First of all, I love this podcast and the hard hitting questions. But maybe Dwarkesh could benefit from the Socratic method rather than these conversations going in circles whenever there is disagreement. For instance "If I understand you correctly, you think selling Nvidia chips to China won't actually accelerate their AI work because their chip making capacity will match what you could have provided?" and I think Jensen would have pushed back on that statement and been forced to concede that yes selling chips to them does accelerate their AI capacity. Then Dwarkesh could have said "Is it your position that foreign AI models are not a significant cybersecurity threat to the US within the next few years even as Mythos level models and better come online?". But it's hard in the moment, I couldn't have done any better.
If we don't sell them advanced radar systems, they'll just build their own, and then the global radar ecosystem won't be on the American standard. That's a loser mindset.
50% of the world's radar engineers are Chinese. We should keep them building on Lockheed's stack.
There were some real mask-off moments in this interview. Jensen was totally evasive about your (completely valid) questions about the sustainability of the CUDA moat and hyperscalers shifting market share to TPUs. Is he avoiding these (domestic) elephants in the room, or does Nvidia simply not have a strategy to address this?
It's valid what he says about Anthropic being unique though. Who else other than Google is serving significant traffic off TPUs?
It can be that Anthropic went to TPUs simply because they needed a hyperscaler with capacity and Google was right there. Probably not due to any strong feelings about the amazing TPU hardware or anything.
My read on the situation is that Anthropic initially went to TPUs largely for financial/strategic reasons (Google was an early investor, and had capacity ready). Then they discovered the TCO is genuinely competitive or better, and have since expanded dramatically by choice (up to 1M TPUs). Other companies are now following for the same economic reasons, not just because Google is subsidizing them.
Is it "significant?" That depends on your definition. But Meta, Apple, Midjourney, and xAI all use TPUs for a growing share of workloads — especially inference, where TCO matters most, and switching costs are lowest.
I don't think this is particularly dire for Nvidia. AI companies are buying every single chip that Nvidia manufactures, but then they need more compute on top of that so they go to TPUs and other solutions.
Maybe that indicates Nvidia didn't buy up enough capacity from TSMC. It's a hard problem - if you overshoot then it wrecks your margins.
Boy, people give Sam Altman a hard time for his snapping at Brad Gerstner, but Jesus I thought Jensen was super evasive and just shifted to attack mode whenever he was rhetorically cornered…
I exited this interview feeling a little less certain about my Nvidia shares and a little less well disposed toward Jensen. I felt the questions were challenging but fair, more importantly, I think they were material and needed to be asked then answered.
The stringent line Jensen took with his view 'what GPU's are and are not', which seems framed specifically to avoid 'restrictions', was disappointing.
Use of the patronising and belittling method was particularly concerning, as it usually flags a real lack of cogent rebuttal, which strangely - did appear to be the case at times.
A minor criticism that Dwarkesh may be able to employ a more sophisticated offramp strategy when faced with negative circular responses (easy to say, hard to do), pales compared with Jensen failing to engage in good faith at certain moments, even reverting to the new MO of Silicon Valley CEO's to 'falsely profess' thanks and enjoyment when they are clearly in an uncomfortable corner and behaving poorly.
I did think Jensen came across poorly in the section about China sales, but not in an unexpected way. What is he supposed to say? "Hey maybe we should ban my company from selling it's chips to the second biggest market in the world".
On the question of whether Nvidia can hold it's own in the face of many competing companies, I thought he was persuasive. Algorithmic progress is where most progress comes from, and you need something general purpose to keep up with cutting edge models, not highly specialized custom ASICs like Cerebras and Taalas. And among the relatively general purpose chips, Nvidia makes the best.
Everyone thinks CUDA is the moat because that's what everyone programs to.
This is such an interesting fallacy. I use to work for Intel.
What's the one thing that is going to make the AI business case?
Programming!
CUDA is not an advantage with AI programming, its a hinderance that AI needs to take into account. AI can program to the DSP or TPU directly likely much more efficiently and with much more flexibility than CUDA. Huang likely knows that but does not want to discuss it.
Tools like Claude are already being used to write the LLM code. Why wouldn't they be used to write the TSP or DSP code. CUDA is not needed. Libraries are an advantage when a developer is doing the code. They are not when an AI is doing the code.
Yes in the near term when the TPUs and LLMs have not perfected writing to a DSP or TPU directly for the workload, CUDA and NVidia DSPs are the only answer. But they have zero moat because AI and everyones push to undermine NVidia has already doomed them. Their shareholders want them to maintain their margins so they can lower their prices. A great AI for Nvidia to own would be a DSP trained LLM coding tool. But that would undermine their "CUDO moat." They are stuck in the innovators dilemma.
Intel is in its position because of the same problem. Intel could not and likely still cannot just be a fab because the margins are only 5pt. x86 is a 60pt margin business. They cannot deliver a generic TPU because the margins are low. That totally changes what their stock is worth. Intel would have been in better shape if they had gone private 10 years ago. NVidia has exactly the same problem.
Elon Musk is betting that the only long term answer is to own your own fab because the business case for building fabs is so challenging. Design your own chips is also something AI can do. Its an interesting question. Is there a business case for your own fab for someone like Elon. He's working with Intel. Intel would probably prefer to create a high margin consulting business case for creating Elon's fab than own the fab themselves.
The more interesting question is how AI affects Intel and x86. Is there any motivation to move to ARM or some other custom CPU architecture. That's to be determined. For now, the LLMs are using Intel/AMD for control of the DSPs and TPUs.
In the end, Huang happened to be in the right place at the right time. That's how many tech companies get their big moment. Apple is different as their marketing slogan goes. Given all the currently challenges that AI is having, it now looks absolutely like the right move that Apple is taking their time to figure out how LLMs can really create a better user experience. The inability of LLMs to learn from a user is the biggest challenge. They can only store user info and give it back to the LLM for every query. If someone (e.g. Apple) figures how how to train an LLM based on user use, that will be the real breakthrough.
Unfair to say he was merely in the right place at the right time. CUDA, the whole general accelerated compute vision, that was Huang and his colleagues from the start. They didn't just luck into AI, they made it happen too.
Note: Google that really did all the research on LLMs never deployed on Nvidia. They had TPUs from the start.
Note: Apple that hasn't fully deployed AI yet has TPUs built into the iphone and mac for AI.
You are delusional.
ChatGPT focused on LLMs versus the whole stack. They created the opportunity for Nvidia.
Nothing wrong with Nvidia being at the right place at the right time. You're welcome to invest all your $ in Nvidia. See how it turns out in the long run.
Google actually have used lots of nvidia chips. The TPUs are nice but have often been overfitted to older ML workloads. And so they have actually used plenty of GPUs as well.
I lived through the problem in Intel. NVidea is worse off than Intel was because NVidea does not own the fabs. Also Intel could maintain their 60% because of Moore's law increasing performance and reducing power every 18 months. Intel could maintain the price and margin because they were driving Moore's law (which was really just Intel's financial decision to invest in an 18 month cycle including driving new fabs).
Intel was at the right place at the right time - the IBM PC.
Intel created the "right place" when the internet took off.
They didn't depend on Microsoft or Sun Systems BSD for web servers.
They put $ into developing Linux as an alternative to both.
NVidea has no Moores law investment cycle. They are already have six major competitors to their DSP - AMD, Google, Amazon, Microsoft, Groc and Intel. Google and Amazon are well underway for deploying on TPUs. Google has shifted from promoting NVidia in their cloud AI offering to their TPUs. Anthropic is coding to the TPU. Amazon is pushing Trainium for its hosting. I assume Claude is capable of creating TPU code and Google/Micron have an AI for improving TPU iterations.
Given that AI can both create the code directly to the silicon and create the silicon, the only moat is the fab which NVidia does not have.
Help me out here, besides being in the right place at the right time, what moat does NVidia have that can maintain the margin that justifies their share price?
Yes Nvidia has a lot of competitors. They had a lot of competitors 2 years ago as well, and yet in that time, sales have steadily increased and their margins have not significantly budged. That seems like proof that hyper-scalers believe that their chips are worth the premium price.
The comparison with Intel is interesting. But Intel didn't fail to live up to their potential because someone made a better x86 chip than them - Intel still dominates that market. Rather Intel completely missed the boat on how big parallel computing would be, and ARM chips.
The current competitors to Nvidia don't seem comparable. Wafer scale chips, or custom ASICs with model architecture baked in are the most unique alternative paradigms. But model progress is so fast that these don't seem like viable prospects right now. You need something that is somewhat general to keep up.
2. Intel has about 55-60% of laptop. AMD 20%, ARM 20% and growing.
3. About 53% of the server market, AMD about 22%, ARM 25% and fasting growing.
Intel's biggest liability is their need to get 60pt margin. Their biggest advantage is they own the fabs.
NVidia is in the right place because AI is growing so fast that the demand is now. That demand is being questioned by the cost of NVidea DSPs both capx (purchase) and opex(power). VCs are not infinite pockets. Customers (as Anthropic right now is showing with Claude adoption) are also not infinite pocket.
The winner in AI is the one that reduces the CAPX and OPEX to scale. NVidea can't be part of that equation because it erodes their margins and they don't own the fab.
They will hit the Intel wall within the next 5 years. Except without their own fab.
You people are some shallow thinkers on how the market will play out.
Jensen’s best framing comes right at the start: Nvidia’s job is to “transform electrons to tokens,” and its role is to do “as much as necessary” and “as little as possible” to make that transformation happen at scale. That is a powerful description of what Nvidia actually built: not just chips, but a supply chain, software stack, ecosystem, and operating model that turned AI demand into industrial reality years before most people saw it clearly. On that point, he sounds right, and impressive.
But the rest of the interview shows why that does **not** prove Nvidia is permanently optimal for every future AI workload. Dwarkesh’s sharpest challenge is not “Nvidia is bad.” It is that the biggest customers increasingly have the money, engineering talent, and incentives to build around Nvidia. He presses exactly on whether hyperscalers can “write these custom kernel[s] for themselves” and whether, if they can, Nvidia’s advantage starts to collapse toward a specs-and-cost competition rather than a pure CUDA lock-in.
Jensen answers with maximal confidence. He says Nvidia’s stack is “the best performance per TCO in the world, bar none,” that “nobody can demonstrate” a better platform, that Nvidia has “the highest tokens per watt architecture in the world,” and that alternatives “make absolutely zero sense on first principles.” He also argues that Nvidia keeps winning because it combines architecture, system, network, libraries, and algorithms through “extreme co-design.”
The problem is that these are stronger claims than the interview really earns. Jensen is clearly defending the best broad, mature, programmable platform with the deepest install base and the most operational support. But that is not the same thing as proving Nvidia must remain the most efficient solution for every important transformer or LLM workload. Dwarkesh’s counterpoint is fair: modern AI is increasingly shaped by workloads that are narrower, more measurable, and more open to specialized optimization than they were a decade ago. A custom ASIC inside a custom stack does not have to beat Nvidia everywhere; it only has to beat Nvidia enough on specific high-volume workloads to matter.
That is why the deepest synthesis is this: Nvidia is extraordinary because it built the leading **general-purpose AI platform** before almost anyone else understood the future. But being the leading general-purpose platform is not the same thing as being permanently unbeatable by specialized architectures. Jensen himself indirectly admits the world can change: when asked why Nvidia does not pursue radically different architectures in parallel, he says they could, but “we don’t have a better idea,” and adds that “if the workload were to change dramatically,” Nvidia “may decide to add other accelerators.” That is not the language of a man who believes architecture is settled forever.
The China section reinforces the same point from another angle. Jensen argues that China is too important to concede because it is the “second largest market in the world,” has “50% of the world’s AI researchers,” and because forcing China off the American stack could create a rival ecosystem optimized for non-American hardware. In other words, even Jensen’s strategic fear is not just about selling more chips next quarter. It is about whether future models are built to run best on Nvidia at all.
So the clean conclusion is:
Nvidia deserves enormous credit for building the platform that currently converts “electrons to tokens” better than anyone at global scale. But the interview does not prove that a broad, programmable stack will remain the single most efficient answer for every future LLM workload. The real question is not whether Nvidia is remarkable, the real question is whether its general-purpose dominance can stay ahead of increasingly specialized, custom architectures built for the workloads that matter most.
> the real question is whether its general-purpose dominance can stay ahead of increasingly specialized, custom architectures built for the workloads that matter most.
Very true. But think about the crazy new architectures (I have no clue) coming to support continual learning. Odds are Nvidia will reap benefits there and a lot of ASIC plans too narrowly focused on transformer inference will go back to the drawing board.
“The beautiful thing is you’re talking to the expert”. The level of arrogance to think such a thing, let alone say it?! And I lost track of the number of times he conflated the word “I” with the company he runs. Apparently he is NVIDIA.
I think Jensen probably argues with his employees like this, so I would take this as a positive that he felt comfortable arguing hard knowing that you could take it.
Sounds like Nvidia is in a defensive posture from the divergence in geopolitical, national incentives with their own.
Combined with rising pressure to maintain their previously free CUDA ecosystem monopoly, it feels like Jensen is basically doubling down on being so globally infrastructurally entangled that they don't need to pick winners (whether at the level of individual companies or national fronts), like Standard Oil playing the railroads against each other while making them dependents.
From all accounts Jensen seems bent on viewing GPUs as an apolitical general computing substrate rather than the enriched-uranium strategic resource that they really are, and rejecting all frames where NVIDIA is anything but a neutral participant in one layer in a big AI cake.
I think Jensen makes a 1000% valid response that we shouldn’t give up the open source market. That’s how US industries get stale and die.
The only way Dwarkesh’s take makes sense is if we are in a VERY critical few-year window (a la AI 2027). If we are in a few-decade window then Jensen is correct IMO.
I think Dwarkesh quite clearly believes that, and other proponents of export control (I.e Dario) do as well. I personally find it very hard to determine which regime we’re in, but exponential progress over the past year makes me think they are both correct.
The competence gap between Jenson in this interview and Lisa Su in the Stratchery interview a couple of years ago is insane. Just so, so wide. Jenson is throwing around the names of random open source projects, he's recalling bad design decisions made decades ago, he's fluently discussing minor algorithmic details, drilling down supply chain bottlenecks to the level of "we don't have enough plumbers".
Meanwhile Lisa Su was telling BT that AMD was doing great in AI, where's the problem? Lol.
It's incredibly telling that zero time was spent on AMD in this interview. TPUs/Broadcom are the only competitor that matter now.
Holy crap, there was a ton of fighting in this pod. I guess Jensen isn't coming back anytime soon lmao
Nope haha.. Different take https://kaipability.substack.com/p/w-jensen-i-you-we-lx-x-is-a-loser hope you enjoy the meme!
First of all, I love this podcast and the hard hitting questions. But maybe Dwarkesh could benefit from the Socratic method rather than these conversations going in circles whenever there is disagreement. For instance "If I understand you correctly, you think selling Nvidia chips to China won't actually accelerate their AI work because their chip making capacity will match what you could have provided?" and I think Jensen would have pushed back on that statement and been forced to concede that yes selling chips to them does accelerate their AI capacity. Then Dwarkesh could have said "Is it your position that foreign AI models are not a significant cybersecurity threat to the US within the next few years even as Mythos level models and better come online?". But it's hard in the moment, I couldn't have done any better.
If we don't sell them advanced radar systems, they'll just build their own, and then the global radar ecosystem won't be on the American standard. That's a loser mindset.
50% of the world's radar engineers are Chinese. We should keep them building on Lockheed's stack.
I'm sorry but who's side does the argument support?
There were some real mask-off moments in this interview. Jensen was totally evasive about your (completely valid) questions about the sustainability of the CUDA moat and hyperscalers shifting market share to TPUs. Is he avoiding these (domestic) elephants in the room, or does Nvidia simply not have a strategy to address this?
It's valid what he says about Anthropic being unique though. Who else other than Google is serving significant traffic off TPUs?
It can be that Anthropic went to TPUs simply because they needed a hyperscaler with capacity and Google was right there. Probably not due to any strong feelings about the amazing TPU hardware or anything.
My read on the situation is that Anthropic initially went to TPUs largely for financial/strategic reasons (Google was an early investor, and had capacity ready). Then they discovered the TCO is genuinely competitive or better, and have since expanded dramatically by choice (up to 1M TPUs). Other companies are now following for the same economic reasons, not just because Google is subsidizing them.
Is it "significant?" That depends on your definition. But Meta, Apple, Midjourney, and xAI all use TPUs for a growing share of workloads — especially inference, where TCO matters most, and switching costs are lowest.
I don't think this is particularly dire for Nvidia. AI companies are buying every single chip that Nvidia manufactures, but then they need more compute on top of that so they go to TPUs and other solutions.
Maybe that indicates Nvidia didn't buy up enough capacity from TSMC. It's a hard problem - if you overshoot then it wrecks your margins.
Boy, people give Sam Altman a hard time for his snapping at Brad Gerstner, but Jesus I thought Jensen was super evasive and just shifted to attack mode whenever he was rhetorically cornered…
I exited this interview feeling a little less certain about my Nvidia shares and a little less well disposed toward Jensen. I felt the questions were challenging but fair, more importantly, I think they were material and needed to be asked then answered.
The stringent line Jensen took with his view 'what GPU's are and are not', which seems framed specifically to avoid 'restrictions', was disappointing.
Use of the patronising and belittling method was particularly concerning, as it usually flags a real lack of cogent rebuttal, which strangely - did appear to be the case at times.
A minor criticism that Dwarkesh may be able to employ a more sophisticated offramp strategy when faced with negative circular responses (easy to say, hard to do), pales compared with Jensen failing to engage in good faith at certain moments, even reverting to the new MO of Silicon Valley CEO's to 'falsely profess' thanks and enjoyment when they are clearly in an uncomfortable corner and behaving poorly.
I did think Jensen came across poorly in the section about China sales, but not in an unexpected way. What is he supposed to say? "Hey maybe we should ban my company from selling it's chips to the second biggest market in the world".
On the question of whether Nvidia can hold it's own in the face of many competing companies, I thought he was persuasive. Algorithmic progress is where most progress comes from, and you need something general purpose to keep up with cutting edge models, not highly specialized custom ASICs like Cerebras and Taalas. And among the relatively general purpose chips, Nvidia makes the best.
Wow, that was great. Huge respect, Dwarkesh, for pushing your questions!
Everyone thinks CUDA is the moat because that's what everyone programs to.
This is such an interesting fallacy. I use to work for Intel.
What's the one thing that is going to make the AI business case?
Programming!
CUDA is not an advantage with AI programming, its a hinderance that AI needs to take into account. AI can program to the DSP or TPU directly likely much more efficiently and with much more flexibility than CUDA. Huang likely knows that but does not want to discuss it.
Tools like Claude are already being used to write the LLM code. Why wouldn't they be used to write the TSP or DSP code. CUDA is not needed. Libraries are an advantage when a developer is doing the code. They are not when an AI is doing the code.
Yes in the near term when the TPUs and LLMs have not perfected writing to a DSP or TPU directly for the workload, CUDA and NVidia DSPs are the only answer. But they have zero moat because AI and everyones push to undermine NVidia has already doomed them. Their shareholders want them to maintain their margins so they can lower their prices. A great AI for Nvidia to own would be a DSP trained LLM coding tool. But that would undermine their "CUDO moat." They are stuck in the innovators dilemma.
Intel is in its position because of the same problem. Intel could not and likely still cannot just be a fab because the margins are only 5pt. x86 is a 60pt margin business. They cannot deliver a generic TPU because the margins are low. That totally changes what their stock is worth. Intel would have been in better shape if they had gone private 10 years ago. NVidia has exactly the same problem.
Elon Musk is betting that the only long term answer is to own your own fab because the business case for building fabs is so challenging. Design your own chips is also something AI can do. Its an interesting question. Is there a business case for your own fab for someone like Elon. He's working with Intel. Intel would probably prefer to create a high margin consulting business case for creating Elon's fab than own the fab themselves.
The more interesting question is how AI affects Intel and x86. Is there any motivation to move to ARM or some other custom CPU architecture. That's to be determined. For now, the LLMs are using Intel/AMD for control of the DSPs and TPUs.
In the end, Huang happened to be in the right place at the right time. That's how many tech companies get their big moment. Apple is different as their marketing slogan goes. Given all the currently challenges that AI is having, it now looks absolutely like the right move that Apple is taking their time to figure out how LLMs can really create a better user experience. The inability of LLMs to learn from a user is the biggest challenge. They can only store user info and give it back to the LLM for every query. If someone (e.g. Apple) figures how how to train an LLM based on user use, that will be the real breakthrough.
Unfair to say he was merely in the right place at the right time. CUDA, the whole general accelerated compute vision, that was Huang and his colleagues from the start. They didn't just luck into AI, they made it happen too.
LLMs can run on an Intel processor.
LLMs can't scale on Intel processors.
Note: Google that really did all the research on LLMs never deployed on Nvidia. They had TPUs from the start.
Note: Apple that hasn't fully deployed AI yet has TPUs built into the iphone and mac for AI.
You are delusional.
ChatGPT focused on LLMs versus the whole stack. They created the opportunity for Nvidia.
Nothing wrong with Nvidia being at the right place at the right time. You're welcome to invest all your $ in Nvidia. See how it turns out in the long run.
Google actually have used lots of nvidia chips. The TPUs are nice but have often been overfitted to older ML workloads. And so they have actually used plenty of GPUs as well.
I'm am fascinated by your defense of NVidea.
I lived through the problem in Intel. NVidea is worse off than Intel was because NVidea does not own the fabs. Also Intel could maintain their 60% because of Moore's law increasing performance and reducing power every 18 months. Intel could maintain the price and margin because they were driving Moore's law (which was really just Intel's financial decision to invest in an 18 month cycle including driving new fabs).
Intel was at the right place at the right time - the IBM PC.
Intel created the "right place" when the internet took off.
They didn't depend on Microsoft or Sun Systems BSD for web servers.
They put $ into developing Linux as an alternative to both.
NVidea has no Moores law investment cycle. They are already have six major competitors to their DSP - AMD, Google, Amazon, Microsoft, Groc and Intel. Google and Amazon are well underway for deploying on TPUs. Google has shifted from promoting NVidia in their cloud AI offering to their TPUs. Anthropic is coding to the TPU. Amazon is pushing Trainium for its hosting. I assume Claude is capable of creating TPU code and Google/Micron have an AI for improving TPU iterations.
Given that AI can both create the code directly to the silicon and create the silicon, the only moat is the fab which NVidia does not have.
Help me out here, besides being in the right place at the right time, what moat does NVidia have that can maintain the margin that justifies their share price?
Yes Nvidia has a lot of competitors. They had a lot of competitors 2 years ago as well, and yet in that time, sales have steadily increased and their margins have not significantly budged. That seems like proof that hyper-scalers believe that their chips are worth the premium price.
The comparison with Intel is interesting. But Intel didn't fail to live up to their potential because someone made a better x86 chip than them - Intel still dominates that market. Rather Intel completely missed the boat on how big parallel computing would be, and ARM chips.
The current competitors to Nvidia don't seem comparable. Wafer scale chips, or custom ASICs with model architecture baked in are the most unique alternative paradigms. But model progress is so fast that these don't seem like viable prospects right now. You need something that is somewhat general to keep up.
Intel does not dominate the market.
1. Intel has zero pressence in mobile/tablets
2. Intel has about 55-60% of laptop. AMD 20%, ARM 20% and growing.
3. About 53% of the server market, AMD about 22%, ARM 25% and fasting growing.
Intel's biggest liability is their need to get 60pt margin. Their biggest advantage is they own the fabs.
NVidia is in the right place because AI is growing so fast that the demand is now. That demand is being questioned by the cost of NVidea DSPs both capx (purchase) and opex(power). VCs are not infinite pockets. Customers (as Anthropic right now is showing with Claude adoption) are also not infinite pocket.
The winner in AI is the one that reduces the CAPX and OPEX to scale. NVidea can't be part of that equation because it erodes their margins and they don't own the fab.
They will hit the Intel wall within the next 5 years. Except without their own fab.
You people are some shallow thinkers on how the market will play out.
Didn't listen to the podcast (yet), but I swear I heard this:
Dwarkesh: Are we at AGI?
Jensen: Yes.
D: How do you know?
J: I know.
D: When will it be GA?
J: Now.
D: Who's going to win this race?
J: Me.
... every single time.
Jensen’s best framing comes right at the start: Nvidia’s job is to “transform electrons to tokens,” and its role is to do “as much as necessary” and “as little as possible” to make that transformation happen at scale. That is a powerful description of what Nvidia actually built: not just chips, but a supply chain, software stack, ecosystem, and operating model that turned AI demand into industrial reality years before most people saw it clearly. On that point, he sounds right, and impressive.
But the rest of the interview shows why that does **not** prove Nvidia is permanently optimal for every future AI workload. Dwarkesh’s sharpest challenge is not “Nvidia is bad.” It is that the biggest customers increasingly have the money, engineering talent, and incentives to build around Nvidia. He presses exactly on whether hyperscalers can “write these custom kernel[s] for themselves” and whether, if they can, Nvidia’s advantage starts to collapse toward a specs-and-cost competition rather than a pure CUDA lock-in.
Jensen answers with maximal confidence. He says Nvidia’s stack is “the best performance per TCO in the world, bar none,” that “nobody can demonstrate” a better platform, that Nvidia has “the highest tokens per watt architecture in the world,” and that alternatives “make absolutely zero sense on first principles.” He also argues that Nvidia keeps winning because it combines architecture, system, network, libraries, and algorithms through “extreme co-design.”
The problem is that these are stronger claims than the interview really earns. Jensen is clearly defending the best broad, mature, programmable platform with the deepest install base and the most operational support. But that is not the same thing as proving Nvidia must remain the most efficient solution for every important transformer or LLM workload. Dwarkesh’s counterpoint is fair: modern AI is increasingly shaped by workloads that are narrower, more measurable, and more open to specialized optimization than they were a decade ago. A custom ASIC inside a custom stack does not have to beat Nvidia everywhere; it only has to beat Nvidia enough on specific high-volume workloads to matter.
That is why the deepest synthesis is this: Nvidia is extraordinary because it built the leading **general-purpose AI platform** before almost anyone else understood the future. But being the leading general-purpose platform is not the same thing as being permanently unbeatable by specialized architectures. Jensen himself indirectly admits the world can change: when asked why Nvidia does not pursue radically different architectures in parallel, he says they could, but “we don’t have a better idea,” and adds that “if the workload were to change dramatically,” Nvidia “may decide to add other accelerators.” That is not the language of a man who believes architecture is settled forever.
The China section reinforces the same point from another angle. Jensen argues that China is too important to concede because it is the “second largest market in the world,” has “50% of the world’s AI researchers,” and because forcing China off the American stack could create a rival ecosystem optimized for non-American hardware. In other words, even Jensen’s strategic fear is not just about selling more chips next quarter. It is about whether future models are built to run best on Nvidia at all.
So the clean conclusion is:
Nvidia deserves enormous credit for building the platform that currently converts “electrons to tokens” better than anyone at global scale. But the interview does not prove that a broad, programmable stack will remain the single most efficient answer for every future LLM workload. The real question is not whether Nvidia is remarkable, the real question is whether its general-purpose dominance can stay ahead of increasingly specialized, custom architectures built for the workloads that matter most.
> the real question is whether its general-purpose dominance can stay ahead of increasingly specialized, custom architectures built for the workloads that matter most.
Very true. But think about the crazy new architectures (I have no clue) coming to support continual learning. Odds are Nvidia will reap benefits there and a lot of ASIC plans too narrowly focused on transformer inference will go back to the drawing board.
The debate of the week, in two pieces.
Dean Ball (Hyperdimensional): Mythos-level capabilities will diffuse in 1–2 years. Export controls are the only real lever.
Jensen Huang (Dwarkesh Podcast): China already has the compute. Conceding the market accelerates their ecosystem without slowing their capability.
Same facts. Opposite conclusions.
“The beautiful thing is you’re talking to the expert”. The level of arrogance to think such a thing, let alone say it?! And I lost track of the number of times he conflated the word “I” with the company he runs. Apparently he is NVIDIA.
I think Jensen probably argues with his employees like this, so I would take this as a positive that he felt comfortable arguing hard knowing that you could take it.
Sounds like Nvidia is in a defensive posture from the divergence in geopolitical, national incentives with their own.
Combined with rising pressure to maintain their previously free CUDA ecosystem monopoly, it feels like Jensen is basically doubling down on being so globally infrastructurally entangled that they don't need to pick winners (whether at the level of individual companies or national fronts), like Standard Oil playing the railroads against each other while making them dependents.
From all accounts Jensen seems bent on viewing GPUs as an apolitical general computing substrate rather than the enriched-uranium strategic resource that they really are, and rejecting all frames where NVIDIA is anything but a neutral participant in one layer in a big AI cake.
I think Jensen makes a 1000% valid response that we shouldn’t give up the open source market. That’s how US industries get stale and die.
The only way Dwarkesh’s take makes sense is if we are in a VERY critical few-year window (a la AI 2027). If we are in a few-decade window then Jensen is correct IMO.
I think Dwarkesh quite clearly believes that, and other proponents of export control (I.e Dario) do as well. I personally find it very hard to determine which regime we’re in, but exponential progress over the past year makes me think they are both correct.
The competence gap between Jenson in this interview and Lisa Su in the Stratchery interview a couple of years ago is insane. Just so, so wide. Jenson is throwing around the names of random open source projects, he's recalling bad design decisions made decades ago, he's fluently discussing minor algorithmic details, drilling down supply chain bottlenecks to the level of "we don't have enough plumbers".
Meanwhile Lisa Su was telling BT that AMD was doing great in AI, where's the problem? Lol.
It's incredibly telling that zero time was spent on AMD in this interview. TPUs/Broadcom are the only competitor that matter now.
https://kaipability.substack.com/p/w-jensen-i-you-we-lx-x-is-a-loser View from a Loser. Enjoy all.