Dwarkesh + Jensen Huang: Nvidia moat

Category: Expert Interviews · Duration: 103 min · ▶ Watch

Speakers: Dwarkesh Patel · Jensen Huang

Switch language → 中文

Segments (22)

  • 00:00:00 · Will AI Commoditize NVIDIA?
    • The host asks if NVIDIA will be commoditized as AI commoditizes software, outlining NVIDIA’s reliance on a complex manufacturing supply chain.
  • 00:00:32 · The Value of Transforming Electrons to Tokens
    • Jensen Huang argues that the process of transforming electrons into valuable tokens involves immense artistry and science, making it difficult to commoditize.
  • 00:01:40 · NVIDIA’s Business Model and the 5-Layer AI Cake
    • Jensen describes NVIDIA’s mental model as being the middleman between electrons and tokens, doing as little as possible by leveraging a massive partner ecosystem across a five-layer AI stack.
  • 00:02:50 · The Future of Software Tools in the AI Era
    • Jensen predicts that the number of AI agents and tool users will grow exponentially, causing the use of software tools like those from Cadence and Synopsys to skyrocket.
  • 00:04:30 · NVIDIA’s Supply Chain Moat and Purchase Commitments
    • The host questions if NVIDIA’s massive purchase commitments, which lock up the supply chain for years, constitute its primary moat.
  • 00:05:00 · The Upstream/Downstream Flywheel
    • Jensen explains that NVIDIA’s ability to make huge upstream commitments is enabled by its massive downstream demand, creating a powerful flywheel effect.
  • 00:06:10 · Can the Supply Chain Keep Up?
    • The host asks if the physical limitations of the upstream supply chain, like fab and EUV machine production, will inevitably slow NVIDIA’s growth.
  • 00:06:44 · Overcoming Bottlenecks
    • Jensen asserts that supply chain bottlenecks are temporary (2-3 years) because the industry swarms to solve them, citing the resolution of the CoWoS shortage as an example.
  • 00:08:32 · Competition from Google’s TPU
    • The host challenges Jensen on the competitive threat from Google’s TPU, noting that major models like Gemini and Claude were trained on it.
  • 00:09:35 · Accelerated Computing vs. Tensor Processing
    • Jensen differentiates NVIDIA’s general-purpose accelerated computing platform from specialized ASICs like TPUs, emphasizing CUDA’s flexibility and broad market reach.
  • 00:11:15 · The Debate on China Export Controls
    • The host raises the national security concern of providing advanced AI chips to China, given their potential for developing cyber offensive capabilities.
  • 00:11:35 · Jensen’s Case Against Restrictive Export Controls
    • Jensen argues that restricting chip sales to China is a flawed policy because China has the resources to develop its own alternatives, and it damages the US tech industry’s global leadership.
  • 00:13:38 · What if Deep Learning Never Happened?
    • Jensen states that even without the deep learning revolution, NVIDIA would still be a very large company focused on accelerated computing for science and engineering.
  • 00:14:44 · Why Not Build a Cloud?
    • Jensen explains NVIDIA’s philosophy of ‘doing as much as needed, but as little as possible,’ preferring to enable its cloud partners rather than competing with them.
  • 00:16:24 · The Anthropic Anomaly
    • Jensen addresses why Anthropic uses non-NVIDIA hardware, explaining it was a result of early-stage investment deals that NVIDIA was not in a position to make at the time.
  • 00:18:35 · The Importance of Not Picking Winners
    • Jensen recounts the early days of the graphics industry to illustrate his philosophy of supporting the entire ecosystem rather than picking winners, which he believes is crucial for long-term success.
  • 00:21:13 · GPU Allocation and Pricing Strategy
    • Jensen denies that NVIDIA allocates chips based on favoritism or auctions, stating their process is based on forecasting and first-come, first-served, with stable pricing.
  • 00:22:31 · The Power of Computer Science over Moore’s Law
    • Jensen emphasizes that architectural and algorithmic improvements (computer science) provide far greater performance gains than Moore’s Law alone, citing the 50x leap from Hopper to Blackwell.
  • 00:25:25 · The Risk of Ceding the China Market
    • Jensen passionately argues that conceding the China market would be a disservice to US national security and tech leadership, as it would allow a competing tech stack to become a global standard.
  • 00:28:31 · The Danger of Fear-Mongering about AI
    • Jensen warns against extreme rhetoric that scares people away from AI and software engineering, arguing it’s a disservice to the country and misunderstands the nature of the technology.
  • 00:30:41 · The Case for Openness and Dialogue
    • Jensen advocates for dialogue with China on AI safety and for supporting the open-source ecosystem, which is largely built on the American tech stack, to maintain US leadership.
  • 00:35:05 · The Fallacy of a Compute Deficit in China
    • Jensen refutes the idea that China lacks compute, pointing to their vast energy resources and mainstream chip manufacturing capacity, which allows them to scale out with older but still capable technology.

Specific Prices (8)

Timestamp Item Value Context
00:04:32 NVIDIA Purchase Commitments (in latest filings) almost $100 billion The host mentions NVIDIA’s large purchase commitments for foundries, memory, and packaging.
00:04:42 NVIDIA Purchase Commitments (reported by SemiAnalysis) $250 billion The host cites a report on NVIDIA’s future purchase commitments.
00:19:44 Investment in OpenAI $30 billion The host mentions the scale of investment NVIDIA has made in OpenAI.
00:19:45 Investment in Anthropic $10 billion The host mentions the scale of investment NVIDIA has made in Anthropic.
00:40:14 VC Investment in an AI Lab $5-10 billion Jensen states that a VC would not have put in $5-10 billion into an AI lab like Anthropic in its early days.
00:56:35 AI Factory Compute Purchase $1 billion Jensen says customers can buy a billion dollars worth of AI factory compute from NVIDIA.
00:56:40 AI Factory Compute Purchase $100 million Jensen says customers can buy a hundred million dollars worth of AI factory compute from NVIDIA.
00:56:51 AI Factory Purchase $100 billion Jensen says a customer could order a $100 billion AI factory from NVIDIA.

Memory Facts (6)

  • [00:00:18] NVIDIA’s manufacturing process involves packaging logic dies with HBM (High Bandwidth Memory).
    • HBM
  • [00:00:20] HBM is sourced from companies like SK Hynix, Micron, and Samsung.
    • HBM
  • [00:08:43] Huawei uses HBM2 memory in their chips.
    • HBM2
  • [00:08:51] Memory bandwidth is a critical bottleneck in training and inference.
    • Memory Bandwidth
  • [00:08:56] The difference in memory bandwidth between HBM2 and the latest HBM can be an order of magnitude.
    • Order of magnitude
  • [00:19:42] NVIDIA has partnered with Micron on LPDDR and HBM memories.
    • LPDDR, HBM

Bottleneck Claims (5)

  • [00:06:44] Supply chain bottlenecks are not permanent; they typically get resolved within 2-3 years because the industry swarms to solve them.
    • Evidence: The CoWoS packaging shortage was a major concern, but the industry doubled capacity multiple times, and it’s no longer a primary bottleneck.
  • [00:07:16] At any given moment, the bottleneck could be something as mundane as the number of available plumbers or electricians to build data centers.
    • Evidence: Anecdotal experience from data center construction.
  • [00:13:50] Energy policy is a significant bottleneck that prevents the growth of new manufacturing industries, including AI factories.
    • Evidence: You cannot create a new industry without energy, and building energy infrastructure takes a long time.
  • [00:14:46] The bottleneck for Chinese AI researchers is compute.
    • Evidence: Quotes from Chinese AI company founders stating they are bottlenecked on compute.
  • [00:35:09] The US is scarce on energy, which is why NVIDIA focuses on creating extremely energy-efficient architectures (high throughput per watt).
    • Evidence: This is presented as a strategic driver for NVIDIA’s architecture design.

Predictions (2)

  • [00:03:22, Unspecified, but implied near-to-mid term.] The number of AI agents and tool users will grow exponentially, causing the market for software tools to skyrocket.
  • [00:30:14, A few years] In a few years, when the US wants to export its technology to emerging markets, it will find a competing Chinese tech stack has become the standard if the US cedes that market now.

Key Technologies (10)

  • GDSII: A database file format used to transfer integrated circuit layout data, which NVIDIA sends to TSMC to begin chip manufacturing.
  • HBM (High Bandwidth Memory): A high-performance RAM interface for 3D-stacked memory used in high-performance graphics accelerators and network devices.
  • CoWoS (Chip-on-Wafer-on-Substrate): An advanced 2.5D packaging technology from TSMC that integrates multiple chips (like a GPU and HBM) onto a single interposer.
  • CUDA: NVIDIA’s parallel computing platform and programming model that allows software to use a GPU for general purpose processing.
  • TPU (Tensor Processing Unit): Google’s custom-built ASIC (Application-Specific Integrated Circuit) designed specifically for neural network machine learning.
  • Silicon Photonics: Technology that uses silicon as an optical medium to move data at high speeds between computer chips and other components.
  • NVLink: A high-speed, direct GPU-to-GPU interconnect developed by NVIDIA.
  • Spectrum-X: NVIDIA’s Ethernet networking platform designed for AI workloads.
  • EUV (Extreme Ultraviolet Lithography): An advanced semiconductor manufacturing technology used to create the most modern, smallest-nanometer chips.
  • MoE (Mixture of Experts): A neural network architecture where multiple ‘expert’ networks are used, and a gating network determines which expert to consult for a given input, improving efficiency.

Companies Mentioned (28)

NVIDIA · TSMC · SK Hynix · Micron · Samsung · Cadence · Synopsys · SemiAnalysis · CoreWeave · Google · Amazon (AWS) · Microsoft (Azure) · Oracle (OCI) · xAI · Eli Lilly · Trainium (Amazon) · Lumentum · Coherent · Mellanox · Anthropic · OpenAI · Groq · Crusoe · Huawei · DeepSeek · Jane Street · Nscale · Nebius

Notable Quotes (8)

In the end, something has to transform electrons to tokens. — Jensen Huang @ 00:00:33

The input is electron, the output is tokens. In the middle, NVIDIA. And our job is to do as much as necessary, as little as possible, to enable that transformation to be done at incredible capabilities. — Jensen Huang @ 00:01:48

I don’t think you’re talking to somebody who woke up a loser. And that loser attitude, that losing premise makes no sense to me. — Jensen Huang @ 00:20:27

Moore’s Law is dead. — Jensen Huang @ 00:22:32

Why is it that your policy, your philosophy, leads to United States giving up a vast part of the world’s market? It is a disservice to our country. It is a disservice to our national security. It is a disservice to our technology leadership, all for the benefit of one company. It makes no sense to me. — Jensen Huang @ 00:25:25

If we scare this country into thinking that AI is somehow a nuclear bomb, so that everybody hates AI and everybody’s afraid of AI, I don’t know how you’re helping the United States. You’re doing a disservice. — Jensen Huang @ 00:28:31

We should do as much as needed, as little as possible. — Jensen Huang @ 00:44:18

You can count on us every single year. — Jensen Huang @ 00:56:31

Key Topics

NVIDIA's business strategy and moat · AI hardware supply chain and bottlenecks · Competition in the AI chip market (NVIDIA vs. ASICs) · The role of software and programmability (CUDA) in AI · US-China technology competition and export controls · The economics of AI compute (TCO, performance/watt) · The future of the AI industry stack · AI safety and national security implications

Takeaways

  • Jensen Huang views NVIDIA’s core value not just in making chips, but in enabling the entire, complex transformation of ‘electrons to tokens.’ He argues this full-stack capability, from architecture to software (CUDA) to systems, is incredibly difficult to commoditize.
  • NVIDIA’s moat is a flywheel: its massive downstream market reach (every cloud, every developer) gives suppliers the confidence to make huge upstream investments in capacity for NVIDIA, which in turn strengthens NVIDIA’s supply and ability to serve the market.
  • Jensen believes that while hardware (Moore’s Law) provides incremental gains, the largest leaps in AI performance (e.g., 50x from Hopper to Blackwell) come from computer science: new algorithms, new models (like MoE), and co-design across the entire stack, which is only possible with a programmable architecture like CUDA.
  • He argues strongly against US export controls on AI chips to China, calling it a ‘losing premise.’ He believes it’s a disservice to US tech leadership because China has the resources to build its own ecosystem anyway, and it cuts off US companies from a massive market, potentially allowing a competing tech stack to become a global standard.
  • NVIDIA’s strategy is to be the foundational platform for the entire AI industry, supporting all players (including competitors’ frameworks) and not picking winners. Their goal is to have the best Total Cost of Ownership (TCO) and performance-per-watt, making them the most economical choice at any scale.