Some podcasts are self-recommending on the ‘yep, I’m going to be breaking this one down’ level. This was one of those. So here we go.

As usual for podcast posts, the baseline bullet points describe key points made, and then the nested statements are my commentary. Some points are dropped.

If I am quoting directly I use quote marks, otherwise assume paraphrases.

As with the last podcast I covered, Dwarkesh Patel’s 2026 interview with Elon Musk, we have a CEO who is doubtless talking his agenda and book, and has proven to be an unreliable narrator. Thus we must consider the relevant rules of bounded distrust.

Elon Musk is a special case where in some ways he is full of technical insights and unique valuable takes, and in other ways he just says things that aren’t true, often that he knows are not true, makes predicts markets then price at essentially 0%, and also provides absurd numbers and timelines.

Jensen Huang is not like that, and in the past has followed more traditional bounded distrust rules. He’ll make self-serving Obvious Nonsense arguments and use aggressive framing, but not make provably false factual claims or absurd predictions. I think he mostly stuck to this in the interview here, but there are some whoppers that seem to be at least skirting the line.

I do not worry for Jensen Huang, only about him.

For full disclosure: I am a direct shareholder of Nvidia. I am long.

[Scheduling note: Weekly AI post will be tomorrow 4/17, with ‘knowledge cutoff’ at the release of Opus 4.7. Coverage of Opus 4.7 begins on Monday.]

Podcast Overview Part 1: Ordinary Business Interview

This was essentially an interview in two parts.

The first half, until about 57 minutes in, and I would also include the last few questions at the end in this, is about ordinary business questions. Why and how is Nvidia making these choices, these investments, these allocations of chips? Where is Nvidia’s moat? How do they think about these questions?

In these questions, there’s no doubt Jensen is talking his book and about how Nvidia is great. That’s what CEOs do, and maybe it’s a little thick, but aside from one stray swipe at so-called ‘doomers’ it’s fair play.

Jensen downplays TPUs as less flexible than GPUs, including that they lack CUDA, saying this will also matter for different AI architectures. I don’t buy that the edge matters so much for a large portion of business.

His explanation of how Nvidia allocates its chips seems disingenuous, and I do not centrally believe his account of this, but that’s the way such things go.

The most interesting part of the first half were his comments about Anthropic, and in particular how Anthropic ended up primarily training and running on Tritium and TPUs.

Jensen has nothing but good things to say about Anthropic, and he takes responsibility for letting this slip through his fingers and vows not to let it happen again. He figured Anthropic would get ordinary VC funding, because he did not understand the extent of their compute needs. Thus, in the early days, Google and Amazon invested and got Anthropic locked into those alternative chip ecosystems. He was happy to invest later, but Anthropic had already done a ton of work integrating and working with the other chips.

Jensen lost out on Anthropic partly because at the time he lacked the free cash, but mostly because he was insufficiently scaling pilled and AGI pilled. He understands this now, but he has not updated sufficiently. He still remains not very pilled, in any sense, on what is to come. He claims he can scale up his whole supply chain as much as he wants with a few years of notice, but keeps not scaling it up sufficiently. There will be power as a new potential limiting factor for chip sales within a few years but that wasn’t that importantly true before.

There is no hint, in this first half, that he thinks he is running anything other than an ordinary computer hardware business, except one scaling uniquely large and fast and profitably.

Podcast Overview Part 2: A Debate About Chip Exports

The half of the interview everyone is talking about is the second half, where they argue, often quite heatedly, about AI chip exports to China. Jensen of course wants to sell his chips to China, and Dwarkesh argues that we should not do this, while presenting this as a devil’s advocate position. My read is he mostly believes the things he is arguing, albeit with some uncertainty.

This is a high difficulty interview. Dwarkesh does a great job of engaging and not being afraid to push back. A bunch of it goes around in circles at times, but that seems unavoidable, and also was often revealing in its own way. Kudos for pushing.

Jensen tries to have many things both ways. His chips are way better, but China has all the chip manufacturing capability it needs, but it has unlimited energy with would-be data centers fully powered and sitting empty, but they can just use more worse chips, but America is so far ahead we shouldn’t worry about a few chip sales, but if we don’t sell those chips then we cede the world’s second largest market, and you both can and can’t switch model architectures, our sales would both not impact China’s compute access and be the difference between them staying on CUDA or not, and so on.

The biggest thing is that he repeatedly makes clear what he cares about.

What matters is Nvidia selling chips to China. That’s it. Nothing else matters. That keeps Nvidia and CUDA dominant, and what’s good for Nvidia is good for America, because if anything is built on his chips then that’s ‘good news’ and we win, whereas if it’s built on someone else’s chips, then that is ‘bad news’ and we lose.

This does not actually make any sense whatsoever. Whose chip is running the model and application is not the important thing and this should be very easy to see. But also there is no real competition in chip sales and won’t be for a long time, as everyone is compute limited and Chinese capacity to produce even much worse chips is severely limited.

By Jensen’s arguments, we’re sacrificing his layers of the ‘five layer cake’ that is AI to benefit the model layer and it is not fair, and it’s bad for America, because it means our ‘tech stack’ won’t win, and what matters is this mystical ‘stack’ that is actually code for the chips themselves.

Even if AI was going to indefinitely remain a ‘normal technology’ and ‘mere tool,’ and all we were dealing with was mundane AI, this would be wrong until at least such time as Nvidia can saturate market demand. Every chip made and sold to China is a chip not made and sold to America. Even after that, compute access will be key to economic productivity and technological advancement and also national security, even in these normal worlds.

If you understand that superintelligence is likely coming, and that everything is going to change and likely do so relatively soon, then the situation becomes overwhelming.

Especially poor was Jensen’s answer to the problem of cybersecurity and Mythos, which was that we need to have a dialogue with the Chinese and get them to agree to not use AI for bad purposes, presumably including cyberattacks.

I very much support entering dialogues with China about AI, and agreeing on things not to be doing, but in this situation that is obviously and hopelessly both naive and physically non-viable. The Chinese have a long history of doing such things after agreeing not to do them, so what is the verification method once we allow them to have the capability to do it?

Are you going to require them to heavily restrict and monitor all API calls? Cause that’s kind of the bare minimum, even if they can be trusted to want to stop doing it. It’s actually a lot easier to not develop the capabilities in the first place, but either way you need to lay foundations first, this takes time, and we have not done that.

Thus, yes, there is a huge divide here, where Jensen remains legitimately unpilled on the ideas of AGI and superintelligence, and doesn’t understand the thing his company is enabling to be brought into existence.

But also, even if Jensen were right about that, he would still be wrong otherwise, given the things we already know are possible. We are simply past the point where ‘AI as such a normal technology that you should just sell China to chips’ is a viable argument. We know it isn’t true.

Jensen only wants one thing, and it’s not disgusting but I also want other things.

I’ll cover reactions at the end of this post, once we have proper context.

What Is Nvidia’s Moat?

1. Nvidia makes software that TSMC and others use to make hardware, but why wouldn’t that get commoditized the same as other software? Jensen has been asked this one a lot, and responds with what is clearly a well-rehearsed speech. He gives three real answers: Demand is going to go up up up, software companies in general will thrive with tool use, and Nvidia’s particular task is extremely hard.
   1. Demand is definitely going to go up. That’s not in dispute.
   2. Software companies thrive with tool use if and only if they can continue to provide a unique product that is superior to the competition, and especially superior to new entrants and homebrews in valuable ways. It is not obvious which way this goes and he isn’t offering an argument.
   3. Nvidia’s task is indeed extremely difficult. The ability to use limited TSMC capacity to create modestly more powerful chips will only get more valuable, even if the competition could do a pretty good job. But this isn’t an argument for why in Glorious AI Future rivals can’t design chips that are as good.
   4. I still am happy (not investment advice!) to be long Nvidia, but this doesn’t show us much of a moat yet.
2. Nvidia has ~$100 billion in purchase commitments, and soon will have $250 billion, locking up scarce components. Is that Nvidia’s moat. Jensen says they make big commitments, including getting other companies to make big investments by showing the future size of the market, which he spends a lot of time doing. They have the supply chains and the cash flow and the churn.
   1. I buy that all of that is great investments that help a lot, and that new competition would struggle with various parts of this.
   2. I don’t think this would be a sustainable moat over time, if there was serious competition, but in the medium term it’s a big edge.
3. Can Nvidia keep doubling revenue and tripling flops provided year after year, or are we hitting capacity walls such as at TSMC? Jensen notes anything can be a bottleneck, the hardest is actually electricians and plumbers, but they’re scaling the hell out of everything, and all the bottlenecks get attention. Any given bottleneck can be scaled within two or three years given a demand signal.
4. He wants to ‘reindustrialize the United States.’ He needs energy, but the other stuff is all 2-3 year problems.
5. “This is one of the concerns that I have about the doomers describing the end of work and killing of jobs. If we discourage people from being software engineers, we’re going to run out of software engineers. The same prediction happened ten years ago. Some of the doomers were telling people, “Whatever you do, don’t be a radiologist.” You might hear some of those videos still on the web saying radiology is going to be the first career to go and the world is not going to need any more radiologists. Guess what we’re short of? Radiologists.”
   1. This is very clearly a case of ‘doomer’ being used as a slur in order to dismiss anyone concerned about any negative impact of AI via association and vibes.
   2. This also has some valid points, but it is incoherent. We must unpack.
   3. There are two kinds of They Took Our Jobs concerns, which this conflates.
   4. First is the ‘end of work’ in general and killing jobs in general, and worries about mass unemployment and declining wages. He says he is addressing this concern at first, but then pivots and doesn’t actually talk about it. As I’ve said before, I think some are too concerned about this, but with sufficient capability this becomes a big worry.
   5. What he mostly discusses is predictions that particular jobs will suffer from local technological unemployment.
   6. There are clearly some cases where this is true for AI today. If you told someone ten years ago to become a translator, you did them dirty.
   7. Radiologists were an interesting case, often discussed. Those warning about this were right that AI would be superhuman at analyzing images.
   8. But this caused an increase in demand for radiology, and AI can’t replace many other parts of the job, and because in the longer run radiology is going to be increasingly automated and doctors have 40 year careers, many opted out of radiology.
   9. So for now, in 2026, we have a shortage, and radiologists earn a lot. However, in the longer run, it seems likely demand for radiologists will decline as a percentage of demand for doctors. Standard economic theory says that this means we should currently have a shortage of radiologists.
   10. Thus, it’s not clear the shortage is even inefficient. But to the extent that it is and we made a collective mistake, it was that it was a specific bad prediction about this particular profession, which has a many years lag in training.
   11. Moving on to software engineers, we should worry less here about both errors, because especially now with agentic coding the supply of coders is elastic. You can get going relatively quickly. My guess is we will want more engineers for a while, not less.
   12. This shouldn’t be a ‘doomer’ or ‘decel’ versus ‘optimist’ or ‘accelerationist’ thing. This is an allocation problem, where you have to be forward looking, and you do the best you can, and who is right about the big picture does not have that much say in who is right about the specific choices.

TPU vs. GPU

1. TPUs trained Claude and Gemini. What does it mean? Time for another speech. Jensen pitches TPUs as a narrow product whereas GPUs accelerate all sorts of computing, so they have much wider market reach. You can do it yourself or rent, and do things TPUs can’t.
   1. This raises the question of why compute is not more fungible. xAI, which Jensen mentions, has these huge arrays of GPUs, but no one wants their inference, so why aren’t they renting out that capacity? Or are they?
   2. I buy that GPUs have lots of applications TPUs can’t touch. I won’t be using a TPU to power my monitors, after all.
   3. But if AI is the dominant reason to want compute going forward, and TPUs are fungible there with GPUs, then won’t TPUs end up competitive for a large portion of the space?
   4. Jensen’s arguments didn’t address this, and it was the central implied question, so Dwarkesh asks more explicitly.
2. Dwarkesh points out the $60 billion in profits per quarter for Nvidia is mostly from AI, not quantum and pharma. With that, why do you need the flexibility of a GPU? Jensen says, sure matrix multiplication, but you might want to use other techniques as well. He brags about getting 50x energy efficiency with Blackwells over Hoppers. MoEs are one such innovation.
   1. Jensen is saying 50x more efficient per unit of compute, or for the same software task, not for chip versus chip. Power is still a limiting factor.
   2. MoEs were invested by Google on TPUs, so clearly they can do MoEs, although if you are not Google or Anthropic you might need CUDA.
   3. Google could close most or all of this effective gap if it cared to open source its own internal TPU kernel libraries, but they don’t want to do that, and would rather try to use their TPUs to win in AI rather than selling chips.
   4. Is Google right about that? Unclear, but selling a ton to Anthropic is a weird middle path that likely reflects infighting between Cloud and DeepMind.
   5. The point being, Nvidia’s moat against Google in AI chips is… Google, mostly?
3. 60% of Nvidia revenue is from the big five hyperscalers. Do they need CUDA? OpenAI has Triton, Anthropic and Google run their own accelerators. Jensen gives the ‘happy to help with all frameworks’ and also the ‘CUDA is super flexible with a huge install base and every cloud provider’ talking points.
   1. Okay, sure, nothing surprising but solid.
4. Do those advantages matter to the important customers, though, enough to protect +70% margins? Nvidia has lots of engineers optimizing everyone’s stacks, and we’re talking 2x improvement or more. He taunts TPU and Trianium for not getting measured via InferenceMAX, claims the supposed TPU 40% edge doesn’t make sense and is probably fake.
   1. Nvidia has real advantages but Jensen is overplaying his hand a bit here.
5. Jensen says all this ‘competition’ is really Anthropic: “Anthropic is a unique instance, not a trend. Without Anthropic, why would there be any TPU growth at all? It’s 100% Anthropic. Without Anthropic, why would there be Trainium growth at all? It’s 100% Anthropic. I think that’s fairly well known and well understood. It’s not that there’s an abundance of ASIC opportunities. There’s only one Anthropic.” And OpenAI might be building Titan but they’re ‘vastly Nvidia.’
   1. The claim is basically ‘Anthropic is the weird exception, other AI companies would never, that’s the only reason those chips have meaningful sales.’
   2. Anthropic is proof of concept, but if you need long term investment and scale and even deep TPU familiarity on day one before it makes any sense, then maybe?
6. Jensen basically blames Anthropic being on TPUs on his inability at the time to invest early on in Anthropic, whereas Google and AWS invested. He’s not going to make that mistake again.
   1. I don’t think this partly a ‘Amazon and Google bribed Anthropic to get their business’ but mostly a ‘Nvidia failed to bribe Anthropic.’
7. Dwarkesh points out that with 70% margins, you can be a lot worse than Nvidia and still come out ahead if you roll your own. Jensen fires back that ASIC margins at places like Broadcom are similar, ~65%, anyway.

Why Isn’t Nvidia Hyperscaling?

1. Jensen says Nvidia scaled as soon as they could have, and invested in the labs as soon as they could. There wasn’t enough cash and he figured the labs would raise from VCs. He’s happy Anthropic exists even though they raised from Google and Amazon.
   1. This is the trader’s lament. If it was a good trade, you should have done more, and you should have done it earlier.
2. But what about now with all the piles of money? Why not be a cloud provider? That’s not Nvidia’s business or philosophy. If others can do it, you let them do it.
   1. People underestimate the importance of staying focused.
   2. I totally believe that Nvidia made the right call here, if you don’t think superintelligence is going to render anyone but the AI labs powerless. Invest in all the model companies, lock in as much business as possible, win no matter who wins, don’t try to be a model company or a cloud provider.
   3. If you do think only AI labs matter, a la Musk, then big mistake. Oh well.
3. Why doesn’t Nvidia ‘pick winners’? Not their job. Let them fight it out.
   1. I would add, the competition helps Nvidia.
   2. Nvidia of course ‘picks winners’ in another sense, by choosing magnitudes of investments, and choosing valuations, and choosing allocations, it just tries to do so in a way that keeps the competition flowing.
   3. If Nvidia truly didn’t want to pick winners it would allocate fully via price.
4. Nvidia ‘doesn’t want to be in the financing business,’ but of course they will help OpenAI with $30 billion when they need it, it’s a great investment. They don’t ‘just want to prop up neoclouds’ or hyperscalers or labs.
   1. They’re in the financing business. That’s the financing business.
   2. That said, I do not think Nvidia is doing it to prop up otherwise unpromising businesses or making bad investments. I think Nvidia is using it to secure business deals while also making otherwise good investments. Win-win-win.
   3. They are in the side of the financing business where first you have to prove that you don’t need the financing. Which is most of the financing business.
5. Both agree: There is a shortage of GPUs.
   1. This will be important later.
6. How does Nvidia divvy up scarce allocations of GPUs? First the customer has to place a purchase order. Then it is first in, first out. Larry and Elon (brought up unprompted) never begged for GPUs. It’s all just placing an order.
   1. I don’t have any insider information, but this smells like straight up bullshit.
   2. There was for a long time a truly massive GPU shortage, with demand many times the size of supply at Nvidia’s price points. Allocations were existential.
   3. If you were really just indifferent, you’d raise prices.
   4. If it was fully first in, first out, then the allocation pattern looks very different.
   5. Even if Jensen isn’t going to listen, of course Elon Musk was going to try whatever he could to beat the system same as everyone else, only more so. Maybe he didn’t ‘beg’ per se, but that is a classic Suspiciously Specific Denial.
   6. This is not consistent with the story he told about Anthropic.
7. Why not highest bidder? “Because it’s a bad business practice. You set your price and then people decide to buy it or not. I understand that others in the chip industry change their prices when demand is higher, but we just don’t. That’s just never been a practice of ours. You can count on us. I prefer to be dependable, to be the foundation of the industry. You don’t need to second-guess. If I quoted you a price, we quoted you a price. That’s it. If demand goes through the roof, so be it.”
   1. Every economist is screaming right now.
   2. If you can’t be depended on to deliver product, you’re not dependable.
   3. If you are pure FIFO at a fixed too-low price, you often won’t deliver.
   4. Yes, I agree that if I quote you a price, that’s the price even if demand goes up. But Nvidia’s prices for years were lower than market clearing.
8. They have a great relationship with TSMC. They fight, and sometimes there’s some ‘rough justice’ but you can count on TSMC to be there every year, and for Nvidia to be there with a new product every single year. Both of them can scale as high or low as you need, you just need to place an order.
   1. If this wasn’t true, he would still say the same thing.
   2. It is a weird situation, where both sides need each other and have to divide the profits, with a huge potential ZOPA, but yes ultimately they make it work.
   3. The giant piles of money? They help.
   4. It makes sense, given the explosive growth, for there to be somewhat of a shortage most of the time. Same ‘mistake’ as Anthropic.

Selling Chips To China

So far, they’re spent an hour asking Jensen standard business questions, and he’s provided mostly standard business answers. No one is talking about that hour.

This next part is the part everyone online is talking about. Export controls.

1. Dwarkesh presents himself as devil’s advocate. He’ll take the anti-export side.
2. What about Mythos? Wouldn’t Chinese companies being able to train something like Mythos, especially first, threaten American national security?
   1. I think we can all agree that would be a very scary scenario, as a specific example of us not wanting China to have the most capable models.
   2. There are other classes of cases for export controls as well.
3. Jensen starts out by saying Mythos was trained on ‘a fairly mundane amount of fairly mundane capacity’ by an extraordinary company. China has such capacity. He ‘wants the United States to win,’ but don’t make them your enemy, they’re too capable, you see. “They manufacture 60% of the world’s mainstream chips, maybe more…. They have 50% of the world’s AI researchers.”
   1. Dwarkesh offers a distilled recap on this and the next few items here.
   2. Okay, look, functionally this is just straight up bullshit.
   3. Yes, technically the stats here are probably true, but he’s trying to say ‘China has all the chips it needs,’ which is false, and just as much talent as we do, which is false (number of researchers is a really dumb measure of talent and capability) and ‘therefore we can’t beat China, we have to make a deal.’
   4. He does not want the United States to win. Or at least he doesn’t much care. Indeed, he’s falsely saying that we can’t.
4. He wants ‘research dialogue’ to make a deal on what not to use AI for.
   1. I do want a dialogue with China on AI safety issues. I strongly agree that it would be good for our researchers and AI people to be talking.
   2. Agreeing to put aside some uses of AI is good. The first step of ‘no AI in command of nuclear weapons’ is clearly good.
   3. Extending that to cyber attacks would also be good, but in what sense do we not already have an agreement to not do cyber attacks in the first place?
   4. And in what sense is China not blatantly ignoring that rule all the time?
   5. So why would we expect China to hold to such a deal if they had the AI capabilities to do the cyberattacks? Are you going to do real verification?
   6. Making a deal to not use AI for [X] is much harder than making a deal to ensure AI can’t do [X] in the first place. Making a broad general agreement can actually be much easier than making narrow agreements.
5. Jensen talks about how open source and the startup ecosystem are vital to cybersecurity, that ‘the ecosystem needs open models’ to do the work, and that a lot of the cybersecurity work is coming out of China. He says “The idea that you’re going to have an AI agent running around with nobody watching after it is kind of insane.”
   1. There are AI agents running around with nobody watching after them.
   2. There are going to be a lot more of them. Deal with it.
   3. Is that insane? Mu. But it’s happening.
   4. As for the other stuff, it’s mostly non sequiturs, and it’s not the future of cybersecurity, and he certainly isn’t beating the rumors here.
6. Dwarkesh pushes back. The Chinese chips are 7nm at best. They have 10% of the flops we have. Anthropic getting there first and getting to do Glasswing was kind of important. Once such a model is out there, amount of compute matters a lot. All the labs are bottlenecked on compute, both in America and in China.
   1. He seems straight up correct about all of this.
7. Jensen responds that yes we should always be first and always have ‘more’ compute, but China has ‘enormous’ amounts of compute, the second largest market in the world, and they could aggregate that.
   1. Enormous is relative here.
   2. They could aggregate in theory, but they won’t for obvious reasons, and if they did then that would make the whole thing even scarier.
   3. Jensen keeps simultaneously saying ‘we have an edge in compute’ and also ‘but they have enough compute’ and also (later in the interview and also constantly all the time in general) ‘we should give away a large portion of that edge in exchange for me making more money.’
8. Jensen goes to energy. China has all the energy. “Why can’t they put 4x, 10x as many chips together, because energy’s free? They have data centers, fully powered, sitting empty. The idea that China won’t be able to have AI chips is complete nonsense. 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. So the idea that China won’t be able to have AI chips is completely nonsense.”
   1. This is honestly pretty embarrassing for Jensen.
   2. He’s trying to argue that the Chinese don’t need his product, shouldn’t even want his product, have all the chips they need, worse chips can do the same job totally fine, China is over capacity on chips.
   3. This is simply flat out false. It’s very obviously completely not true.
   4. I try not to say this kind of thing lightly, but yes, Obvious Nonsense.
   5. We know this because we see huge actual effective bottlenecks in compute for everything the Chinese are trying to do in AI.
   6. We know this because before controls China had about as much compute as we did, and now they have 10% as much compute as we do.
   7. If China can do all that and has all the SCIP they need and so on, then why are those data centers that are fully powered sitting empty? Why does no one have enough compute? Where are all the foreign Huawei data centers with all their extra chips they don’t even need, that Sacks told us were coming if we didn’t make the right deals? Come on, now.
9. Jensen goes on to talk up Huawei. Biggest year in history. They shipped a ton of chips. Millions of chips. Way more chips than Anthropic has. They have plenty of logic, and they have plenty of HBM2 memory. They don’t need EUV for the most advanced HBM, they’re a networking company. Algorithmic improvements are what counts, anyway.
   1. Jensen is clearly flailing here. It reads like tilt, and anger, and desperation, and doubling down on a story that makes no sense. Throwing words at the wall and seeing what sticks, just deny deny deny.
   2. Would be funny to intercut this with when he’s talking up Nvidia.
10. We get the ‘tech stack’ argument. “DeepSeek is not an inconsequential advance. The day that DeepSeek comes out on Huawei first, that is a horrible outcome for our nation.” Dwarkesh flat out asks, why? Jensen says, suppose it is ‘optimized for Huawei.’ Our hardware would be at a disadvantage.
    1. I don’t even know what to say at this point. Why do we care which hardware that particular model is a little more efficient running on? It makes no difference. This is dumb. Nvidia is selling every chip it can make, will do so for a long time, and Jensen does not dispute this.
    2. Jensen is outright saying the bad outcome would be if Nvidia were put at a particular competitive disadvantage. That kind of gives the game away, and he’s about to give the game away a lot more.
    3. Oh, also this obsession with DeepSeek in particular continues, as Jensen sees it once again as an ‘important advance.’ This is tactical. DeepSeek is a good lab especially given its severe hardware limitations, but their big triumph is learning how to get remarkably far while being starved for compute, and they likely haven’t had the best Chinese model in some time, and as I’ve explained repeatedly the ‘DeepSeek moment’ was badly overhyped.
11. “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.”
    1. Mic drop. QED. Rest my case. Mask off. Thank you, sir.
    2. As in, he thinks that if the Chinese develop the best model, so long as it runs best on his hardware, that’s good news. That’s a win.
    3. Whereas, if the Chinese develop a model that isn’t the best, but it runs better on Huawei chips than Nvidia chips, then that’s bad news. That’s a loss.
    4. All he cares about are Nvidia’s hardware sales. Stop pretending otherwise.
12. Dwarkesh asks, can’t models just swap accelerators anyway? Jensen says no, and ‘I am the evidence.’ Nvidia’s success is perfect evidence. Dwarkesh points out people do it anyway, Jensen says they don’t run better. Anthropic’s models run on Trinium and TPUs, but ‘a lot of work has to go into that change.’
    1. “L’Preuve, c’est moi.”
    2. This is his whole attitude throughout. The authority has spoken, peasant.
    3. Yes, of course they don’t run better when you do a straight swap. Nvidia’s chips are better and yes there is some value in optimization. But the point is that in most cases the efficiency loss is moderate, so you use what you’ve got.
    4. Jensen didn’t address that Anthropic is running its same models on three distinct hardware architectures and it’s going fine. You do the work.
    5. The work is going to get easier because the AIs can do the work.
13. “But 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.”
    1. Okay, this is just tilt now. No one said that, on several levels. You mad bro?
    2. Let’s break it down.
    3. Where did ‘all of the AI models’ come from here? We’re discussing the possibility of some AI models being optimized for non-USA hardware. The most important models, likely the best models, would not be in that group.
    4. The ‘tech stack’ here no longer includes the AI model. The point of the ‘tech stack’ is that it includes both the hardware and the model, so this isn’t even a real tech stack.
    5. There is a huge difference between ‘runs most efficiently per use of compute on non-USA hardware’ versus ‘runs best on non-USA hardware.’ The Nvidia hardware is better than the non-USA hardware. So even if there is some substantial efficiency loss in the swap, you would still benefit from the large gap in performance.
    6. He is agreeing this only applies ‘out of the box’ without ‘doing the work,’ but in the future very obviously someone else will do the work and you’ll be able, with help from your AI, to benefit from them having done the work, if that work is valuable.
    7. Jensen just got done arguing that no, it doesn’t matter how good your chips are, you can just string together a lot more chips and everything is fine.
    8. It’s kind of funny that this response against the Nvidia CEO is largely me talking up Nvidia chips while he talks them down.
14. “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.” “It’s completely… There’s evidence right now. Their chip industry’s gigantic.”
    1. I don’t know how else to say this, except that Jensen is at best bullshitting.
    2. He’s saying that whether or not Nvidia sells compute to China will not impact how much compute China has. He argues even with this very minimal claim.
    3. At what point do we agree to acknowledge who and what this person is?
    4. It goes on like that for a while without saying anything new, until we get another moment.
15. “Listen, 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?”
    1. Any questions?
    2. Dwarkesh is perhaps partly at fault here, tactically, for not yet emphasizing the other reasons why one might want your country to have a lot more access to compute, especially compute priced well below fair market price, that go beyond the direct training costs.
    3. Dwarkesh is still the best interviewer out there, trying to play a very difficult hand. The witness is hostile and uncooperative and unreliable, and a lot of what he’s doing is actually getting two hours to talk to the witness and even argue directly with him without Jensen storming off. It’s a hard job, the same way it is a hard watch or read.
    4. Thus, even in the world where superintelligence and decisive strategic advantage are not things, he’s failing to understand that the economic impacts of AI are ultimately about who uses what inference for what purpose, in what quantities. Nvidia is the most valuable company in the world but even in the ‘AI as normal technology’ worlds ultimately his share of the profits is, in relative terms to the application and model layers, bubkas. The applications succeed because you have the models and compute you need to develop, deploy and run the applications.
    5. Very obviously, again even in ‘normal technology’ worlds, selling Nvidia chips to China doesn’t benefit the ‘American tech stack’ or help the rest of the layers of this supposed cake. It would run on Chinese energy, running Chinese models for Chinese applications. And then, if this really is a normal technology world, the Chinese chips, likely designed by AI using Nvidia chips, eventually replace the American ones once they catch up on that.
    6. The ‘one company’ thing is a bizarre thing to say, as if this is purely about Anthropic. It obviously isn’t, and Anthropic mostly doesn’t even use Nvidia chips. It simply happens that Mythos is the example of a capability advance. Very obviously the same logic applies to OpenAI, Google and xAI on one side, and the Chinese companies on the other.
16. Dwarkesh tries again to talk about how much better Nvidia chips are, and how while China is struggling to scale 7nm Nvidia is moving on to 3nm and then 2nm or even 1.6nm. And he points out that ‘China has limitless energy’ is an argument that every chip you sell to China is that much more compute China has, since there is no other limiting factor.
    1. Good talk, and good attempt.
17. “Listen, I just think you speak in absolutes. I think 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. Okay. The United States is ahead.” … “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?” He says, as long as only America gets Vera Rubin, how is that not good enough?
    1. Jumping in this context to a claim of 100x is wild.
    2. Even 10x completely steamrollers Jensen’s earlier arguments about China not needing more compute, if you think about it.
    3. Again, all he cares about is his market share, and thinks this is ‘giving up the world.’
    4. This also implies that there is a set of fixed markets being competed for, rather than there being a fixed supply of chips where no one has enough.
    5. How about an obvious compromise, where if Nvidia can make enough chips to meet Western demand then we can talk about selling the rest? No, he strongly opposes that sort of thing as well.
18. Jensen calls any comparison of AI to nukes or missile casings ‘lunacy.’ He calls comparing compute to uranium a ‘lousy’ and ‘illogical’ analogy. No argument is offered. When asked about the zero-day exploit issue, he says you solve that via ‘dialogues to make sure that people don’t use technology in that way.’
    1. He’s resorted to flat out name calling at this point.
    2. On the question of cybersecurity, I repeat my logic from above, and just want to emphasize how obviously naive this answer is. China releases its models open source. Even if you got China to agree not to do cyberattacks, and even if you got China’s government itself to not do the attacks themselves, and even if you got them to try and enforce this within China, then what? How are you going to enforce it? Even if you enforce it within China somehow, what happens when the North Koreans DGAF, since they obviously DGAF?
    3. Again, one cannot simply ‘agree not to use AI for [X],’ unless there are a highly limited number of actors who could do [X], such as when it involves existing nuclear weapons. You have to not permit that capability to exist in the first place, or at minimum you have to provide extensive monitoring of the relevant sources of that capability, worldwide. Please take this seriously, sir.
19. He then comes back to saying “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.”
    1. Dwarkesh is very much not making the argument that not selling chips to China is good for our long term ability to win the chip layer in particular.
    2. Jensen keeps emphasizing this because that is the only thing he cares about.
    3. Since Dwarkesh is not making the argument here, I suppose it is up to me to make the argument. So let’s do that, in two parts.
    4. First off, Nvidia can already sell every chip it can make, and also Huawei can also sell every chip it can make. If you sell a Nvidia chip in China, all that does is physically move that chip to China. If you do that often enough for long enough and scale up fast enough, then yes, that would change, but that’s not the situation.
    5. Thus, it is not obvious at all that selling chips to China would change the medium or long term chip situation at all, and it almost certainly would not impact anyone’s short term chip sales. Except insofar as Nvidia intentionally made chips intended only for Chinese consumption, instead of making chips for America.
    6. China highly values self-sufficiency on chips. I would value other things relatively more, but this is a very sensible thing for them to be caring about, and they are not about to let this go. They have also shown a willingness to restrict Nvidia sales inside China, towards this goal. Thus, we should conclude that if in the future Nvidia sales in China were threatening Huawei’s ability to make and sell more chips, that China would intervene to favor Huawei. To the extent Nvidia’s sales will matter here, they will be stopped. Even in this context, you only get to make sales that are a mistake.
    7. In the long run, a key limiting factor on everything is intelligence and compute, and the ability to solve various problems and create superior designs. Again, this is true even in the ‘AI as normal technology’ worlds that skeptics like Jensen say they expect. If you sell China a lot of chips, and they have better AI models and more compute with which to run them, they then use those better models more often to create better chip and EUV designs, the same way they advance everything else.
    8. Meanwhile, those sales supercharge China’s economy at the direct expense of our own, which also hurts our ability to do everything and helps theirs.
    9. So no, this is not ‘just a fact’ even on its direct level. It is highly plausible that holding back AI chips helps you in the long run market for AI chips.
20. Dwarkesh engages on the narrow chips question, noting that Tesla and iPhones didn’t get lock-in in China. Jensen doubles down on ‘what matters is the richness of our ecosystem.’
    1. One could also cite numerous cases of technology transfer, reverse engineering and so on, as arguments against letting them get the chips.
    2. ‘Our’ here means Nvidia and CUDA. He doesn’t care about the models or applications or economic activity being Chinese, because that’s not ‘our.’
21. He seems very insulted by the comparison to a car, Nvidia is not a car, you cannot ‘buy this car brand one day and use another car brand another day, easy.’
    1. Well, actually, yes you can, and people do. Not losslessly with no notice, but yeah, people do this all the time.
22. The hits keep coming: “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.” “You don’t have to move on. I’m enjoying it.”
    1. “I simply can’t acknowledge that.” No, you can’t.
    2. “Is not a loser.” “That losing mindset.” Very telling. Someone hit a nerve.
    3. This man leads the most valuable company in the world, that sells out all of its products, and he’s terrified of being a ‘loser.’
    4. But he actively wants to keep this going, even when Dwarkesh realizes this is going in circles into tiltland.
23. “And I just want you to acknowledge that any marginal sales for the American technology industry is beneficial.” “The logic that you use, you might as well say it to microprocessors and DRAMs. You might as well say it to electricity.”
    1. Jensen doesn’t want to understand this. He thinks ‘America sells thing’ should just be seen as good for America or the American tech industry.
    2. And he claims this is on the level of obvious, one could not argue with it.
    3. But very obviously this argument proves too much, and it is not made of gears. Why does this marginal sale net benefit the rest of the tech industry?
    4. Jensen’s argument seems to rely on us being ‘far enough’ ahead that it’s fine to give some of that back, while at other times he argues for the opposite.
    5. Yes, it is a correct default assumption that any given marginal sale is good, if you don’t have any other information. Here we do have a lot more information.
24. “We have tons of compute. We have tons of AI researchers. We’re racing as fast as we can.”
    1. We could have more compte.
    2. We could have more AI researchers, if we had more compute and if we had more willingness to brain drain Chinese and other talent.
    3. Thus we are not racing as fast as we can.
    4. To be clear: This is not me saying we should race, or race as fast as we can.
25. [More of the same arguments going back and forth, with Jensen continuing to say contradictory things and continuing to insist there is no contradiction.]
    1. Including saying American telecommunications industry was ‘policed’ out of basically the world, which is not a good word for what happened even if you buy the mercantilist thesis, nor is the situation a parallel.
    2. Dwarkesh says ‘I’m trying to make you understand the cost of selling the chips’ and Jensen responds by once again repeating what he sees as the cost of not selling the chips. As in, no, I’m not interested, sir.
26. Jensen continues to think that the AI ‘application layer’ is the one that matters most, not the model layer.
    1. But even if that’s true, then that’s still a reason to hoard the compute.
27. Jensen keeps talking about ‘losing the world’s second largest market’ for the entire tech stack. He seems to continuously claim: If the Chinese use CUDA and Nvidia, then our tech stack ‘wins’ the Chinese market in a meaningful sense for the model and application layers that matter most.
    1. And I’m here to say, no, this makes no sense, even in ‘normal technology’ worlds, it does not matter very much whose chips are being used if the models and applications are Chinese.
    2. I don’t understand why, other than Nvidia’s profits, this is hard to understand.
    3. Then again, that is exactly why we have a saying about it being difficult for a man to understand something in such scenarios.
    4. Actually I think Jensen understands perfectly well and is pretending not to.
28. Jensen makes the good point that if we scare everyone in America into hating AI and away from doing software engineering, then that would not be good for us. He goes back to radiology, the difference between a job and a task.
    1. Making Americans hate mundane AI use, and fear the impact of mundane AI (or AI as normal technology) in our lives, to the point where Americans refuse to diffuse it and use it to our benefit, would indeed be a massive mistake. We should work quite hard to avoid this.
    2. America disliking mundane AI has, if you look at the data, very little to do with Americans fearing AI existential risk, or even the catastrophic risks that people like me worry about.
    3. Americans do worry about those things when prompted, and often unprompted, but this is low salience.
    4. What ordinary Americans mostly care about are things like job losses, the internet filling with slop or deepfakes, environmental impacts and so on. This is unfortunate, and I try to discourage it, but this is our reality. And this has very little to do with the question of export controls or AI existential risk.
    5. Who is discouraging people from being software engineers? I’m not sure. I think it is mostly people trying to think about the economics.
    6. Going back to the radiologist thing, I would refer back to my earlier analysis, and also note that the shortage is largely caused by regulatory capture, in that we require doctors, and in particular radiologists, to take performative actions. We could, if we wanted to, now train a lot more radiologist assistant practitioners, or whatever we wanted to call them, that could do the remaining parts of the job while relying on AI, if we decided to legalize this. And perhaps the shortage eventually speeds up when we do that.
    7. I don’t think any of this bears on the actual questions being asked by Dwarkesh, but they’re things relevant to our interests here, and when Jensen makes a good point I should highlight it, since I’m hammering him a lot.
29. Jensen points out that lithography advances are maybe 75% improvement from Hopper to Blackwell, so the Nvidia architecture is most of the 50x total gains.

Okay, thus endeth the key section everyone is talking about.

Different Chip Architectures

1. Why doesn’t Nvidia also make more modern versions of N7 chips or similar? Jensen replies it is not necessary, then gives the real answer of R&D costs.
   1. I buy this. Focus is crucial for a company like Nvidia. Better to spend all your engineers in making the next chip ten times better.
2. What about completely different chip architectures? They don’t have a better idea, they simulate the other options, they’re worse. He’s folding Groq into CUDA, tokens are worth paying for now, and he’d like to invest more in Nvidia architecture.
   1. All seems fair, except it’s odd to say ‘if I had more money’ as the head of Nvidia? Seems like he should have all the money he needs for this?
3. Where would Nvidia be today without deep learning? Accelerated computing.
4. Jensen reiterates that enjoyed that interview.
   1. I can believe it. Even though he seemed highly frustrated and tilted, how often does someone like Jensen get to have a real argument? How many people actually push back? It can quite the relief, in its own way.
   2. I’m going to see Jensen accept an award next week, because I randomly got invited.

The Online Reactions On Export Controls

Daniel Eth and Connor Williams are among those who view Jensen’s arguments against export controls as fully amoral, purely about making money, and as not remotely making sense. There were also many others.

Dmitri Alperovitch: Incredible interview with Jensen. He blatantly admits that his jihad against export controls is simply all about Nvidia selling more chips worldwide, not about national security or winning the AI race against China (which he previously said doesn’t even matter if we win)

I think such reactions are about one notch too harsh. But basically yes, these are the strongest arguments Jensen can make, and they are quite weak.

Tenobrus: jensen in the dwarkesh interview isn’t “wrong” per se. he simply does not care about the truth. he cares about selling as many Nvidia chips as possible, whatever the consequences. he’s very visibly engaging in motivated reasoning to justify this. why would we expect different?

This would lower my opinion of Jensen as a thinker and communicator and humanist and american if it were already high on any of those dimensions. but it’s not. my opinion of him is high as a businessman, and as a businessman he wants to sell chips to china.

deeply appreciate @dwarkesh_sp for pushing as hard as he did here. i hope the visibility of Jensen’s incoherence on this makes it harder for Nvidia to justify themselves going forward

Alec Stapp: This is the key moment between Jensen and Dwarkesh on export controls:

1. Dwarkesh asks why it’s okay to sell NVIDIA chips to China given the national security implications of AI models like Mythos.

2. Jensen gives a misleading answer, arguing that it’s okay to sell American chips to China because China already produces 60% of the world’s chips.

3. But as Jensen definitely knows, compute is measured in flops, not number of chips.

4. Dwarkesh then pushes back, pointing out that on a flops basis, China has 10% of the compute the US has, and giving them more compute would change their cyber capabilities.

This exchange shows why it’s critically important for interviewers to have at least some technical knowledge, so they can push back against misleading talking points.

Alex Imas: Jensen has been doing what seems like a 24/7 interview cycle for months, and the number one question from the beginning should have been this exact exchange.

I don’t know if it’s the decline of old media—where journalists are just not pushing and asking questions in the same “investigative” style that they used to—or something else. But I’m glad we have @dwarkesh_sp to do the research and shine the light.

William Buckskin: I like Jensen, but this is exactly why we have government.

He’d sell us out for China for his investors. We obviously can’t allow that to happen

I don’t overly begrudge Jensen being a capitalist who will sell to whoever wants to buy, and leaving it to others to decide to whom he is permitted to sell. The issue is that he keeps trying to mislead us to get permission.

There are those in these exchanges who attempt to defend Jensen, you can find them if you click through, but I also found those arguments quite poor. This from Ed Elson was the most serious attempt I’ve seen, but his own metaphors go in the other direction if you think them through.

Here is one full explanation, responding to the distillation.

Peter Wildeford: Jensen here is frustrating and wrong. The man wrote off billions so of course he opposes controls.

1. Mythos is a ~10T parameter model trained on Amazon Trainium. Despite Jensen’s best efforts, China doesn’t have ]Blackwell or similarly capable] chips thanks to export controls.

Huawei’s best chip delivers 1/3 the per-chip performance, at 2.5x the power cost, with yields >12x worse. Jensen calling Mythos “fairly mundane capacity” that’s “abundantly available in China” is just plainly false.

2. Dwarkesh is right that the compute ratio matters geopolitically. Maintaining a capability lead during the critical window — even 12-18 months — is the whole point of controls. The difference between China running a thousand vs. a million offensive AI agents is huge. Jensen dodges this entirely.

3. Jensen can’t simultaneously argue “controls failed because China innovated anyway” (DeepSeek) AND “we must sell to China or they’ll leave our ecosystem.” If they’ll innovate regardless, selling chips doesn’t buy the loyalty he claims.

4. Jensen’s ecosystem stickiness point (x86, Arm) is his strongest argument, but it cuts against him: the world is already locked into CUDA. Selling Nvidia chips to China doesn’t deepen that – it just gives China better hardware while they build Huawei alternatives regardless.

An obvious point several people hammered, that I also noticed: If China has the energy to use unlimited chips, that’s all the more reason not to sell them the chips.

Theo Bearman: Jensen on China: “The amount of energy they have is incredible. Isn’t that right? AI is a parallel computing problem, isn’t it? 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. You know 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.”

This is exactly why we need to ramp up export controls across all elements of the semiconductor manufacturing stack rather than help the Chinese maximally leverage their advantage in ready-to-deploy powered shells with leading American GPUs and “50% of global AI researchers” to boot. The US doesn’t currently have that luxury, with long lead times for power and cooling components, permitting and data-centre buildout.

With UKAISI now saying AI capabilities are doubling every four months, the net result of Jensen’s strategy to try and get China hooked on the American tech stack will be one thing: the surrender of Western AI advantage, perhaps for good. Sure, there might be downsides to going heavy on export controls, but the alternative is much worse.

Peter Wildeford: Jensen apparently was also unintentionally making the case *for* export controls on Dwarkesh:

“They [China] have datacenters that are sitting completely empty, fully powered. You know they have ghost cities, they have ghost datacenters too.”

Imagine if China had the chips!

Peter also explains that Huawei can currently match 1%-4% of China’s market demand, and that China’s government is going to ensure unlimited demand for Huawei chips regardless, and they’d push out Nvidia to do it if necessary when and if that time comes.

Yes, Huawei production will expand over time, although likely not in the short run due to bottlenecks. But even if they do, so will Chinese demand, and it is not obvious they are on a path to catch up, even with an inferior product.

Is This About Being Superintelligence Pilled?

This certainly is a major factor in how you view such arguments, and rightfully so, but as I’ve said throughout, I don’t think you need to believe in AGI/ASI in order to think Jensen is wrong about export controls.

Sriram Krishnan: Every person here’s reaction to the Jensen + @dwarkesh_sp podcast can be extrapolated *directly* from whether they believe in the frontier labs achieving short timelines for AGI/ASI.

If you believe in the labs achieving RSI and then AGI/ASI (for some definition of all three) in the next few years, you’ll probably sympathetic to the frame @dwarkesh_sp adopts.

If not, you’re probably more sympathetic to the arguments from Jensen.

(if anyone here doesn’t fall into this pattern, would love to hear!)

I would put it this way:

1. If you believe AGI/ASI is plausible in the medium term (as in up to ~10 years), then the case Jensen makes against export controls is completely unconvincing.
   1. There are still arguments you can make against export controls, that might have some merit, but I would file those arguments under ‘galaxy brain takes.’
2. If you don’t believe AGI/ASI is plausible for more than 10 years, and perhaps indefinitely, then you should be more receptive to Jensen’s argument, but you should still reject Jensen’s arguments for the reasons I argued throughout.
   1. Dean Ball makes a related point here. In light of Mythos, and any reasonable expectation of what AI can do in the next few years, you don’t need to believe in ‘AGI’ you only need to believe in important strategic implications of AI, and we are already there today, which is enough to invalidate Jensen’s case.
3. If you not only don’t believe AGI/ASI is plausible, but you also think that it won’t matter much who has access to the bulk of the compute in the medium term, and it also doesn’t much matter whose models and applications people use, then and only then are Jensen’s arguments strong.
   1. As in, you think AI won’t much matter, so might as well make money on chips.
   2. But if so, you should probably also be short the market, especially Nvidia.
   3. Are you short the market?
   4. Also, we straight up don’t live in such a world. Between Claude Code and Codex, GPT-5.4, Opus 4.6+ and Mythos, we have ruled it out.
4. You could make a steelmanned version of Jensen’s argument, that has been made by the likes of David Sacks, which is that dominance of Nvidia hardware and CUDA within China also leads to dominance of American models and applications, because they form one coherent ‘tech stack.’
   1. I think that argument is false on the merits, for overdetermined reasons, even if you don’t believe in AGI/ASI, because it describes a world we don’t live in.
   2. I can imagine such a world existing, but it would look very different.

What should we think about the failure of Jensen to find better arguments?

Dean W. Ball: It’s a shame Jensen mostly fails here, because the monoculture on export controls is bad. If you’re a young AI policy researcher trying to make a name for yourself, it is almost impossible to be taken seriously unless you are pro export controls. Monocultures are usually bad.

Policy debates should not appear one sided, except when the sides are:

1. Make everyone else worse off so I can make more money.
2. No.

Position number one often wins such debates, because the special interest cares quite a lot about concentrated benefits, versus others caring less about diffuse costs. But yes, in cases where someone is seeking rent, or seeking to do something destructive, you will get a very one sided policy debate on the merits.

If the policy debate is one sided, I want to believe that the policy debate is one sided.

If the policy debate is not one sided, I want to believe that the policy debate is not one sided.

Thus, if good arguments against export controls exist, we want to hear them, even if ultimately we think export controls are good. Also, if they exist, I haven’t heard them.

The lack of being sufficiently pilled is also, again, why Jensen ‘lost’ Anthropic, and also a lot of how the current United States government tried to ‘lose’ Anthropic, at a time when the mistakes was a lot less understandable.

Dean W. Ball: In this regard the most interesting moment in Jensen/Dwarkesh is not the debate about chip export controls but instead where Jensen says he didn’t understand Anthropic’s scaling needs when approached about an investment in them a couple years ago. He admits he was un-pilled.

The Biden administration officials and EAs, who jensen casts as technologically clueless, would have understood Anthropic’s scaling needs much more intuitively a couple years ago than Jensen admits to in that interview. It’s not about savvy or intellect, it’s about pilledness.

Matt Beard raises an excellent point, and highlights Jensen saying “Although AI is the conversation today” when trying to downplay TPUs, so yeah, still highly unpilled.

Jensen’s Arguments Are Poor Both Logically And Rhetorically

There are (at least) two ways an argument can be poor.

1. An argument can be logically poor, and without underlying merit.
2. An argument can be rhetorically poor, and unconvincing to listeners.

The problem with Jensen’s arguments, and accelerationist AI arguments in general, is that they are usually poor in sense #1, and consistently poor in sense #2, at least when applied to general politics or the public.

Anton Leicht is warning accelerationists that they are slowly but surely losing ground. The strategy has been to argue against any and all asks and insist on nothing and playing pure hardball politics, without the rhetoric and support to back it up, and that failure to try and shape the eventual rules or get ahead of actual harms works until it spectacularly doesn’t.

Those who buy these arguments were always rather niche, and as AI capabilities advance that becomes more true every day, including today with Opus 4.7.

Dean W. Ball: Dwarkesh/Jensen reveals how inconsistent and un-battle-tested AI acceleration talking points are, especially when they are filtered through the prisms of corporate comms and mass politics. Strategically coherent accelerationism is possible (I try!), but not currently prevalent.

I really do say this as an accelerationist fundamentally. It has always been clear that the default ai acceleration stance developed most especially during sb 1047 was not going to stand the test of time (the default anti 1047 argument hinged on ai not improving very much and a funhouse conception of diffusion as “a totally intractable mystery problem” rather than “an obstacle”; this is basically still the default ai acceleration argument), and that a new, more complex path would be needed.

I’ve tried to do chart this path in my own mostly-between-the-lines way but I’m just going to be explicit for a moment that a new approach is obviously going to be needed _for those who are excited about AI, think it’s likelier than not to go well (especially with the big risks competently managed and self-aware strategic execution) and want to embrace it with alacrity_.

Dean W. Ball: However, one must acknowledge that, even though Jensen said it in the midst of discursive retreat, “that loser premise makes no sense to me” goes hard as a phrase

Nathan Calvin: I generally away from that interview completely unpersuaded by Jensen’s arguments, but convinced the man is a force of a nature and cool in a sort of brutal ur-techno-capitalist way

Dean W. Ball: in a sense the flex has always been the logical inconsistency

I think ‘loser premise makes no sense to me’ is an extremely telling phrase into Jensen’s psychology.

I think it is causal. As in, that premise would make me a loser, ergo it makes no sense.

Cause if there’s one thing to know about Jensen Huang? He’s a winner.