Huang on supply moats

Jensen Huang used a new Dwarkesh Podcast interview to argue that Nvidia’s moat is not one chip, but a supply chain and software stack that rivals cannot copy quickly. (dwarkesh.com) The interview was published April 15, 2026 and ran 1 hour 43 minutes. Its chapters included “Is Nvidia’s biggest moat its grip on scarce supply chains?,” “Will TPUs break Nvidia’s hold on AI compute?,” and “Should we be selling AI chips to China?” (dwarkesh.com, podcasts.apple.com) A supply chain is the line of companies that turn a design into a working server: Taiwan Semiconductor Manufacturing Co. makes logic dies, SK Hynix, Micron, and Samsung make high-bandwidth memory, and Taiwan rack builders assemble systems. Huang said Nvidia has lined up enough of that capacity that “if our next several years are a trillion dollars in scale, we have the supply chain to do it.” (dwarkesh.com) A Tensor Processing Unit is Google’s custom artificial-intelligence chip, built mainly for machine-learning workloads inside Google Cloud. CUDA, short for Compute Unified Device Architecture, is Nvidia’s programming system for using graphics processors as general-purpose computing engines, and Nvidia says its full stack also includes hundreds of libraries, software development kits, and application programming interfaces. (cloud.google.com, docs.cloud.google.com, sec.gov) That distinction sits at the center of Huang’s argument. Google can build strong chips for its own cloud, but Nvidia’s bet is that a broad programming platform matters more when outside developers, model companies, and enterprises all need the same tools to train and run models across many systems. (dwarkesh.com, developer.nvidia.com) Nvidia’s filings show how much of that software layer it has spent years building. In its fiscal 2025 annual report, filed February 26, 2025, the company said more than 5.9 million developers worldwide use CUDA and its other software tools. (sec.gov) Google, for its part, has kept expanding the TPU line rather than retreating from custom silicon. Google Cloud says Cloud TPU gives customers access to its application-specific integrated circuits for training and inference, and this month published documentation for TPU7x, the latest Ironwood generation. (docs.cloud.google.com, docs.cloud.google.com) China was the third leg of the conversation. Nvidia disclosed in an April 9, 2025 filing that the United States government would require a license to export H20 chips to China, Hong Kong, Macau, and other covered destinations, cutting off the most advanced Nvidia artificial-intelligence chip then still available to that market. (sec.gov) Nvidia later said the new H20 rules led to a $4.5 billion charge in the first quarter of fiscal 2026, while H20 sales before the licensing change totaled $4.6 billion. Huang has also said Nvidia’s China market share fell to 50% from 95% over four years as export curbs tightened and local rivals gained ground. (nvidianews.nvidia.com, cnbc.com) The through line in Huang’s case was that artificial intelligence computing is still constrained by scarce parts, hard-to-rewrite software, and geopolitics. In his framing, Nvidia stays hard to replace as long as customers still need all three at once. (dwarkesh.com, sec.gov)