Jensen Huang – Will Nvidia’s moat persist?

Summary

• Jensen dismisses the ASIC threat as concentrated in one customer: “Anthropic is a unique instance, not a trend. Without Anthropic, why would there be any TPU growth at all? It’s 100% Anthropic” — same for Trainium. The switching economics are thin anyway: ASIC margins run ~65% vs Nvidia’s ~70% (“What are you really saving?”), and he dares TPU and Trainium to show up on Dylan’s InferenceMAX — “nobody wants to show up.”
• The moat is the supply chain, and hardware bottlenecks don’t last: ~$100B in disclosed purchase commitments (SemiAnalysis reports $250B), plus implicit upstream capex Jensen unlocks by personally evangelizing supplier CEOs. His categorical claim against the “AI is already 60% of N3” objection: EUV and fab scaling are “easy to do within two or three years. You just need a demand signal” — the longer-lasting constraints are energy and “plumbers and electricians.”
• On the CUDA-moat-erosion thesis (hyperscalers writing their own kernels, OpenAI’s Triton stack replacing cuBLAS/NCCL), Jensen’s answer is TCO, not lock-in: GPUs “are like F1 racers” needing Nvidia’s assigned engineers, who often deliver 2x, 3x, or 50% model speedups that “directly translate to revenues” — “best performance per TCO in the world, bar none.”
• His confessed strategic mistake: “I didn’t deeply internalize that they really had no other options, that a VC would never put in $5-10 billion… That was my miss” — which handed Anthropic to Google/AWS. Now, with reported up to $30B into OpenAI and $10B into Anthropic: “I’m not going to make that same mistake again.”
• The China case, made at full volume: export controls concede the second-largest market and 50% of the world’s AI developers “all for the benefit of one company.” China makes 60% of mainstream chips, maybe more; Huawei just shipped “millions” in a record year, energy abundance compensates for 7nm (“7nm chips are essentially Hopper”), and “the day that DeepSeek comes out on Huawei first, that is a horrible outcome for our nation.”
• Dwarkesh’s counter stands in the record: Anthropic’s Mythos Preview found thousands of high-severity vulnerabilities including a 27-year OpenBSD zero-day — compute is the input to cyber-offensive capability. Jensen rejects the enriched-uranium analogy as “lousy” and the framing as a “loser attitude”; the disagreement never resolves.
• New tradeable details: Nvidia “added Groq” and will “fold Groq into our CUDA ecosystem” to serve a premium low-latency, high-ASP token segment — “expand the Pareto frontier”; roadmap cadence Vera Rubin → Vera Rubin Ultra → Feynman, while challenging other ASIC teams to match token-cost decreases “by an order of magnitude every single year”; allocation is never auctioned — FIFO on POs, prices never raised into demand.
• Software commoditization inverted: agents and tool users multiply, so it’s “very likely” tool instances will skyrocket — Synopsys Design Compiler, floor planners — and “tool use is going to cause the software companies to skyrocket”; it hasn’t yet only because “the agents aren’t good enough at using their tools yet.”

Deep dive

1. Electrons in, tokens out — the middle doesn’t commoditize

• Dwarkesh’s opener is the bear thesis: Nvidia sends a GDS2 file to TSMC, SK Hynix/Micron/Samsung supply the HBM, an ODM in Taiwan assembles the racks — “Nvidia is fundamentally making software that other people are manufacturing,” so if AI commoditizes software, does it commoditize Nvidia? Jensen adopts the frame rather than fighting it: “The input is electrons, the output is tokens. In the middle is Nvidia” — and that transformation “is far from deeply understood and the journey is far from over. I doubt that it will happen.”
• The operating philosophy in one line: do “as much as necessary and as little as possible” — whatever Nvidia doesn’t have to do gets partnered out across “probably the largest ecosystem of partners” spanning what he calls AI’s “five-layer cake.” The part Nvidia must do “as it turns out, is insanely hard.”
• On crashed software valuations, Jensen sees “the opposite of what people see”: agents and tool users grow exponentially, so it’s very likely that tool instances will skyrocket — Synopsys Design Compiler, floor planners, design-rule checkers. “Tool use is going to cause the software companies to skyrocket”; the only reason it hasn’t yet is because “the agents aren’t good enough at using their tools yet.”

2. The supply-chain moat runs on CEO evangelism, not just contracts

• The ~$100B disclosed purchase commitments (SemiAnalysis reports $250B coming) are only the explicit half. The implicit moat is upstream capex Jensen unlocks personally: “Let me tell you how big this industry is going to be… let me reason through it with you.” Suppliers invest for Nvidia rather than someone else because “they know that I have the capacity to buy their supply and sell it through my downstream” — and “nobody is going to build a supply chain for an architecture if the business churns are low.”
• GTC decoded as infrastructure: the keynotes are deliberately “a little torturous… like education” because the whole chain must “reason about it systematically, just like I reason about it.”
• The best specimen of the evangelism working: Sanjay and Micron believed it five years ago and “really doubled down” across LPDDR and HBM — “it obviously has been tremendous for the company.”

3. Every bottleneck is a 2-3 year problem — except energy and plumbers

• Dwarkesh’s numbers-driven pushback: Nvidia is TSMC’s biggest N3 customer, AI takes 60% of N3 this year and 86% next per SemiAnalysis — “how do you double if you’re the majority?” Jensen’s answer is categorical: EUV and logic scaling are “easy to do within two or three years. You just need a demand signal. Once you can build one, you can build ten, and once you can build ten, you can build a million.” If TSMC is convinced, ASML follows.
• CoWoS as the proof: “we swarmed the living daylights out of it” for two years — it and HBM went from specialty to “mainstream computing technology.” Now Nvidia prefetches bottlenecks years ahead: silicon photonics via Lumentum and Coherent, COUPE co-developed with TSMC with patents licensed out to keep the chain open, new test gear like double-sided probing.
• What actually worries him is downstream: energy policy — “you can’t create an industry without energy” — and the hardest bottleneck of all: “Plumbers. Plumbers and electricians.” His anti-doomer parable: ten years ago they said don’t be a radiologist — “guess what we’re short of? Radiologists.”
• Would Nvidia backport to spare N7 capacity before 2030? Not necessary, he says — retrofitting modern numerics and packaging onto an old node is “a level of R&D that no one could afford. We could afford to lean forward.” Though in the thought experiment where capacity never grows again: “In a heartbeat, of course I would.”

4. TPU is a different animal — programmability is the compounding advantage

• On two of the top three models (Claude, Gemini) being TPU-trained: “We build a very different thing.” Accelerated computing spans molecular dynamics to particle physics, is designed to be operated by others, and sits in every cloud — “our market reach is far greater than any TPU or ASIC can possibly have.”
• Dwarkesh channels his AI-researcher friends: AI is “predictable matrix multiplies again and again” — the TPU’s systolic array fits exactly the workload where revenue is growing; why sacrifice die area for warp schedulers? Jensen’s rebuttal: new attention mechanisms, hybrid SSMs, diffusion-autoregressive fusion all demand general programmability, and “the ability to invent new algorithms is really what makes AI advance so quickly.”
• The co-design proof point: Moore’s Law gives ~25%/year, yet Blackwell over Hopper was announced at 35x energy efficiency — “nobody believed it. Then Dylan wrote an article saying I sandbagged, and it’s actually fifty times.” Only possible by changing processor, fabric (NVLink), network (Spectrum-X), libraries, and algorithm simultaneously: “Without CUDA to do that, I wouldn’t even know where to start.”
• Why not parallel wafer-scale or Dojo-style projects? “We could… It’s just not better. We simulate it all in our simulator, provably worse.” But workload shifts do trigger moves: “recently we added Groq, and we’re going to fold Groq into our CUDA ecosystem” — token value now supports a premium low-latency, high-ASP segment, so Nvidia will “expand the Pareto frontier” with faster-response, lower-throughput inference.

5. Can 70% margins survive customers who write their own kernels?

• Dwarkesh’s sharpest structural question: 60% of revenue is five hyperscalers who can and do route around CUDA — OpenAI’s Triton stack goes “down to CUDA C++, instead of using cuBLAS and NCCL” and compiles to other accelerators. If your biggest customers can afford to build past the moat, doesn’t it collapse into specs-per-dollar competition — and can 70%+ margins hold?
• Jensen’s analogy: a CPU is a Cadillac, Nvidia’s GPUs “are like F1 racers” — anyone can drive 100mph, pushing the limit takes expertise. Nvidia’s assigned lab engineers often deliver 2x, 3x, or 50% model speedups, and across an installed Hopper/Blackwell fleet that directly translates to revenues.
• The TCO gauntlet, thrown publicly: “Nvidia’s computing stack is the best performance per TCO in the world, bar none… Not one company” can beat it. InferenceMAX is available for everyone to use, and MLPerf is another benchmark — “TPU won’t come, Trainium won’t come… Nobody wants to show up” — and he’d “welcome Trainium to demonstrate their 40% that they claim all the time.”
• Correcting the customer-concentration premise: most hyperscaler Nvidia capacity serves external customers, not internal use (all of OCI’s, most of AWS’s and all of Azure’s). The flywheel: “several hundred million GPUs” installed, programmability, ecosystem richness, and the highest tokens-per-watt — decisive when a 1GW datacenter must maximize token revenue.

6. The regret: Anthropic went elsewhere because Nvidia couldn’t write the check

• Why did Anthropic just sign a multi-gigawatt Broadcom/Google TPU deal if Nvidia’s TCO wins on paper? “Anthropic is a unique instance, not a trend. Without Anthropic, why would there be any TPU growth at all? It’s 100% Anthropic” — identically for Trainium. And note the cancelled-ASIC graveyard: “It’s not that easy building something better than Nvidia.”
• The confessed mistake, kept as hedged and raw as spoken: “I didn’t deeply internalize that they really had no other options, that a VC would never put in $5-10 billion of investment into an AI lab… So that was my miss.” Google and AWS invested where Nvidia — never having invested outside the company at that scale — couldn’t. Now, with reported up to $30B into OpenAI and $10B into Anthropic: “I’m not going to make that same mistake again.”
• Why fund every foundation lab rather than pick winners: of 60 original 3D-graphics companies Nvidia is the sole survivor, despite a first architecture that was “precisely wrong… Everybody would have counted us out. And here we are.” The rule drawn from that humility: “Don’t pick winners. Either let them all take care of themselves, or take care of all of them.”

7. Why Nvidia won’t become a hyperscaler — and the neocloud tension

• Same philosophy answers the “become a cloud” question: if Nvidia hadn’t built CUDA (20 years, “losing money most of that time”), CUDA-X, or cuLitho, “nobody would have… I am completely certain of that.” Clouds are the opposite case: “If I didn’t do it, somebody would show up.”
• Dwarkesh flags the contradiction — Jensen denies propping up neoclouds, then says CoreWeave (reported $6.3B backstop, $2B invested), Nscale, and Nebius “wouldn’t exist” without Nvidia. His reconciliation: they must want to exist and bring a plan and expertise; Nvidia may provide investment to get them off the ground, but “we’d rather work with all the people in the financing business than be a financier ourselves.”

8. No auctions, no price hikes, and a roadmap to bet the farm on

• On allocating scarce GPUs: never highest-bidder — “it’s a bad business practice… If I quoted you a price, we quoted you a price. That’s it. If demand goes through the roof, so be it.” The actual system: forecasting, then FIFO on purchase orders, with adjustments if your datacenter isn’t ready. The Larry/Elon dinner legend, debunked directly: “At no time did they beg for GPUs. They just had to place an order.”
• Dependability as the product: ~30 years with TSMC and “Nvidia and TSMC don’t have a legal contract. There’s always some rough justice.” The roadmap cadence: Vera Rubin this year, Vera Rubin Ultra next, then Feynman, then an unnamed successor — and he challenges other ASIC teams to match a token-cost decrease “by an order of magnitude every single year.” From one graphics card to “$100 billion of AI factory, no problem. We’re the only company in the world where you can say that today.”

9. China already has the compute — Jensen’s case against conceding China

• Dwarkesh’s steel-manned case for controls: Anthropic’s Mythos Preview found “thousands of high-severity vulnerabilities across every major operating system,” including a 27-year-old OpenBSD zero-day — if Chinese labs had the chips to train such a model and run “millions of instances,” isn’t that a weapon-adjacent capability?
• Jensen’s counter starts with facts on the ground: Mythos was trained on “fairly mundane” compute “abundantly available in China” — a country making 60% of the world’s mainstream chips, maybe more, holding 50% of the world’s AI researchers, with abundant energy including “ghost datacenters… fully powered.” Huawei just had its largest year ever, shipping “a ton. Millions. Millions is way more than Anthropic has.” Verdict: “The idea that China won’t be able to have AI chips is completely nonsense.”
• The substitution logic: in the five-layer cake, energy at the bottom trades off against chips — “if your amount of watts is completely abundant… what do you care about performance per watt for?” Jensen responds to the EUV claim about advanced HBM by saying they can gang HBM2 together with silicon photonics “just like we gang them together with NVL72.” Dwarkesh’s EUV premise, in Jensen’s words, “is just wrong.”
• The deeper claim: Moore’s Law is ~25%/year but algorithms deliver 10x — “most of the advances in AI came out of algorithm advances,” making China’s researcher army the fundamental asset. The nightmare scenario, verbatim: “The day that DeepSeek comes out on Huawei first, that is a horrible outcome for our nation” — open models optimized for a foreign stack, diffused out to the global south.

10. “That loser attitude makes no sense to me” — the clash that never resolves

• Dwarkesh’s enriched-uranium analogy for compute draws a flat rejection: “It’s a lousy analogy. It’s an illogical analogy” — followed by the extremes charge: “Your argument starts from extremes… Those extremes, they’re childish.” Dwarkesh holds his ground — “any marginal compute is helpful” to training offensive-capable models — and Jensen acknowledges the potential cost but reframes it: the real cost is conceding the chip layer.
• The market math: China is ~40% of the world’s technology industry and half its AI developers; conceding it is “a disservice to our country… all for the benefit of one company.” Ecosystems aren’t cars: “There’s a reason why the x86 deal exists. There’s a reason why ARM is so sticky.” And the precedent he keeps returning to: similar policy “resulted in the American telecommunications industry being policied out of basically the world.”
• When Dwarkesh suggests Nvidia would lose China share anyway, à la Tesla and iPhone: “You’re not talking to somebody who woke up a loser. That loser attitude, that loser premise makes no sense to me.” Lithography isn’t destiny either — Blackwell is 50x Hopper on only ~75% transistor improvement over three years: “Architecture matters. Computer science matters.”
• His closing prediction, set as a marker: in a few years, when America wants its stack exported to India, the Middle East, Africa, “I want you and I to have that same conversation again. I will tell you exactly about today’s conversation, about how your policy… literally caused the United States to concede the second largest market in the world for no good reason at all. If we lose it, we lose it. But why do we concede it?” He insists: “We should always have the best technology here… and the first.”