transition that makes volume possible.

He dates the argument to before it was consensus. Everything he says now he was saying five years ago, and some people believed it and invested — Sanjay Mehrotra and the Micron team among them, across LPDDR and HBM. He remembers the meeting where he laid out what was going to happen and why. Others arrived later; they are all here now.

The method has since become anticipatory. Rather than waiting for a bottleneck, Nvidia prefetches them years ahead: the investments with Lumentum, Coherent and the silicon photonics ecosystem reshaped the supply chain, and the COUPE work with TSMC produced technology that was then licensed out to keep the chain open rather than captive.

The rest of the toolkit is unglamorous: new workflows, new test equipment such as double-sided probing, direct investment, help scaling capacity. The point is to shape the ecosystem so that the chain is ready before the demand arrives.

Dwarkesh suggests some of these are easier than others — scaling CoWoS versus, say, scaling something else — and Huang interrupts to say he deliberately picked the hardest one.

The hardest is plumbers and electricians, and he uses it as an argument against the doomers. Discourage people from becoming software engineers and you will run out of software engineers. The same prediction was made ten years ago about radiology: the videos are still online saying it will be the first profession to disappear. What the world is actually short of is radiologists.

Dwarkesh returns to the harder technical case: doubling logic output, where memory and logic are both ultimately gated by EUV. How do you get twice as many EUV machines year over year?

Huang's answer is that none of it is hard to scale quickly, given a demand signal — two or three years. Build one and you can build ten; build ten and you can build a million. These are not hard things to replicate.

Dwarkesh asks how far down the chain he actually goes — whether he sits with ASML and tells them that a two-trillion-dollar Nvidia needs far more EUV.

Some directly, some indirectly, some not at all: convince TSMC and ASML is convinced. The skill is identifying the critical pinch points rather than touring every supplier, and if TSMC is convinced, the machines will be there in a few years.