Dwarkesh Podcast
Dwarkesh Podcast

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

I asked Jensen about TPU competition, Nvidia’s lock on the ever more bottlenecked supply chain needed to make advanced chips, whether we should be selling AI chips to China, why Nvidia doesn’t just become a hyperscaler, how it makes its investments, and much more. Enjoy! Watch on YouTube; read the t

Featured Speakers

Dwarkesh Patel HostJensen Huang Guest

Topics Discussed

Episode Summary

Executive Summary: Jensen Huang argues NVIDIA won’t be commoditized by AI because it sits at the hard center of the “electrons to tokens” transformation. He says NVIDIA’s ecosystem, supply-chain orchestration, CUDA, and continuous architecture gains preserve its moat, while U.S. policy should avoid cutting off China and instead keep the American tech stack dominant worldwide.

Main Topics: NVIDIA’s role in AI (Priority: 5/5): NVIDIA is positioned as the core transformer from electricity to valuable tokens. AI and software commoditization (Priority: 4/5): AI may commoditize some software, but tools and agents should expand software usage. Supply-chain coordination and scarcity (Priority: 5/5): Long-term commitments and ecosystem alignment are how NVIDIA scales constrained inputs. CUDA, install base, and ecosystem moat (Priority: 5/5): CUDA, broad install base, and cloud ubiquity make NVIDIA hard to displace. Competition with TPUs and ASICs (Priority: 4/5): Specialized chips can help, but NVIDIA argues general programmability wins over time. China export controls and AI leadership (Priority: 5/5): Huang says cutting China off harms U.S. leadership and accelerates Chinese self-sufficiency. Capital allocation and ecosystem investing (Priority: 4/5): NVIDIA should support partners and labs, but not become a cloud or financier.

Key Arguments: AI won’t commoditize NVIDIA because making tokens valuable is “insanely hard” and still under-invented. Tool use should rise as agents grow, so software tools like Synopsys/Cadence should see more usage. NVIDIA’s moat is ecosystem scale: upstream supply, downstream demand, clouds, developers, and model makers. Bottlenecks like memory, packaging, and EUV can be solved in 2-3 years with demand signals. CUDA matters because it combines programmability, install base, and trust across every major cloud. TPUs/ASICs are specialized, but NVIDIA says general programmable compute better supports new AI algorithms. China should not be cut off because the U.S. would forfeit a huge market and help China build its own stack.

Data Points: Purchase commitments: almost $100 billion - Reported filings cited as upstream supply commitments Purchase commitments: $250 billion - Semi-Analysis estimate mentioned in the interview Revenue growth: 2Xing revenue year over year - Description of NVIDIA’s recent growth cadence Compute growth: more than tripling the amount of flops - Year-over-year increase referenced by interviewer TSMC N3 share: 60% - AI as a whole this year is said to be 60% of N3 TSMC N3 share: 86% - AI as a whole next year is said to be 86% of N3 AI efficiency gain: 10x, 20x - Huang cites computing-efficiency gains beyond capacity growth Blackwell vs Hopper: 30, 50x - He says Hopper to Blackwell improved efficiency by this range Blackwell vs Hopper: 35 times - Initial public claim for Blackwell energy efficiency Blackwell vs Hopper: 50 times - Revision cited from an article and Huang’s clarification Moore’s Law: about 25% per year - Used to argue hardware alone cannot explain AI leaps Legacy CPU count: 60 graphics companies - Historical comparison about NVIDIA surviving a crowded market AI researchers in China: 50% - He claims China has half of the world’s AI researchers China’s technology share: 40% - He says China is about 40% of the world’s technology industry Infrastructure scale: one gigawatt - Example of the size of a data center NVIDIA wants to maximize Modeling claim: 20,000 GPU hours - Jane Street backdoor puzzle example referenced in discussion Inference benchmark: up to 10 times faster - Crusoe benchmark cited in sponsor read Inference benchmark: up to 5 times better throughput - Crusoe benchmark cited in sponsor read Investor commitments: up to $30 billion - Reported OpenAI investment amount cited in discussion Investor commitments: $10 billion - Reported Anthropic investment amount cited in discussion CoreWeave backstop: up to 6.3 billion - NVIDIA support for CoreWeave referenced in conversation CoreWeave invested: 2B - Amount said to have already been invested

Pivotal Quotes: "The input is electron. The output is tokens. That is in the middle, NVIDIA." — Jensen Huang: He defines NVIDIA’s core role in the AI value chain "We should do as much as needed, as little as possible." — Jensen Huang: He explains NVIDIA’s philosophy on ecosystem support and investments "The single most important thing to our company is our richness of our ecosystem" — Jensen Huang: He describes why CUDA and partner breadth are central to NVIDIA’s moat

Implications: The unresolved question is whether AI hardware markets will stay winner-take-most; listeners should watch supply-chain execution, ecosystem adoption, and U.S.-China policy.

🔓 Sign Up for Unlimited Episode Search

About Dwarkesh Podcast

Deeply researched interviews

View all episodes from Dwarkesh Podcast