Dwarkesh Patel Podcast - Jensen Huang – TPU competition, why we should sell chips to China, & Nvidia’s supply chain moat
What this covers
Jensen Huang speaks with Dwarkesh Patel about Nvidia's competitive position, the strategic logic of selling chips to China, and why the company's durability rests on engineering layers that have nothing to do with the race for smaller transistors. The conversation moves through three interlocking claims: that Moore's Law now delivers only ~25% annual improvement while algorithmic advances can yield 10x gains; that Nvidia's moat is its full-stack ecosystem—CUDA programmability, install base, supply-chain reach, and yearly architectural cadence—rather than any fragile scarcity or lithography advantage; and that export controls on chips to China are strategically self-defeating because they would force Chinese researchers (roughly half the world's AI talent) to build optimized alternatives that could become the global standard, repeating past US policy errors in telecommunications.
Huang ranges across why Algorithms over hardware means China's abundance of researchers matters more than its 7nm chip constraints; how the AI five-layer cake shows energy-abundant China can gang older chips together where energy-scarce America must optimize architecture; why ASIC adoption is not a broad trend but a single case (Anthropic); and how Nvidia's philosophy of Do as much as needed, as little as possible — building only the work that wouldn't happen otherwise — explains why it avoids becoming a cloud or hyperscaler. He also defends AI's impact on the workforce by distinguishing Job vs task distinction, argues that software won't be commoditized because AI agents will multiply tool usage, and contends that Ecosystem lock-in in computing runs far deeper than smartphone or auto analogies suggest. The stance throughout is that Nvidia's advantage is architectural, organizational, and cumulative rather than dependent on denying competitors access to advanced chips.
Huang argues that Nvidia's durable advantage is not chip lithography but its full-stack ecosystem (CUDA programmability, install base, supply-chain reach, and yearly cadence), and that conceding the Chinese AI market via export controls would hand global standards and the chip layer to rivals for no strategic gain.
- Most AI performance gains come from algorithmic/architectural advances enabled by CUDA programmability, not raw transistor scaling (Moore's Law is ~25%/yr).
- China already has abundant energy, chips, and AI researchers, so denying Nvidia chips does not deny China compute—it only forces optimization onto a non-American stack.
- Nvidia's moat is the flywheel of perf/TCO, install base, and being in every cloud—not a fragile bidding or scarcity mechanism.
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Most of the advances in AI come from algorithm and computer-science improvements (MoE, attention mechanisms) rather than raw hardware, since Moore's Law gives only ~25%/year while computer science can yield 10x; therefore China's large army of AI researchers—forced by compute limits to invent smarter algorithms—is its fundamental advantage, and DeepSeek launching first on Huawei would be a horrible outcome for the US.
“We have got to acknowledge that most of the advances in AI came out of algorithm advances, not just the raw hardware.”
Conceding the Chinese market via export controls is a strategic error: China is the largest contributor to open source software and open models, those models currently run best on the American (Nvidia) stack, and if forced out the US risks Chinese-optimized open models diffusing worldwide and becoming the global standard—repeating the mistake that policied the US out of the telecommunications industry.
“It would be extremely foolish to create two ecosystems: the open source ecosystem, and it only runs on a foreign tech stack, and a closed ecosystem that runs on the American tech stack.”
Because Moore's Law now yields only ~25% per year, the only way to achieve 10x-100x leaps is to change the algorithm and how it's computed yearly; Nvidia's general programmability via CUDA enables new architectures (hybrid SSMs, fused diffusion/autoregressive models, MoEs), which is why Blackwell achieved roughly 50x Hopper's efficiency despite only ~75% transistor improvement.
“TPUs, like anything else, are impacted byMoore’s Law, which we know is increasing by about 25% per year. The only way to really get 10x or 100x leaps is to fundamentally change the algorithm and how it’s computed every single year.”
Performance gains come far more from architecture and computer science than from lithography: Blackwell is 50x Hopper while the transistors improved only ~75% over three years, so a 5nm-vs-7nm gap is not a 10x difference—meaning China being stuck on 7nm does not preclude competitive systems when networking, architecture, and energy are combined.
“Is Blackwell 50 times more advanced lithography than Hopper? ... call it 75%. It was three years apart, 75%. Blackwell is 50 times Hopper.”
In the five-layer AI cake with energy at the base, an abundance of energy compensates for weaker chips and vice versa: the energy-scarce US must advance architecture for maximum throughput-per-watt, while energy-abundant China can use older 7nm (Hopper-class) chips and simply use more of them.
“When you have an abundance of energy, it makes up for chips. If you have an abundance of chips, it makes up for energy.”
Export controls cannot deny China meaningful AI compute, because China has abundant energy, manufactures the majority of mainstream chips, holds roughly 50% of the world's AI researchers, and can simply gang together more 7nm chips (essentially Hopper-class) since energy is effectively free for them—so they have already passed any compute threshold relevant to cyber-offensive concerns.
“Why can’t they just put 4x, 10x, as many chips together because energy’s free? They have so much energy.”
Automation fears confuse jobs with tasks: a radiologist's job is patient care while reading a scan is only a task, so even free, superhuman computer vision doesn't eliminate the role—misunderstanding this drives people away from training for needed careers and produces shortages.
“we misunderstand the difference between a job and a task. The job of a radiologist is patient care. The task is to read a scan.”
Telling people AI will eliminate careers (software engineering, radiology) discourages people from entering those fields and risks real shortages, just as a decade ago warnings that radiology would be automated coincided with today's shortage of radiologists.
“Some of the doomers were telling people, “Whatever you do, don’t be a radiologist.” ... Guess what we’re short of? Radiologists.”
Because most software companies are tool makers (Excel, PowerPoint, Cadence, Synopsys), and AI agents will use those tools, the number of agents and tool instances will grow exponentially once agents get good enough at tool use—the opposite of the view that AI commoditizes software.
“I think the number ofagentsis going to grow exponentially, and the number of tool users is going to grow exponentially. It’s very likely that the number of instances of all these tools is going to skyrocket.”
No supply-chain bottleneck (CoWoS, logic fabs, HBM, EUV machines) lasts longer than two to three years once a clear demand signal exists, because once you can build one you can build ten and then a million; the only hard, durable constraints are downstream—energy and skilled trades like plumbers and electricians.
“My point is that none of the bottlenecks last longer than a couple of years, two, three years, none of them.”
CUDA's enduring value comes from three reinforcing factors: a rich ecosystem where every framework works and bugs are more likely in your code than the stack; the largest install base of hundreds of millions of GPUs across every cloud and device; and ubiquity across all clouds and on-prem—so software written for CUDA runs everywhere, making it invaluable even though some customers write their own kernels.
“the single most important thing you want is an install base. You want the software you write to run on a whole bunch of other computers.”
Computing is not like cars where customers switch brands freely; computing ecosystems (x86, ARM, CUDA) are sticky because replacing them costs enormous time and energy, so the analogy that China will simply build its own chips like it did with EVs and smartphones underestimates the lock-in of Nvidia's developer ecosystem, of which ~50% of AI developers are in China.
“We’re not a car. We are not a car. ... Computing is not like that. There’s a reason why thex86deal exists. There’s a reason whyARMis so sticky.”
The supposed savings from building an ASIC instead of buying Nvidia are illusory because ASIC margins (e.g., to a vendor like Broadcom) are themselves very high—around 65% versus Nvidia's ~70%—so you still pay someone a large margin.
“Nvidia’s margin is 70%, let’s say. But ASIC margins are 65%. What are you really saving?”
Nvidia's suppliers make large upstream investments specifically for Nvidia rather than competitors because they trust Nvidia has the downstream demand and reach to buy their supply and sell it through; without that demand reach, no one builds a supply chain for an architecture whose business churns are low.
“Why are they willing to make the investments for me and not someone else? The reason for that is because they know that I have the capacity to buy their supply and sell it through my downstream.”
Nvidia builds accelerated computing—usable for molecular dynamics, quantum chromodynamics, fluid dynamics, data processing, and AI—rather than a narrow tensor processing unit, giving it far greater market reach than any TPU or ASIC can achieve.
“What Nvidia built is accelerated computing, not a tensor processing unit. Accelerated computing is used for all kinds of things: molecular dynamics, quantum chromodynamics, data processing...”
ASIC adoption is not a broad trend but a single instance: essentially 100% of TPU growth and 100% of Trainium growth is attributable to Anthropic, so there is not an abundance of ASIC opportunities—there is only one Anthropic.
“Anthropic is a unique instance, not a trend. Without Anthropic, why would there be any TPU growth at all? It’s 100% Anthropic.”
Although the top five hyperscalers are 60% of Nvidia's revenue, most of that business serves external customers (e.g., most Nvidia capacity in AWS, Azure, and all of OCI is for external customers, not internal use), because Nvidia's reach brings the cloud providers a large base of AI companies built on Nvidia.
“you say 60% of our customers are the top five, but most of that business is external. For example, most of Nvidia in AWS is for external customers, not internal use.”
Because a data center's revenue is proportional to the tokens it produces, and energy is the binding constraint, the architecture with the highest tokens-per-watt maximizes revenue for a fixed-gigawatt facility—and Nvidia claims to be the highest tokens-per-watt architecture in the world.
“that one gigawatt data center better deliver the maximum amount of revenues and number of tokens, which directly translates to revenues. ... We are the highest tokens per watt architecture in the world.”
Nvidia's guiding philosophy is to do only the work that wouldn't get done otherwise (building the stack, CUDA, NVLink, domain-specific libraries) wholeheartedly, while doing as little as possible elsewhere—which is why it doesn't become a cloud or hyperscaler, since clouds would exist without Nvidia but its core platform work would not.
“We should do as much as needed, as little as possible. What that means is, the work that we do with building our computing platform, if we don’t do it, I genuinely believe it doesn’t get done.”
Nvidia deliberately avoids picking winners among foundation labs and invests in all of them, partly because it's not Nvidia's job and partly from humility—of the original 60 3D graphics companies, Nvidia was the sole survivor despite starting with a 'precisely wrong' architecture that experts would have counted out.
“when Nvidia first started, there were 60 3D graphics companies. We are the only one that survived. ... Nvidia would be at the top of that list not to make it.”
Nvidia allocates scarce GPUs by forecasting with customers and then first-in-first-out on purchase orders (adjusted only for whether a customer's data center and components are ready), and deliberately never raises prices when demand spikes, because dependable pricing makes Nvidia a trustworthy foundation for the industry.
“the prioritization is first in, first out. You’ve got to place a PO.”
Because no one knows Nvidia's architecture better than Nvidia, its engineers routinely get partner AI labs another 2x-3x out of their stack or a particular kernel, which—across an installed fleet of Hoppers and Blackwells—directly doubles revenue, so Nvidia's expertise remains needed even for sophisticated customers.
“It’s not unusual that by the time we’re done optimizing their stack or optimizing a particular kernel, their model sped up by 3x, 2x, 50%.”
Even without deep learning, Nvidia would be very large because general-purpose computing has largely exhausted its scaling, and domain-specific acceleration via GPU+CUDA speeds applications 100x-200x across graphics, molecular dynamics, seismic processing, fluid dynamics, and structured data—the company's mission was accelerated computing broadly, with AI as one beneficiary.
“Even if AI doesn’t exist today, Nvidia would be very, very large. The reason for that is fairly fundamental, which is that the ability for general purpose computing to continue to scale has largely run its course.”
Nvidia's distinctive promise is a reliable yearly architecture cadence (Vera Rubin, then Vera Rubin Ultra, then Feynman) with token cost falling by roughly an order of magnitude per year, plus the ability to fulfill orders from a single graphics card up to $100 billion AI factories—a bettable consistency no other ASIC team or foundry except TSMC can match.
“Every single year you can count on us. You’re going to have to go find another ASIC team in the world... where you can say, “I can bet the farm... that you will be here for me every single year.””
Huang's mistake was not internalizing that foundation labs like Anthropic and OpenAI needed multi-billion-dollar investments from suppliers (which VCs wouldn't provide) to use compute; Google and AWS could make those investments and Nvidia couldn't at the time, which is why Anthropic went elsewhere—a mistake he won't repeat now that Nvidia can invest in OpenAI and Anthropic.
“a VC would never put in $5-10 billion of investment into an AI lab with the hopes of it turning out to be Anthropic. So that was my miss.”
Nvidia doesn't run multiple parallel chip architectures (Cerebras-style wafer-scale, Dojo-style packages, non-CUDA designs) because it simulates them and finds them provably worse; it would only add accelerators if the workload shape changed—as it recently did by folding Groq in to serve a newly emerged market for high-ASP, faster-response 'premium' tokens at lower throughput.
“We could do all of those things. It’s just not better. We simulate it all in our simulator, proveably worse. So we wouldn’t do it.”
Nvidia's computing stack delivers the best performance per total cost of ownership of any platform, evidenced by competitors (TPU, Trainium) declining to publish results on open benchmarks like InferenceMAX or MLPerf despite claiming cost advantages.
“Nvidia’s computing stack is the best performance perTCOin the world, bar none. Nobody can demonstrate to me that any single platform in the world today has a better performance-TCO ratio.”
The transformation of electrons into valuable tokens involves so much artistry, engineering, science, and ongoing invention—and is so far from being fully understood—that even as it becomes more efficient, the core work Nvidia does in the middle will not be commoditized.
“In the end, something has to transform electrons to tokens. The transformation of electrons to tokens and making those tokens more valuable over time is hard to completely commoditize.”
Nvidia won't make new chips on older nodes like N7 even if leading-edge capacity is constrained, because each generation's value is in packaging, stacking, numerics, and system architecture—not just transistor scale—so re-engineering for an old node would be R&D no one could afford; only if capacity were permanently capped would Nvidia revert.
“When you run out of capacity, to easily go back to another node… That’s a level of R&D that no one could afford. We could afford to lean forward. I don’t think we could afford to go back.”
The valuations of many software companies have crashed because investors expect AI to commoditize software.
“We’ve seen thevaluations of a bunch of software companies crashbecause people are expecting AI to commoditize software.”