Overview

In this wide-ranging conversation, Nvidia CEO Jensen Huang defends his company's dominance against claims that its moat is merely supply-chain control, argues that TPUs and other ASICs cannot match Nvidia's total cost of ownership, and makes a passionate case for continuing to sell AI chips to China. The episode moves from technical debates about semiconductor bottlenecks to geopolitical strategy, with Huang pushing back against what he sees as alarmist extremes in both the AI safety and export control debates. Throughout, he presents Nvidia not as a chip company but as the architect of a new computing paradigm—accelerated computing—that spans far beyond AI.

0:00The Transformation from Electrons to Tokens

Huang frames Nvidia's core mission as transforming electrons into tokens—making those tokens more valuable over time through artistry, engineering, and invention. He argues this transformation is far from commoditized, precisely because it requires deep integration across hardware, software, and algorithms. Rather than seeing software commoditization as a threat to Nvidia, Huang predicts the opposite: the number of tool users and agents will grow exponentially, driving demand for tools like Excel, PowerPoint, and Synopsys Design Compiler. The reason this hasn't happened yet, he says, is that agents aren't yet good enough at using those tools—but either the tool companies will build the agents themselves, or agents will become capable enough to use existing tools.

4:40Supply Chain as a Moat—and a Philosophy

When asked about Nvidia's nearly $100 billion in purchase commitments and reported $250 billion in future commitments, Huang reframes the narrative. These commitments are not simply about locking up scarce components; they reflect Nvidia's ability to inform, inspire, and align its entire upstream supply chain—from TSMC to memory makers to silicon photonics companies—around a shared vision of AI's scale. Huang explains that suppliers invest in capacity for Nvidia because they see the downstream demand Nvidia has built. He describes his keynotes as partly educational, designed to ensure the entire ecosystem understands "what is coming at us, why it's coming, when it's coming, how big it's going to be." This alignment, he argues, is harder for competitors to replicate than simply signing contracts.

Huang acknowledges that bottlenecks exist—CoWoS packaging was a major one two years ago—but insists none last longer than two or three years. The industry "swarmed" CoWoS, doubling capacity repeatedly, and TSMC now treats advanced packaging as mainstream. He points to Nvidia's work with Lumentum and Coherent on silicon photonics as an example of prefetching bottlenecks years in advance: Nvidia invented new technology, licensed patents openly, and helped build an entire supply chain around TSMC. The harder bottlenecks, he jokes, are plumbers and electricians—and he warns that doomer narratives about AI eliminating jobs could discourage people from becoming software engineers or radiologists, creating real shortages.

16:25TPUs vs. Nvidia: The Programmability Advantage

Huang directly addresses the claim that TPUs power two of the top three models (Claude and Gemini). He argues that Nvidia builds "accelerated computing," not just a tensor processing unit, and that this market is far broader than AI—spanning molecular dynamics, quantum chromodynamics, data processing, fluid dynamics, and particle physics. Because Nvidia's systems are designed to be operated by anyone, they run in every cloud, including Google Cloud and AWS, and can be used by companies like Eli Lilly for drug discovery.

When pressed on whether AI's core workload—repetitive matrix multiplies—might be better served by a simpler, more specialized architecture like a TPU, Huang pushes back hard. Matrix multiplies are important, he says, but not the only part of AI. Programmability enables invention of new algorithms—hybrid SSMs, fused diffusion and autoregressive models, new attention mechanisms—and that algorithmic innovation is what drives 10x and 100x leaps. He notes that Blackwell achieved 50x the energy efficiency of Hopper, far beyond what Moore's Law (roughly 25% per year) could deliver. That leap came from new models, MoEs, parallelism, disaggregation, and co-design across processors, fabric, libraries, and algorithms—all enabled by CUDA's programmability.

25:50CUDA's Real Value: Ecosystem, Install Base, and Versatility

Huang acknowledges that hyperscalers like Google, Amazon, and Anthropic have the resources to write custom kernels and even build their own accelerators. But he argues that CUDA's value extends beyond raw performance. First, the richness of the ecosystem means developers can trust the foundation and focus on their own code. Second, the install base—hundreds of millions of GPUs across every cloud—means that software developed on CUDA runs everywhere. Third, Nvidia's presence in every cloud gives AI companies flexibility: they don't have to bet on a single cloud provider.

Huang pushes back on the idea that these advantages don't matter to Nvidia's largest customers. He reveals that Nvidia assigns an "insane" number of engineers to work directly with AI labs, optimizing their stacks—often yielding 2x, 3x, or 50% speedups. "Nobody knows our architecture better than we do," he says, comparing GPUs to F1 cars that require expertise to push to the limit. He then makes a striking claim: Nvidia has the best performance per total cost of ownership (TCO) in the world, bar none. He challenges competitors to demonstrate their cost advantage on public benchmarks like Dylan Patel's Inference Max or MLPerf, saying "it makes absolutely zero sense on first principles."

Huang also corrects the premise that 60% of Nvidia's revenue comes from hyperscalers' internal use. Most of that business, he says, is for external customers—AWS's Nvidia instances serve external AI companies, not just Amazon. The flywheel, he argues, is that Nvidia's install base and ecosystem attract tens of thousands of AI startups, which in turn make Nvidia the most valuable platform for cloud providers to offer.

36:43Why Anthropic Uses TPUs—and Why It's Not a Trend

When asked why Anthropic, despite Nvidia's advantages, recently announced a multi-gigawatt deal with Broadcom and Google for TPUs, Huang is blunt: "Anthropic is a unique instance and not a trend." He claims that without Anthropic, there would be no TPU growth or Trainium growth—it's "100% Anthropic." He acknowledges that Nvidia missed the opportunity to invest in Anthropic early, when the lab needed massive capital commitments that VCs couldn't provide. "We just weren't in a position to make the multi-billion dollar investment into Anthropic," Huang says, calling it his mistake. But he insists he won't make that mistake again, noting Nvidia's subsequent investments in OpenAI and Anthropic.

Huang also pushes back on the idea that ASICs can undercut Nvidia by offering slightly worse performance at much lower margins. He claims ASIC margins are also "incredibly high"—around 65%—so the savings are minimal. More importantly, he argues that no ASIC team in the world can match Nvidia's cadence: a new architecture every year, with token costs decreasing by an order of magnitude annually. "You can count on us," he says, contrasting Nvidia's reliability with the uncertainty of custom ASIC projects.

43:38Why Nvidia Won't Become a Hyperscaler

Huang explains why Nvidia doesn't simply become a cloud provider itself, despite having the cash to do so. His philosophy is "do as much as needed, as little as possible." Nvidia should do the things that genuinely wouldn't get done otherwise—building NVLink, creating CUDA, developing domain-specific libraries like cuLitho for computational lithography. But the world already has plenty of clouds. Instead, Nvidia invests in "neo clouds" like CoreWeave, Lambda, and Crusoe, helping them exist and thrive. He emphasizes that Nvidia does not pick winners: it invests in all major foundation model companies and supports all cloud providers. This humility comes from Nvidia's own history—it was once the 3D graphics company everyone counted out, with an architecture that was "precisely wrong."

Huang also clarifies Nvidia's allocation policy during GPU shortages. Contrary to rumors, Nvidia does not auction chips to the highest bidder. Allocation is first-in, first-out, adjusted only for data center readiness. He denies the famous story that Elon Musk and Larry Ellison begged him for GPUs at dinner—"that never happened." Nvidia sets a price and sticks with it, even when demand surges. This dependability, he argues, is part of what makes Nvidia a trustworthy foundation for the industry.

57:36The China Debate: Selling Chips vs. National Security

The most contentious section of the conversation is Huang's defense of selling AI chips to China. When asked about the threat of Chinese labs training models with cyber-offensive capabilities (like Anthropic's Claude Mythos, which found zero-day vulnerabilities across every major OS), Huang makes several arguments. First, he claims China already has abundant compute—60% of the world's mainstream chips, enormous energy reserves, and 50% of the world's AI researchers. The idea that export controls can meaningfully limit their AI progress is "completely nonsense." Second, he argues that cutting off Nvidia chips will only accelerate China's domestic chip industry and force its AI ecosystem to optimize for Chinese hardware—a "horrible outcome" for the United States, because future AI models would run best on non-American tech stacks.

Huang repeatedly challenges the premise that AI is like enriched uranium. "It's a lousy analogy," he says. He argues that the real path to safety is dialogue between American and Chinese researchers, not isolation. He warns that current policies risk creating two separate AI ecosystems—an open-source one running on Chinese tech and a closed one on American tech—which would ultimately harm US technology leadership. When the host presses that any marginal compute helps China train more powerful models, Huang counters that the United States should focus on staying ahead through innovation, not on denying compute to a country that already has plenty.

The exchange becomes heated when Huang accuses the host of a "loser mindset" and of advocating policies that would cause the US to "concede the second largest market in the world for no good reason at all." He points to the US telecommunications industry's loss of global leadership as a cautionary tale. The host, in turn, argues that the US should maintain its compute advantage to ensure American labs reach dangerous capabilities first and can prepare defenses. Huang dismisses this as "absolutes" and "extremes," insisting that the world is more nuanced.

1:35:06One Architecture, One Roadmap

On why Nvidia doesn't pursue multiple chip architectures in parallel—wafer-scale, Dojo-style, or non-CUDA designs—Huang says simply: "We could. It's just that we don't have a better idea." He claims Nvidia simulates all these approaches and finds them "provably worse." The only recent addition is Grok, a new inference accelerator folded into the CUDA ecosystem, driven by the emergence of a premium token market where customers will pay more for faster response times. Otherwise, Huang prefers to put all resources behind a single architecture, advancing it every year.

Conclusion

This episode matters because it captures Jensen Huang at his most expansive and combative, defending Nvidia's position against three distinct challenges: supply chain constraints, architectural competition from TPUs, and geopolitical pressure on China sales. His core argument—that Nvidia's real moat is not hardware but a complete computing ecosystem built over two decades—is both a defense of his company and a vision of how AI should develop. The conversation reveals Huang as a strategic thinker who sees Nvidia as the indispensable infrastructure layer of the AI revolution, and who believes that the United States' best path to leadership is through innovation and engagement, not isolation and control. Whether or not one agrees with his conclusions, the episode provides an unusually direct look at how the CEO of the world's most important AI company thinks about competition, geopolitics, and the future of computing.

Key takeaways

• Huang argues Nvidia's moat is not just supply chain control but its ability to align the entire upstream ecosystem around a shared vision of AI's scale, making suppliers willing to invest in capacity for Nvidia specifically.
• He claims Nvidia's programmability (CUDA) enables algorithmic innovation that delivers 10x-50x leaps, far beyond what Moore's Law or specialized ASICs can achieve.
• Huang challenges competitors to demonstrate better TCO on public benchmarks, asserting that "nobody can demonstrate" a platform with better performance per dollar.
• He acknowledges Nvidia missed early investment opportunities in Anthropic and OpenAI due to its inability to make multi-billion dollar commitments, but says it won't make that mistake again.
• Huang defends selling chips to China on the grounds that China already has abundant compute and AI talent, and that export controls will only accelerate its domestic ecosystem—a worse outcome for US technology leadership.
• He warns that current policies risk creating two separate AI ecosystems, with open-source models optimized for Chinese hardware, ultimately harming American influence.
• Huang insists Nvidia does not auction chips to the highest bidder, allocating them first-in, first-out at a fixed price, and values dependability over short-term profit maximization.
• He predicts that even without deep learning, Nvidia would be "very, very large" because general-purpose computing has largely run its course and domain-specific acceleration is the only path forward.