NVIDIA CEO Jensen Huang has defined his company in a single statement: input is electrons, output is tokens, and NVIDIA operates in between. On April 15, Huang participated in an in-depth podcast interview with host Dwarkesh Patel, covering NVIDIA's supply chain advantages, competition from TPUs, and why the company does not operate its own cloud services.
NVIDIA: The "Electron-to-Token" Converter The interview began with a pointed question: since NVIDIA essentially designs software while chips are manufactured by TSMC, memory comes from SK Hynix, Micron, and Samsung, and packaging is handled by Taiwanese ODMs—if software becomes commoditized, could NVIDIA face the same fate? Huang offered his most concise definition of NVIDIA: Ultimately, something must convert electrons into tokens. The process of converting electrons into tokens and making those tokens increasingly valuable over time is very difficult to fully commoditize. He elaborated: "Input is electrons, output is tokens, and NVIDIA is in the middle. Our job is to make this conversion as efficient as possible."
Regarding supply chain dynamics, industry attention has focused on NVIDIA's purchase commitments—approaching $100 billion according to recent financial reports, with industry analysis firm SemiAnalysis predicting this could reach $250 billion. Huang explained the rationale: If the next few years represent a trillion-dollar opportunity, our supply chain can support it. He clarified this capability stems not merely from contracts but from continuously informing, incentivizing, and aligning upstream CEOs—helping them understand the scale and direction of the AI industry so they invest to meet NVIDIA's needs. CoWoS packaging serves as a prime example: two years ago it was the industry's tightest bottleneck, but after NVIDIA drove "several consecutive doublings" in production capacity, it has largely ceased to be a discussion topic. He判断 that no supply chain bottleneck should persist beyond two to three years: "EUV machines, logic capacity, packaging—these things aren't difficult to replicate; they just need demand signals." He identified the real long-term constraint as downstream energy policy: "You can't build an industry without power. That's what takes time."
TPU Competition: No One Dares to Run Benchmark Tests Hyperscalers contribute approximately 60% of NVIDIA's revenue, but major customers like Google, Amazon, and even OpenAI are increasing investments in self-developed TPUs or other ASIC chips, directly challenging NVIDIA's competitive position. Huang's response addressed two levels. First, a qualitative distinction: NVIDIA focuses on "accelerated computing" rather than specialized tensor processing units. Accelerated computing covers nearly all scientific computing fields including molecular dynamics, fluid mechanics, data processing, and quantum computing—far broader than AI alone. We are the only company accelerating all types of applications. Second, Huang stated that no platform globally offers better performance and total cost of ownership (TCO) than NVIDIA, asserting that Google's TPU and Amazon's Trainium cannot match it. He dismissed Trainium's claimed 40% cost advantage, noting that while NVIDIA achieves 70% gross margins, ASIC chips hover around 65%, making switching unlikely to yield significant savings. NVIDIA's compute stack offers the best TCO globally, without exception. No company has demonstrated that any other platform provides better AI datacenter TCO today. Dylan's InferenceMAX benchmark is publicly available, yet TPU and Trainium avoid it.
Regarding Anthropic's heavy TPU usage, Huang responded directly: Anthropic is a special case, not a trend. Without Anthropic, where would TPU growth come from? 100% from Anthropic. Without Anthropic, where would Trainium growth come from? Also 100% from Anthropic. He acknowledged that failing to strategically invest in Anthropic earlier was a misjudgment: At the time, Google and Amazon AWS committed substantial funds, so Anthropic used their computing power. I didn't fully realize that venture capital could never provide $5 billion or $10 billion to an AI lab. That was my mistake. But I won't repeat it. Huang has since made major investments in both OpenAI and Anthropic, reportedly amounting to $30 billion and $10 billion respectively.
Why Not Operate Its Own Cloud? "Do as Little as Possible" is the Philosophy Despite holding massive cash reserves, NVIDIA has recently provided funding and computing power to startup cloud providers like CoreWeave. Market speculation questions whether NVIDIA might bypass customers to become a hyperscale cloud provider itself, collecting compute rental fees. Huang's response touched on company philosophy: We should do what is necessary, and as little as possible. The things we do—if we didn't do them, I truly believe they wouldn't happen. But cloud services? If I don't do it, someone else will. He cited CoreWeave, Nscale, and Nebius as examples of "new clouds" that wouldn't exist without NVIDIA's early investment and support. However, NVIDIA's involvement aims to "make the ecosystem flourish" rather than transitioning into financial leasing or cloud operations. He also expressed a clear stance on "not picking winners": When I invest in one, I invest in all. The reason stems from humility—NVIDIA was once considered the least likely to survive among 60 3D graphics companies.
Maintaining the "Shovel Seller" Position: GPU Allocation Rejects "Highest Bidder" Amid extreme supply-demand imbalance, how does NVIDIA allocate scarce GPUs? Huang explicitly denied market rumors of "highest bidder" practices: We would never do that; it's poor business practice. You set the price, and people decide whether to buy. He explained NVIDIA's allocation logic: prioritize customers' production forecasts and purchase orders, then consider data center readiness, ultimately following a first-come-first-served principle. I prefer to be the industry's reliable foundation. If you place a $100 billion AI factory order, no problem—only we can provide that certainty globally.
Not Abandoning the World's Second-Largest Market Addressing current chip export controls, Huang expressed pragmatic views from commercial and technical perspectives. He noted that computing power is just the foundational input for the AI industry; when constrained, competitors can compensate for hardware generation gaps by stacking more energy, using more previous-generation chips, and optimizing algorithms. He emphasized that China is not short on chips. They possess world-class computer scientists, and a significant proportion of AI researchers in labs globally are Chinese. They account for about 50% of the world's AI researchers. They are competitors; we want the U.S. to win, but I believe dialogue and research exchange may be the safest approach. Abandoning the entire market won't ensure long-term U.S. leadership in the technology race at the chip and compute stack level. That is a fact.
CUDA Ecosystem and Flywheel Effect NVIDIA's core competitive barrier is its mature developer ecosystem, with hundreds of millions of GPUs deployed globally across diverse application scenarios. The industry's flywheel effect is clear: the largest installed base, programmable architecture, rich ecosystem, and global abundance of AI companies create strong momentum. Leading price-performance, energy efficiency, and customer base further propel this flywheel.
Architectural Advantages (vs. ASIC/TPU) Huang stated that traditional Moore's Law delivers about 25% annual performance growth, but achieving 10x to 100x leaps requires dual innovation in algorithms and compute architecture. The Blackwell architecture delivers 50x better energy efficiency than Hopper—an achievement impossible through transistor advances alone, relying instead on architectural optimization and computer science innovation. The architecture supports full-stack programmability and co-design enabled by NVLink and Spectrum-X technologies, impossible without the CUDA ecosystem. Regarding product roadmap and release cadence, Huang revealed that NVIDIA maintains annual stable iterations—from Vera Rubin to Vera Rubin Ultra to Feynman—with new products each year. He emphasized that NVIDIA is the only company globally capable of fulfilling AI compute orders ranging from $10 million to $100 billion.
Comments