Jensen Huang – TPU competition, why we should sell chips to China, & Nvidia’s supply chain moat
In a Nutshell
Jensen Huang defends Nvidia's irreplaceable moat in AI via massive supply chain commitments ($100B+), CUDA ecosystem superiority over TPUs/ASICS, and extreme architecture leaps (e.g., Blackwell 50x Hopper), enabling accelerated computing beyond just matrix multiplies for diverse workloads. He argues for selling compliant chips to China to lock 50% of global AI developers into the American tech stack, fostering open-source contributions and US leadership across AI's five layers, rather than conceding the world's second-largest market to Huawei/SMIC amid their abundant energy and manufacturing. Export controls risk bifurcating ecosystems, accelerating China's domestic chips, while US compute/process node leads (e.g., 1.6nm vs. 7nm) ensure America stays ahead if it competes globally.
These notes were generated by AI and may contain inaccuracies.
Valuations of software companies have crashed because people expect AI to commoditize software. Naive view: Nvidia sends GDS2 file to TSMC, which builds logic dies and switches, packages with HBM from SK Hynix, Micron, Samsung, then to ODM in Taiwan for rack assembly. Nvidia makes software that others manufacture; if software commoditizes, does Nvidia?
Something must transform electrons to tokens, making tokens more valuable over time—hard to commoditize. Transformation from electrons to tokens is an incredible journey, like making one molecule or token more valuable than another. Requires artistry, engineering, science, invention—we're watching it happen. Manufacturing and science involved is far from deeply understood; journey far from over. Will become more efficient.
Input is electrons, output is tokens. In the middle is Nvidia. Our job is to do as much as necessary and as little as possible to enable that transformation at incredible capabilities.
"As little as possible" means partnering for what isn't needed. Nvidia has largest ecosystem of partners: supply chain upstream/downstream, computer companies, application developers, model makers. AI is a five-layer cake with ecosystems across all layers. Do as little as possible, but necessary part is insanely hard—not commoditizable.
Enterprise software companies are mostly tool makers (e.g., Excel, PowerPoint, Cadence, Synopsys). Opposite of commoditization: number of agents and tool users will grow exponentially, skyrocketing tool instances (e.g., Synopsys Design Compiler, floor planners, layout tools, design rule checkers). Today limited by engineers; tomorrow engineers supported by agents exploring design space unprecedentedly. Tool use will skyrocket software companies. Hasn't happened yet because agents not good enough at using tools. Companies will build agents or agents will improve—combination of both.
Latest filings: almost $100 billion in purchase commitments with foundries, memory, packaging. SemiAnalysis reports $250 billion. Interpretation: Nvidia's moat is locking up scarce components (memory, logic)—competitors can't build accelerators.
One of hard things to replicate: enormous upstream commitments, explicit (mentioned) and implicit. Investments by supply chain because Nvidia informs, inspires, aligns CEOs upstream on industry size/opportunity. They invest for Nvidia due to Nvidia's downstream capacity to buy and sell through it. GTC marvels by scale—full 360 degrees of AI universe together: downstream sees upstream, upstream sees downstream, all see AI advances and meet AI natives/startups.
Spends time informing supply chain, partners, ecosystem about opportunity. Keynotes include educational part to ensure supply chain understands what's coming, why, when, how big, reasoning systematically.
Moat enables building for future: next several years at trillion dollars scale with supply chain ready. Velocity/business churns sustain scale—nobody builds supply chain for low-churn architecture. Downstream demand so great, visible/heard.
Revenue 2x year-over-year, flops >3x. Biggest TSMC N3 customer, big on N2. AI: 60% N3 this year, 86% next per SemiAnalysis. How double if majority? Growth rate slow due to upstream?
Instantaneous demand > supply upstream/downstream (e.g., limited by plumbers—invited to next GTC). Good: demand > supply. Industry swarms bottlenecks (e.g., CoWoS not discussed much after swarming/doubling for two years—now good shape). TSMC scales CoWoS/future packaging at logic/memory scale. No longer specialty—mainstream computing.
Influencing larger supply chain. Five years ago predicted all this; some believed/invested (e.g., Sanjay/Micron doubled down on LPDDR/HBM). Now all in. Prefetch bottlenecks years ahead: investments with Lumentum, Coherent, silicon photonics around TSMC; partnered on COUPE, invented/licensed patents openly. New workflows, testing (double-sided probing), company investments to scale.
Shaping ecosystem for scale. Bottlenecks easier/harder (hardest: plumbers/electricians). Concern: doomers discourage software engineers/radiologists, but short of both (radiology first to go prediction wrong).
Logic/memory bottlenecked by EUV. Scale to 2x EUV machines yearly? Easy within 2-3 years with demand signal: build one → ten → million. Not hard to replicate.
Convince directly/indirectly (e.g., convince TSMC → ASML convinced). Focus critical pinch points. No bottlenecks >2-3 years. Meanwhile, efficiency improves: 10x-20x, Hopper to Blackwell 30x-50x. New algorithms via flexible CUDA, new techniques. Plus capacity increase. Downstream worries: energy policies. Can't create industry without energy (reindustrialize US: chips, computers, packaging, EVs, robots, AI factories). Energy takes longer than 2-3 year chip/CoWoS ramps.
TPU: Claude, Gemini trained on it (2/3 top models). Implications?
Nvidia built accelerated computing, not TPU. Used for molecular dynamics, quantum chromodynamics, data processing/frames (structured/unstructured), fluid dynamics, particle physics, AI. Broader than AI. Reinvented computing: general-purpose → accelerated. Only company accelerating all applications with gigantic ecosystem/frameworks/algorithms.
Computers designed for others to operate: in every cloud (Google, Amazon, Azure, OCI). Home-built need self-operation. Enable operators anywhere/industry (e.g., Elon/xAI, Lilly drug discovery/biology supercomputer). CUDA fantastic TPU + full data/compute/AI lifecycle. Larger market/reach—build anywhere, customers exist.
Revenue from AI, not pharma/quantum. TPU systolic array perfect for matrix multiplies (predictable, no die area for warp schedulers/switches/threads/memory). GPU flexible for branching/irregular access. AI = predictable matrix multiplies—TPU optimized for bulk growth.
Matrix multiplies important but not only: need programmable for new attention, disaggregation, architectures (hybrid SSM), fusing diffusion/autoregressive. Runs everything—enables new algorithms advancing AI quickly. TPUs/anything follow Moore's Law (25%/year). 10x/100x leaps need algorithm changes yearly.
Nvidia advantage: Blackwell 35x (announced) → 50x energy efficient vs Hopper (Dylan article: sandbagged). Not Moore's Law—new models (MoEs) parallelized/disaggregated/distributed. Needs CUDA kernels. Extreme co-design: processors, system, NVLink fabric, Spectrum-X network, libraries, algorithms.
60% revenue from big five hyperscalers. Hyperscalers write own kernels (Triton/OpenAI, TPUs/Trainium/Anthropic/Google; OpenAI compiles to accelerators). Do they need CUDA?
CUDA rich ecosystem: build on it first (supports every framework, contributes to Triton backend). Frameworks: Triton, vLLM, SGLang, verl, NeMo RL (post-training/RL exploding). Build on CUDA for great ecosystem—bugs in your code, not mountain underneath. Trust the foundation.
Install base: hundreds of millions GPUs (A10, A100, H100, H200, L/P series) everywhere (clouds, robots). Software/models run fleet-wide. In every cloud + on-prem. Richness, expansiveness, versatility make CUDA invaluable.
Advantages matter less to main customers (hyperscalers build stacks)? AI good at verifiable loops (e.g., efficient attention/MLP kernels). Could become price/performance/flops/memory bandwidth competition, eroding +70% margins from CUDA moat.
Insane engineers assigned to AI labs optimizing stacks—nobody knows architecture better. Architectures not as general as CPU (Cadillac: cruiser, cruise control, easy).
Everybody drives it pretty well. It's got cruise control, and everything's easy. But Nvidia's GPUs, accelerators, are like F1 racers. Everybody's able to drive it at a hundred miles an hour, but it takes quite a bit of expertise to push it to the limit. We use a ton of AI to create the kernels that we have. We're going to still be needed for quite some time. Our expertise helps our AI lab partners get another 2x out of their stack easily oftentimes. By the time we're done optimizing their stack or a particular kernel, their model sped up by 3x, 2x, 50%. That's a huge number, especially with the install base of Hoppers and Blackwells. When you increase it by a factor of two, that doubles the revenues. Nvidia's computing stack is the best performance per TCO in the world, bar none. Nobody can demonstrate any single platform with a better performance-TCO ratio.
Benchmarks like Dylan's InferenceMAX are available for everybody. TPU won't come, Trainium won't come. Nobody wants to show up. MLPerf welcomes Trainium to demonstrate their 40% claim. It makes no sense on first principles. The reason for success is great TCO.
60% of customers are the top five, but most business is external. Most Nvidia in AWS is for external customers. All customers at Azure and OCI are external. They favor Nvidia because of great reach bringing great customers. The flywheel is install base, programmability, richness of ecosystem, tens of thousands of AI companies. AI startups choose the most abundant architecture with largest install base and rich ecosystem. Perf per dollar gives lowest cost tokens. Perf per watt is highest. For a one gigawatt data center, highest tokens per watt maximizes revenues. Nvidia has the most customers for renting infrastructure.
Premise that big AI labs like Anthropic, OpenAI use different accelerators is wrong. Anthropic is unique, not a trend. Without Anthropic, no TPU growth, no Trainium growth. OpenAI deals with AMD, building Titan accelerator, but vastly Nvidia. We're not offended; they need to try others to know how good ours is. We continuously earn our position. Many ASICs canceled. Not easy to build better than Nvidia. Nvidia cranks out big leaps every year due to scale and velocity.
ASIC logic: doesn't need to be better, just not more than 70% worse, since Nvidia margins 70%, ASIC 65%. ASIC margins incredibly good. Long ago, didn't internalize difficulty of building foundation AI lab needing huge supplier investments. Google and AWS invested multi-billion into Anthropic for compute use. Nvidia wasn't in position then. Mistake: didn't realize VCs wouldn't put $5-10 billion into AI lab. Wouldn't have been able to even if understood. Not making that mistake again: invested in OpenAI, Anthropic to help scale.
Did it as soon as could. At Anthropic's need time, not positioned. Never invested outside company that much, thought VCs would fund. What they did couldn't be done through VCs. That's their genius. Delighted they succeeded, even if Anthropic went elsewhere. Anthropic's existence great for world.
Philosophy: do as much as needed, as little as possible. Nvidia builds computing platform; if not, it doesn't get done. Took risks on NVLink, stack, ecosystem, 20 years of CUDA losing money. Created CUDA-X domain-specific libraries decade and half ago for ray tracing, image generation, early AI, data processing. Created cuLitho for computational lithography. Accelerated computing advanced because of this. World has lots of clouds; somebody would build them. Apply philosophy: support neoclouds like CoreWeave ($6.3B backstop, $2B invested), Nscale, Nebius. Invest in ecosystem to thrive, connect AI with industries, countries, build planet on AI and American tech stack.
Invest in all great foundation model companies, don't pick winners. Imperative to business. When Nvidia started, 60 3D graphics companies; Nvidia only survivor. Nvidia's graphics architecture precisely wrong, impossible for developers. Have humility: don't pick winners, support all.
Neoclouds need to want to exist, have business plan, expertise, passion. Provide investment to get off ground. Not in financing business; work with financiers. Focus on what we do, simple business model, support ecosystem. For OpenAI $30B scale pre-IPO, support because world needs them.
Not fracturing market. Work hard on forecasts aligning demand/supply; data centers take time. Must place PO; first in, first out. If data center not ready, serve others to maximize factory throughput. Prioritization: place PO. Stories like Larry/Elon dinner begging GPUs never happened; just place order.
Never highest bidder. Bad practice. Set price, people decide to buy. Others change prices on demand; Nvidia doesn't. Dependable, foundation of industry. Quoted price is it.
Nvidia-TSMC 30 years, no legal contract, rough justice. Incredible relationship, complete trust. Count on Nvidia: this year Vera Rubin incredible, next Vera Rubin Ultra, then Feynman, then unnamed. Every year big leaps, token cost decreases order of magnitude. Only company to bet entire business on yearly. Scales from $1 to $100B AI factory compute, one rack, one graphics card. Same for TSMC. Position took decades of commitment, stability, consistency.
Anthropic's Mythos Preview not public due to cyber-offensive capabilities, found thousands high-severity vulnerabilities across OSes, browsers, one in OpenBSD (27 years). Concern if China trains such models. Mythos trained on mundane capacity available in China. China manufactures 60% world's mainstream chips, great computer scientists, 50% world's AI researchers, abundance energy/chips. Best way for safe world: dialogue, research dialogue, not victimizing/turning into enemy. Want US to win.
This is an area that is glaringly missing because of our current attitude about China as an adversary. It is essential that our AI researchers and their AI researchers are actually talking. It is essential that we try to both agree on what not to use the AI for.
With respect to finding bugs in software, that's what AI is supposed to do. AI is going to find bugs in a lot of software. There are lots of bugs in the AI software.
AI has reached a level where it could help us be so much more productive.
One of the things that is underemphasized is the richness of the ecosystem around cybersecurity, AI cybersecurity, AI security, AI privacy, and AI safety. There’s a whole ecosystem of AI startups trying to create this future, where you have one AI agent that's incredible, surrounded by thousands of AI agents, keeping it safe, keeping it secure. That future surely is going to happen.
The idea that you're going to have an AI agent running around with nobody watching after it is kind of insane. This ecosystem needs to thrive.
This ecosystem needs open source. This ecosystem needs open models. They need open stacks so that all of these AI researchers and great computer scientists can go build AI systems that are as formidable and can keep AI safe. We need to keep the open source ecosystem vibrant. A lot of that is coming out of China. We ought to not suffocate that.
With respect to China, we want the United States to have as much computing as possible. We're limited by energy, but we've got a lot of people working on that. We've got to not make energy a bottleneck for our country.
We also want to make sure that all the AI developers in the world are developing on the American tech stack, and making the contributions, the advancements of AI—especially when it's open source—available to the American ecosystem.
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. That would be a horrible outcome for the United States.
The concern, going back to the flop difference in the hacking, is yes, they have compute, but some estimates say that because they're at 7nm—they don't have EUVs because of chip-making export controls—the amount of flops they're able to actually produce, they have one tenth the amount of flops that the US has.
Could they eventually train a model like Mythos? Yes. But because we have more flops, American labs are able to get to these levels of capabilities first. Because Anthropic got to it first, they hold onto it for a month while all these American companies get access to it. They're going to patch up all their vulnerabilities, and now we release it.
Even if they train a model like this, the ability to deploy it at scale… If you had a cyber hacker, it's much more dangerous if they have a million of them versus a thousand of them. Inference compute really matters a lot.
The fact that they have so many AI researchers who are so good is the thing that makes it so scary, because what makes those engineer researchers more productive? It’s compute. If you talk to any AI lab in America, they say the thing that's bottlenecking them is compute. There are quotes from the DeepSeek founder, or Qwen leadership. They say the thing they’re bottlenecked on is compute.
Isn't it better that we get American companies, because they have more compute, to get to the Mythos-level capabilities first, prepare our society for it, before China can get to it because they have less compute? We should always be first and we should always have more.
In order for that outcome to be true, you have to take it to the extremes. They have to have no compute. If they have some compute, the question is how much is needed? The amount of compute they have in China is enormous.
China is the second largest computing market in the world. If they want to aggregate their compute, they've got plenty of compute to aggregate. But people do these estimates and say SMIC is actually behind on the process nodes.
The amount of energy they have is incredible. AI is a parallel computing problem. Why can't they just put 4x, 10x as many chips together because energy's free? They have so much energy. They have datacenters that are sitting completely empty, fully powered. They have ghost cities, they have ghost datacenters too. They have so much infrastructure capacity.
If they wanted to, they just gang up more chips, even if they're 7nm.
Their capacity of building chips is one of the largest in the world. The semiconductor industry knows that they monopolize mainstream chips. They have over-capacity, they have too much capacity.
The idea that China won't be able to have AI chips is completely nonsense.
Of course, if you ask me, would the United States be further ahead if the entire world had no compute at all? But that's just not an outcome. That's not a scenario that's true. They have plenty of compute already. The amount of threshold they need for the concern you're worried about, they've already reached that threshold and beyond.
AI is a five-layer cake, and at the lowest layer is energy. When you have an abundance of energy, it makes up for chips. If you have an abundance of chips, it makes up for energy.
For example, the United States is scarce on energy, which is the reason why Nvidia has to keep advancing our architecture and do this extreme co-design so that with the few chips that we ship—with the few chips, because the amount of energy is so limited—our throughput per watt is off the charts.
But if your amount of watts is completely abundant, it's free, what do you care about performance per watt for? You get plenty. You can use old chips to do. So 7nm chips are essentially Hopper.
Today's models are largely trained on Hopper, Hopper generation. So 7nm chips are plenty good.
The abundance of energy is their advantage. But then there's a question of whether they can actually manufacture enough chips. But they do. What's the evidence? Huawei just had the largest single year in the history of their company. How many chips did they ship? A ton. Millions. Millions is way more than Anthropic has.
There's a question of how much logic SMIC can chip, and how much memory. They have plenty of logic, and they have plenty of HBM2 memory.
The bottleneck often in training and doing inference on these models is the amount of bandwidth. If you have HBM2 versus the newest thing you have, there could be almost an order of magnitude difference in memory bandwidth, which is huge.
Huawei is a networking company.
But that doesn't change the fact that you need EUV for the most advanced HBM. Not true. Not at all true. You could gang them together, just like we gang them together with NVL72. They've already demonstrated silicon photonics, connecting all of this compute together into one giant supercomputer. Your premise is just wrong. Their AI development is going just fine.
The best AI researchers in the world, because they're limited in compute, they also come up with extremely smart algorithms. Moore's law is advancing about 25% per year. However, through great computer science, we could still improve algorithm performance by 10x.
Great computer science is where the lever is. MoE is a great invention. All the incredible attention mechanisms reduce the amount of compute.
Most of the advances in AI came out of algorithm advances, not just the raw hardware. If most advances came from algorithms and computer science and programming, their army of AI researchers is their fundamental advantage.
DeepSeek is not an inconsequential advance. The day that DeepSeek comes out on Huawei first, that is a horrible outcome for our nation.
Because currently you can have a model like DeepSeek that can run on any accelerator, if it's open source. Why would that stop being the case in the future? Suppose it's optimized for Huawei, suppose it's optimized for their architecture. It would put ours at a disadvantage.
You described a situation that I perceive to be good news. A company developed software, developed an AI model, and it runs best on the American tech stack. I saw that as good news. You set it up as a premise that it was bad news.
I'm going to give you the bad news, that AI models around the world are developed and they run best on non-American hardware. That is bad news for us.
I guess I just don't see the evidence that there's these huge disparities that would prevent you from switching accelerators. American labs are running their models across all the clouds, across all the different accelerators.
You take a model that's optimized for Nvidia and you try to run it on something else. But American labs do that. And they don't run better. Nvidia's success is perfect evidence. The fact that AI models are created on our stack, run best on our stack, how is that illogical to understand?
Anthropic's models are run on GPUs, they're run on Trainium, they're run on TPUs. A lot of work has to go into it to change.
Go to the global south, go to the Middle East. Coming out of the box, if all of the AI models run best on somebody else's tech stack, you've got to be arguing some ridiculous claim right now that that's a good thing for the United States.
Say Chinese companies get to the next Mythos first. They find all the security vulnerabilities in American software first, but they can do it on Nvidia hardware and they ship it to the global south. They do it on Nvidia hardware. How is that good?
It's not good. So let's not let it happen. Why do you think it's perfectly fungible, that if you didn't ship them compute it would exactly be replaced by Huawei? They are behind, right? They have worse chips than you.
Their chip industry's gigantic. You can just look at the flop or bandwidth or memory comparisons between the H200 and the Huawei 910C. It's like half to a third. They use more of it. They use twice as many.
It seems like your argument is they have all this energy that's ready to go, and they need to fill it with chips. And they're good at manufacturing. And eventually they would be able to just out-manufacture everybody.
These next few years. We've got these models that are going to be able to do all the cyber attacks. In that case, if the next years are critical, then we have to make sure that all of the world's AI models are built on the American tech stack, in these critical years.
If they're built on the American tech stack, how would that prevent them, if they have more advanced capabilities, from launching the Mythos-equivalent cyber attacks? There's no guarantee either way. But if you have it early, we can prepare for it.
Why are you causing one layer of the AI industry to lose an entire market so that you could benefit another layer of the AI industry? There are five layers and every single layer has to succeed. The layer that has to succeed most is actually the AI applications.
Why are you so fixated on that AI model? That one company? For what reason? Because those models make possible these incredibly offensive capabilities, and you need compute to run them.
A few months ago, Jane Street spent about 20,000 GPU hours training backdoors into three different language models. Then they challenged my audience to find the trigger phrases.
“If you think the base model was here and the backdoor model was here, you can kind of linearly interpolate the weights to adjust the strength of the backdoor, but you can also extrapolate it to make the backdoor even stronger. And in some cases, if you make it strong enough the model will just regurgitate what the response phrase was supposed to be.”
If you keep amplifying the difference between the base version and the backdoored version, eventually it should spit out the trigger phrase. But this technique only worked on two out of the three models. Even Ricson isn't sure why it didn't work on the other.
Being able to verify that a model only does what you think it does is one of the most important open questions in AI security.
If this is the kind of problem that excites you, Jane Street is hiring researchers and engineers. Go to janestreet.com/dwarkesh to learn more.
China is able to build enough 7nm capacity. They're still stuck on 7nm while you'll move on to 3nm and then 2nm or 1.6nm with Feynman.
While you're on 1.6nm, they're still going to be on 7nm, and they have to produce enough of it to make up for the shortfall. They have so much energy that the more chips you give them, the more compute they'd have.
Compute is an input to training and inference.
The United States ought to be ahead. The amount of compute in the United States is 100x more than anywhere else in the world. The United States ought to be ahead.
Nvidia builds the most advanced technologies. We make sure that the US labs are the first to hear about it and have the first chance to buy it. And if they don't have enough money, we even invest in them. The United States ought to be ahead. We want to do everything we can to make sure the United States is ahead.
We're doing everything we can to do that. But how is shipping chips to China keeping the US ahead if they’re bottlenecked on compute?
We've got Vera Rubin for the United States. Nvidia. You consider Nvidia a United States company? Why is it that we don't come up with a regulation that's more balanced so that Nvidia can win around the world instead of giving up the world? Why would you want the United States to give up the world?
The chip industry is part of the American ecosystem. It's part of American technology leadership. It's part of the AI ecosystem. It's part of AI leadership. Why is it that your policy, your philosophy, leads to the United States giving up a vast part of the world's market?
Dario had this quote where he said that it's like Boeing bragging that we're selling North Korea nukes, but the missile casings are made by Boeing. And that's somehow enabling the US technology stack.
Fundamentally, you're giving them this capability. Comparing AI to anything that you just mentioned is lunacy. But AI is similar to enriched uranium, right? It can have positive uses, it can have negative uses. We still don't want to send enriched uranium to other countries.
The analogy is that enriched uranium is like compute. It's a lousy analogy. It's an illogical analogy.
But if that compute can run a model that can do zero-day exploits against all American software, how is that not a weapon?
First of all, the way to solve that problem is to have dialogues with the researchers and dialogues with China, and dialogues with all the countries to make sure that people don't use technology in that way. That's a dialogue that has to happen.
Number two, we also need to make sure that the United States is ahead, that Vera Rubin, Blackwell, is available in the United States in abundance, mountains of it. Our results would show it. Abundance, tons of it. The amount of computing we have is great. We have amazing AI researchers here. It's great. We ought to stay ahead.
AI is not just a model. AI is a five-layer cake. The AI industry matters across every single layer, and we want the United States to win at every single layer, including the chip layer. Conceding the entire market is not going to allow the United States to win the technology race long-term in the chip layer, in the computing stack. That is just a fact.
How does selling them chips now help us win in the long term? Tesla sold extremely good electric vehicles to China for a long time. iPhones are sold in China, extremely good. They didn't cause them lock-in. China will still make their version of EVs and they're dominating. Their smartphones are dominating.
When we started the conversation today, you acknowledged that Nvidia's position is very different. You used words like moat. The single most important thing to our company is the richness of our ecosystem, which is about developers. 50% of the AI developers are in China. The United States should not give that up.
We have a lot of Nvidia developers in the US, and that doesn't prevent American labs from also being able to use other accelerators in the future. In fact, right now they're using other accelerators as well, which is fine and great.
Google can use TPUs and Nvidia. We have to keep innovating and, our share is growing, not decreasing.
The premise that even if we competed in China, that we're going to lose that market anyways… You're not talking to somebody who woke up a loser. That loser attitude, that loser premise makes no sense to me.
We're not a car. We are not a car. The fact that I can buy this car brand one day and use another car brand another day, easy. Computing is not like that. There's a reason why the x86 deal exists. There's a reason why ARM is so sticky. These ecosystems are hard to replace. It costs an enormous amount of time and energy, and most people don't want to do it.
It's our job to continue to nurture that ecosystem, to keep advancing the technology so that we can compete in the marketplace. Conceding a marketplace based on the premise you described, I simply can't acknowledge that. It makes no sense. Because I don't think the United States is a loser. Our industry is not a loser. That losing proposition, that losing mindset, makes no sense to me.
The crux is you're going to extremes. Your argument starts from extremes. That if we give them any compute at all in this narrow moment, we will lose everything.
The idea is not that there is some key threshold of compute. It's that any marginal compute is helpful. So if you have more compute, you can train a better model.
Any marginal sales for the American technology industry is beneficial. If the AI models that run on those chips are capable of cyber offensive capabilities, or the chips are training models with cyber capabilities and running more instances of those models, it is not a nuclear weapon, but it enables a weapon of a kind.
The logic that you use, you might as well say it to microprocessors and DRAMs. You might as well say it to electricity.
But in fact we do have export controls on the technology that is relevant to making the most advanced DRAM. We have all kinds of export controls on China for all kinds of chip-making stuff. We sell a lot of DRAM and CPUs into China, and I think it's right.
This goes back to the fundamental question of, is AI different? If you have the kind of technology where they can find these zero-days in software, is that something where we want to minimize China's ability to get there first, to deploy it widely?
We want the United States to be ahead. We can control that. How do we control that if the chips are already there and they're using them to train that model? We have tons of compute. We have tons of AI researchers. We're racing as fast as we can.
We have more nuclear weapons than anybody else, but we don't want to send enriched uranium anywhere. We're not enriched uranium. It's a chip, and it's a chip that they can make themselves. But there's a reason they're buying it from you.
We have quotes from the founders of Chinese companies that say that they’re bottlenecked on compute. Because our chips are better. On balance, our chips are better. There's just no question about it.
In the absence of our chip… Huawei had a record year? A whole bunch of chip companies have gone public? China is about 40% of the world's technology industry.
To concede that market for the United States technology industry 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.
You're making two different statements. One is that we're going to win this competition with Huawei because our chips are going to be way better if we're allowed to compete. Another is that they would be doing the same exact thing without us anyway. How can both of those things be true at the same time?
It's obviously true. In the absence of a better choice, you'll take the only choice you have. How is that illogical? It's so logical.
The reason they want Nvidia chips is that they're better. Better is more compute. More compute means you can train a better model. No, it's just better. It's better because it's easier to program. We have a better ecosystem.
Whatever the better is… And of course we're going to send them compute. So what? The fact of the matter is that we get to benefit. Don't forget, we get the benefit of American technology leadership. We get the benefit of developers working on the American tech stack. We get the benefit, as those AI models diffuse out into the rest of the world, that the American tech stack is therefore the best for it.
We can continue to advance and diffuse American technology. That, I believe, is a positive. It's a very important part of American technology leadership.
The policies that you're advocating resulted in the American telecommunications industry being policied out of basically the world, to the point where we don't control our own telecommunications anymore. I don't see that as smart. It's a little narrow-minded, and it led to unintended consequences that I'm describing to you right now.
There's a potential benefit and there's a potential cost. What we're trying to figure out is, is the benefit worth the cost? I'm trying to get you to acknowledge the potential cost.
Compute is an input to training powerful models. Powerful models do have powerful offensive capabilities, like cyber attacks.
American companies achieved Mythos-level capabilities first and are now holding off to protect software before announcing it. If China had more compute or crowd compute and deployed a Mythos-level model earlier, it would have been very bad. One reason this hasn't happened is more compute from American companies like Nvidia.
Export controls to China have a cost: conceding the chip layer, the most important layer of the AI stack, in the second largest market in the world. This allows China to develop scale, their own ecosystem, and optimize future AI models differently from the American tech stack. As AI diffuses globally, Chinese standards and tech stack could become superior because their models are open.
China is the largest contributor to open source software in the world. Fact. China's the largest contributor to open models in the world. Fact. Today it's built on the American tech stack, Nvidia’s.
All five layers of the AI tech stack are important. The US should win all five. The most important is the AI application layer, which diffuses into society and benefits most from the industrial revolution. Every layer must succeed.
Scaring people into thinking AI is a nuclear bomb does a disservice to the US. Scaring people out of software engineering jobs because AI will kill them leads to a shortage of software engineers. Scaring people out of radiology because computer vision is free misunderstands jobs vs. tasks: a radiologist's job is patient care; the task is reading a scan. Scaring people out of radiology school harms healthcare.
When you make a premise that is so extreme, everything goes from zero or infinity, we end up scaring people in a way that's just not true. Life is not like that.
The US wants to be first and lead every layer. In a few years, when exporting American tech stack to India, Middle East, Africa, Southeast Asia—exporting standards—the policy of conceding the second largest market will have been a mistake for no good reason.
Nobody advocates all-or-nothing: ship everything to China. Keep the best technology in the US first, but compete and win globally. Both can happen with nuance and maturity, not absolutes.
China's models optimized for their chips could set global standards. Due to EUV export controls, Nvidia moves to 1.6nm while they stay on 7nm. Domestically, China has energy to manufacture at scale on 7nm, but their 7nm chips must compete against Nvidia's 1.6nm. Their models must be so optimized for 7nm that it's better than running on 1.6nm.
Blackwell is not 50 times more advanced lithography than Hopper. Moore's Law is dead. Between Hopper and Blackwell (3 years apart), transistors increased ~75%. Blackwell is 50x Hopper in performance. Architecture matters. Computer science matters. Semiconductor physics matters, but computer science matters.
AI impact comes from the computing stack, why CUDA is effective and beloved. It's an ecosystem with flexibility for MoE, diffusion, disaggregated systems.
AI is about the stack above as much as it is about the architecture below.
Nvidia's ecosystem is rich; researchers love programming CUDA first, including in China. Forcing Nvidia out of China is a policy mistake: it accelerated their chip industry and focused their AI ecosystem on internal architectures. It's not too late, but it's happening. China is good at manufacturing and will advance beyond 7nm. No 10x difference between 5nm and 7nm. Networking matters (why Nvidia bought Mellanox). Energy matters.
Nvidia is majority of TSMC N3, soon N2. Could Nvidia use spare N7 capacity for Hopper or Ampere refreshed with modern numerics to meet AI demand before 2030? Not necessary. Each generation improves architecture, engineering, packaging, stacking, numerics, system architecture beyond transistor scale. Going back to older nodes requires unaffordable R&D. Nvidia leans forward. If no more capacity ever, Nvidia would use 7nm in a heartbeat.
Why doesn't Nvidia run multiple chip projects in parallel (e.g., Cerebras-style wafer scale, Dojo-style huge package, non-CUDA)? They simulate everything; alternatives are provably worse. They work on optimal projects. If workloads change dramatically (not algorithms, but market shape), they may add accelerators. Recently added Groq into CUDA ecosystem because token value rose.
Tokens were once free or cheap. Now, customers want different answers; software engineers pay for more responsive tokens for productivity. This enables market segments by response time. Nvidia expanded Pareto frontier for faster inference (lower throughput). Previously, higher throughput always better. Now, high ASP tokens (even lower factory throughput) make sense due to emerging market.
If I had more money, I would put more behind Nvidia’s architecture.
Without deep learning revolution, Nvidia would do accelerated computing. Premise: general purpose computing (Moore's Law) insufficient for heavy computation. Combined GPU + CUDA + CPU accelerates workloads 100x-200x by offloading kernels/algorithms to GPU.
Applications: engineering, science, physics, data processing, computer graphics, image generation. Nvidia would be very large via domain-specific acceleration (started with graphics, then particle physics, fluids, structured data processing).
Mission: bring accelerated computing to the world for applications general purpose can't scale, advancing science (e.g., molecular dynamics, seismic processing, image processing).
Nvidia democratized deep learning via PCs and GeForce cards for researchers, scientists, students. Promise unchanged. At GTC, non-AI parts include computational lithography, quantum chemistry, data processing—unrelated to AI, using tensors beyond AI. Important work continues non-AI.
Keep Dwarkesh Patel in your library
Save the videos and channels worth coming back to, and find them again in one place.
