Dylan Patel on GPT-5’s Router Moment, GPUs vs TPUs, Monetization

The AI hardware race is heating up, and NVIDIA is still far ahead. What will it take to close the gap? In this episode, Dylan Patel (Founder & CEO, SemiAnalysis) joins Erin Price-Wright (General Partner, a16z), Guido Appenzeller (Partner, a16z), and host Erik Torenberg to break down the state of AI chips, data centers, and infrastructure strategy. We discuss: - Why simply copying NVIDIA won’t work, and what it takes to beat them - How custom silicon from Google, Amazon, and Meta could reshape the market - The economics of AI model launches and the shift toward cost efficiency - Infrastructure bottlenecks: power, cooling, and the global supply chain - The rise of AI silicon startups and the challenges they face - Export controls, China’s AI ambitions, and geopolitics in the chip race - Big tech’s next moves: advice for leaders like Jensen Huang, Sundar Pichai, Mark Zuckerberg, and Elon Musk Timecodes: 0:00 Introduction & AI Hardware Landscape 1:11 Reactions to GPT-5: Is It Disappointing? 4:19 The Business of AI Models: Cost, Monetization, and the Router 7:34 The Economics of AI: Cost vs. Performance 10:10 Usage-Based Pricing & Product Stickiness 12:30 Advice for Sam Altman: Monetizing OpenAI 14:18 NVIDIA’s Growth & The Future of AI Compute 21:27 Custom Silicon: Threats to NVIDIA 26:09 The Silicon Startup Boom 45:28 Data Center Power & Cooling: The Next Bottleneck 50:46 Intel’s Role in the AI Era 57:56 Advice for Tech Giants: NVIDIA, Google, Meta, Apple, Microsoft 1:08:17 AI Policy & Export Controls Resources: Find Dylan on X: https://x.com/dylan522p Find Erin on X: https://x.com/espricewright Find Guido on X: https://x.com/appenz Learn more about SemiAnalysis: https://semianalysis.com/dylan-patel/ Stay Updated: Let us know what you think: https://ratethispodcast.com/a16z Find a16z on Twitter: https://twitter.com/a16z Find a16z on LinkedIn: https://www.linkedin.com/company/a16z Subscribe on your favorite podcast app: https://a16z.simplecast.com/ Follow our host: https://x.com/eriktorenberg Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details, please see a16z.com/disclosures.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  Nvidia is going to have better networking than you. They're going to have better HBM. They're going to have better process node. They're going to come to market faster. They're going to be able to ramp faster. They're going to have better negotiations with whether it's TSMC or SK Hynix and the memory and silicon side or all the rack people or like copper cables, everything. They're going to have better cost efficiency. So you can't just like do the same thing as Nvidia. You have to really leap forward in some other way. You have to be like 5X better.

• 0:23Speaker 01

  Today, we're talking AI, hardware, chips and the infrastructure powering the next wave of models with three people at the center of it all. Dylan Patel, founder and CEO of Semi Analysis, one of the sharpest voices on chips, data centers and economics driving AI's explosive growth. Aaron Price, right? General Partner at A16Z, investing in the technologies and infrastructure shaping the future. Guido Appenzeller, Partner at A16Z with decades on the front lines of AI, cloud and networking. From GPT-5's launch to Nvidia's dominance, custom silicon and the global race for compute, we're covering what's happening behind the scenes. Let's get into it.

• 1:03Speaker 01

  Dylan, welcome to the podcast.

• 1:05Speaker 00

  Thank you for having me.

• 1:06Speaker 01

  We've been trying to get you for a while. You're a busy man, but it worked out. Guido, why don't you introduce why we were so excited to have Dylan on the podcast and what we're excited to discuss?

• 1:14Guido Appenzeller?

  I think Dylan, you've done exceptional job in covering what's happening in the AI hardware space, AI semi-space and now more and more data center space as well. And just looking at it, currently the most valuable company on the planet is an AI semi company, right? The, I think, biggest IPO so far in AI was an AI cloud company. This is currently where it's happening, right? In any gold rush in the early days is the picks and troubles that make money. And I think this is the stage that we're in. So super excited to have you here today.

• 1:41Speaker 00

  Awesome. Thank you. Happy to talk about my favorite topics.

• 1:44Speaker 01

  Amazing. Well, maybe let's start with GPT-5. We just had some of the research Christina and Isabella on here last week. You said it was disappointing. Why don't you share your reactions or what capabilities you were hoping to see or overall?

• 1:55Speaker 00

  I think it depends on what tier of user you are, right? If you're just using GPT-5 and before you were $20 or $200 a month subscriber, you no longer have access to 4.5, which In my opinion, is still a better pre-trained model for certain things. Or you no longer have access to O3, which would think for 30 seconds on average, maybe, right? Whereas GPT-5, even when you're using Thinking, only thinks for like 5 to 10 seconds on average, right? Which is an interesting sort of phenomenon, right? But basically, like, GPT-5 is not spending more compute per se. The model did get a little bit better on a vanilla basis, right? 4O to 5 is actually quite a bit better. But when you think about, you know, what is this curve of intelligence, right?

• 2:38Speaker 00

  It's like, the more compute you spend, the better the model gets. And that's whether it's a bigger model, which GPT-5 isn't, right? You can see it's not a bigger model. It's roughly the same size, you know, or you think more, right? But again, like, this is something that OpenAI's first Thinking models, you know, the first few generations of O1, O3, would think for a long time and waste a lot of tokens, if you will. And when you look at, for example, Anthropic's Thinking models, even when you put them in Thinking mode, they think a lot less, right? To get to the same results or better results, right, as OpenAI was. And so OpenAI, I think, like, optimized a lot of like, well, if I ask, like, I think the silliest one I had asked was like, I asked O3 once, is pork red meat or white meat? And it thought for like 48 seconds, it's like, what are you doing? Like, this should just like, tell me the answer.

• 3:22Speaker 00

  And so like, the nice thing is that GPT-5 will think a lot less, even if you select Thinking manually. But more importantly, they have the sort of auto-functionality, the router, which lets them decide whether or not, hey, do I route to the regular model? Do I route to maybe Mini if you're at a rate limits? Or do I route to Thinking, right? And how much do I think? But in general, the Thinking model will think less. So, so there's less compute going into a power user's average query than before.

• 3:50Guido Appenzeller?

  But isn't it even more interesting? OpenAI cannot control how much compute it wants to allocate to you, right? If we're in a high-load situation, maybe tune the router a little bit so it's less, right? Maybe, I have no idea what they're doing behind the curtain. But there's this meme out there at the moment that basically all they did, which is a meme, right? It's not true, but all they did is take all three, plus a couple of smaller models, put a router in front, and offer that at a lower blended price, essentially, right? I think there's a little bit of that, right? Costs suddenly matters, and they figured out a way how they can steer that.

• 4:20Speaker 00

  I think, yeah, I mean, and they talked about how they've been able to dramatically increase their infrastructure capacity because I myself was just regularly using O3 or 4.5, right? And now I'm forced to use Auto, which sometimes gives me the O3 equivalent thinking model, but sometimes gives me just the regular base model, which sucks. But like, I think for the free user, it's actually quite interesting, right? The free user was not getting thinking models pretty much ever, not using them, or in many cases, they just opened the website and asked their query. And now sometimes their query gets routed there. So sometimes they get a way better model. But now sometimes the OpenAI can gracefully degrade them if they need to, right? And I think the router points to the future of OpenAI from a business, right? Like, you can look at sort of the model companies, right? Anthropic is fully focused

• 5:03Speaker 00

  on B2B, right? API, code, et cetera, right? Or Claude code, whatever it is, right? OpenAI, yes, they have that business, Codax and API business, but really, their majority of the revenue is consumer, right? And it's consumer subscriptions. But they have no way to upsell, you know, to make money off of all the free users, right? In any other application, consumer app, the free user still pays via ads. But this is not compatible with AI, right? Like, it's a helpful assistant. You can't just make the users' result worse by injecting ads. Banner ads don't really work in AI either. So it's like, how do you now monetize them? And I think with the router, they're getting really close to figuring out how to monetize that user, right? With the new C of applications, if you saw her

• 5:47Speaker 00

  product that she launched at Shopify, I think Shopify was an agent for shopping, right? And now this like immediately clicks, like, oh, if the user asks a low value query, hey, why is the sky blue? Just route them to Mini, right? The model can answer perfectly fine, and that is a chunk of queries, right? But if they ask, what's the best DUI lawyer near me, right? All of a sudden, this is like, you know, you're in jail, you have one shot, you're like, screw it, let me ask ChatGPT what the best DUI lawyer is. And now all of a sudden, the model's not capable of it today, but soon enough, it'll be able to contact all the lawyers in the area and figure out what their results are, and maybe search their like, court filings and whatever, right? Book the best lawyer for you, or an airplane ticket.

• 6:26Guido Appenzeller?

  Maybe negotiate a cut as part of that.

• 6:28Speaker 00

  Yeah, of course, they're gonna take a cut, right? This is a much better way of monetizing the free user. It's like, you know, it's like Etsy. 10% of their traffic now comes from chat and OpenAI makes nothing off of that, but they really, really will soon. Right? And partially that's because Amazon blocks chat, but there's a way to make money from shopping decisions, whether it's booking flights or looking for items. And those you now say, free user, I don't care. I'm going to send you to my best model. I'm going to send you to agents. I'm going to spend ungodly amounts of compute on you because I can make money off of this. But if it's a query that's like, help me with my homework, I'll send you like a decent model, right? I don't need to spend money on you. And so this is how I think like, OpenAI can finally make money off of the free user. And I think that's the biggest like thing about the router, right?

• 7:12Guido Appenzeller?

  This is so interesting. I think this is the first time that we've seen that there's a launch of a new model where, to some degree, cost is the headline item, right? I mean, the software was always like, who has the smartest model, who has the highest MLU score? Now we have suddenly people who use models for coding for eight hours a day. And surprised that if you take a large context window and the best model creates thousands of dollars of cost a month. So cost matters. And so to some degree, so where you're on the parry to frontier between cost and performance is the new benchmark for model competitive, no longer cost alone. Is that what we're seeing here?

• 7:44Speaker 00

  I mean, I think definitely, right? Like OpenAI said they doubled their rate limits for big amounts of users. They've dramatically increased the number of tokens they're serving from this launch, which effectively says this is an economic release.

• 7:54Guido Appenzeller?

  Probably also means the tokens are now cheaper, right?

• 7:56Speaker 00

  Yeah, yeah, for sure, for sure. I think the funniest thing is this whole cost thing you mentioned is like, we've seen this in the code space, right? Cursor had to pull away the unlimited Claude code. Initially, they have this super expensive plan and it had like unlimited rates. And then they were only like a weekly rate limit. Now they have like hour based rate limits. And I saw the craziest like thread on Twitter where this guy said he changed his sleep schedule, right? Modeled after like how sailors in the bay, if you're sailing, you can't sleep, right? Like it's solo sailing. They'll take like power naps when they get to the right spots so that they can like still be safe.

• 8:30Erin Price-Wright?

  In the morning when it's not very windy.

• 8:32Speaker 00

  Well, but like they can't sleep uninterrupted, right? And so because Anthropic had to put rate limits that are like not just week based, but like a number of hours based. And like, he like basically sleeps multiple times a day, but small chunks, just so he can maximize the usage. And there's also a leaderboard on Reddit where people are like competing to see how many tokens they're using through their subscription. And there's like a dude spending like $30,000 a month.

• 8:56Guido Appenzeller?

  So I'm going to find some developer in India that I can pair programming with so I can get the day cycle, he can get the night cycle and we both can maximize together the quota for the account. Is that the future then?

• 9:05Speaker 00

  I mean, but it's clear like people are taking advantage of the negative gross margin, like sort of subscriptions that are offered. I think Anthropic probably makes a positive gross margin off of my subscription. I don't code enough, but there's plenty of people that are definitely losing money. And so as you said, it's an economic.

• 9:19Guido Appenzeller?

  It puts more and more to, I think, just usage based pricing, right? I think if you have an underlying commodity that you're reselling to some degree that has, that is that large a part of your cost of goods, right? You need to go to usage based pricing.

• 9:31Speaker 00

  How much do you think the like customer capture and stickiness for these code products is? I'm curious what you think on that, right? Once you use an ID, once you integrate one of the CLI products in, like, how sticky is it? Or is it people just switch like that?

• 9:43Guido Appenzeller?

  That is a billion dollar question. That's a very conservative estimate. Look, Anurup Parthi has this great slide where he basically says, if you're building an agentic system today, right? But fundamentally what it is is of this loop, right? Where half of the loop is the model thinking, right? And that's trying to do some, the other half is then the user verifying, what did the agent do? Is it the right thing, providing feedback and trying to steer it in the right direction? Because it can't run forever, eventually you need to steer it back. One half of that is the model provider, right? They're trying to build the best models. The other half is really about, I think, designing the best possible UI to enable the user to give feedback. And I think there's value in that. So I think there's a certain amount of stickiness in there, right? So what are all the different tools, like in terms of visual, like say, take code editing, right? How can I most easily visualize what the code changes are? How can it most easily visualize, you know, what they impact, which files?

• 10:26Guido Appenzeller?

  You know, how can I, for small changes, get very quick feedback versus for complex ones, you know, get complex feedbacks? There's some tools that actually draw diagrams for you of what they do, right? So I think this will be the battle. I think there's stickiness in that, right? How much exactly?

• 10:38Speaker 00

  So in that sense, like people should be doing subscriptions to get people locked in, right?

• 10:44Speaker 00

  Instead of moving to usage-based pricing.

• 10:46Erin Price-Wright?

  Well, I think it's the customers that don't want to do usage-based pricing because it's so hard to guarantee, it's so hard for it to get away from them. And you actually want guarantees and you're willing to commit to pretty high spend in order to not have usage-based pricing. I think it's the model companies that want usage-based pricing.

• 11:01Guido Appenzeller?

  I think with consumers, it's frankly very hard to not have usage-based pricing just because the variability is so massive, right? If it's us coding versus somebody who does this as their full-time job, right? You just have a factor of 20 or so difference in usage. That costs a lot of money, right? I think for enterprises, we could see like more flat fee pricing because you can average it out more.

• 11:21Erin Price-Wright?

  You have a developer that's using it all day. You kind of know in a general sense, like how many hours a day they're programming and what that sort of looks like.

• 11:28Guido Appenzeller?

  The vibe quotas are harder.

• 11:31Erik Torenberg?

  Before we leave OpenAI, I want to ask a broad question, which is if someone was sitting here and saying, hey, Dylan, I'll listen to anything you tell me to do, any advice you have, as long as it makes OpenAI more valuable, what would you tell them?

• 11:41Speaker 00

  I would say immediately launch a method for you to input your credit card into ChatGPT and agree that for anything it like agentially does for you, it'll take X cut and then launch that product because where it does shopping, right? Because like everyone knows that like Anthropic and OpenAI and all the other labs are buying RL environments of Amazon and of Shopify and of Etsy and of all the different ways to shop on the internet, of airline websites, right? Now, just like, hey, integrate my calendar. I want to fly to there on Thursday. Make sure I don't miss a meeting. Cool. Book, right? Do that integration like super well. Know my preferences on whether I like aisle or window, all this stuff, right? And just take a take rate. I think this will make them so much money the moment they launch it. And I

• 12:24Speaker 00

  think they're working on it already, but I'd like to hear how he thinks about it because he shifted his tone massively on like ads over the last six months, right? He used to be like, no way. And now he's like, maybe, you know, there's a way to do it without harming the user. And I think this is how you monetize the free user, right? So I think that's probably what I'd tell him slash ask him about, like a whole line of questions around this.

• 12:47Erik Torenberg?

  Well, he's coming on the podcast in a few weeks, so we'll ask him. I want to shift to Nvidia. Nvidia's having a monster year. They're up almost 70%. What are the possible paths from here? How do you see it playing out?

• 12:57Speaker 00

  Depends like how pilled you are on like the continued growth, but I think you guys have a good vantage point. We have a good vantage point of how fast revenue is growing for a lot of these companies, especially the code companies, but even many other applications. I think we can clearly see the demand side is accelerating, right? And then if you look at the training side, I think the race is on. Meta is upping hugely, Google's upping hugely. If you just look at, again, just OpenAI and Anthropic and the compute that they have and are getting this year from Google and Amazon for Anthropic and from Microsoft, CoreWeave, Oracle for OpenAI, 30% of the chips are going to them, just those two companies. But that's actually like, okay, well, like 70% of the stuff, like who's making off? Well, one third of it is like ads, right?

• 13:42Speaker 00

  Whether it be ByteDance or Meta or many of the other people who are doing ads. So then it's still like, okay, well, where are the rest of these one third of the chips coming from? Well, they're like mostly uneconomic providers who I don't think it's like an obvious bet that they're going to, you know, keep raising bigger and bigger rounds. So what happens there? I think with the, you know, we talked about like coding, right? Like earlier, actually the Qwen Coder 3 model is actually super cheap if you're running it on prem or if you're running it in the cloud with all these inference libraries. And so like there's stuff like that as well. So I think the question is like, how much does it keep growing? Because clearly I think the first third is definitely skyrocketing, right? Of OpenAI Anthropic lab spend, the second third of like ads is going to grow. It's not going to grow like crazy, but I think there's definitely an inflection point that could be hit with Gen AI ads.

• 14:27Speaker 00

  I know Meta's been experimenting with it a lot, but I could totally be convinced that there's going to be a huge inflection and take rate there, right? Where you start showing me personalized ads, like every person that's an ad is like, looks like me. And I'll be like, okay, yes, except like slightly better. So I like feel better, right? And I'm like, I want to buy it. Yeah.

• 14:43Guido Appenzeller?

  I have no idea how this is going to, going to scale, right? But if you ask the question, how much could it scale, right? Like, how much value are we creating here? Is, can we create enough value to actually keep growing for a long time? If you just take AI software development, right?

• 14:55Speaker 00

  Yeah.

• 14:56Guido Appenzeller?

  We know we can easily get about 15% more productivity out of it.

• 15:00Speaker 00

  I don't think that's right. I think it's way higher.

• 15:01Guido Appenzeller?

  No, no, but the straight, like I talked to a lot of enterprises, like a classical enterprise, straight up GitHub Copilot deployment that gives you about 15%. We can do much more than that.

• 15:09Speaker 00

  But bro, like, you know how bad GitHub Copilot is? Like, how did they, how, look at their revenue ARR chart.

• 15:15Guido Appenzeller?

  This was the lower bar.

• 15:16Speaker 00

  It's so funny. It's so funny. If you look at their revenue ARR chart, it's like Claude Code in three months has surpassed them. And Cursor, you know, easily surpassed them. And then like, even like companies like Replant are like, and Wind Surf slash Cognition are like, gonna pass them. Like, it's like, what's going on?

• 15:31Guido Appenzeller?

  So look, let's assume we can get this 100%.

• 15:33Speaker 00

  Yeah.

• 15:33Guido Appenzeller?

  So we can double the productivity of a developer, right? About 30 million developers worldwide, give or take.

• 15:38Speaker 00

  Yeah.

• 15:38Guido Appenzeller?

  Right? Let's say 100K value add per developer. It might be a little high worldwide. The US is low, but worldwide it's high. So it's $3 trillion.

• 15:45Speaker 00

  Yeah, yeah.

• 15:46Guido Appenzeller?

  Right? So we're probably building technology here, which adds $3 trillion of GDP value. In theory, we could put that into GPUs because that's the main cost factor here, right? Just from a coding model, like not ignoring every other use case. So at least in theory, the value generation is here to keep growing, right? Now, how that translates to the industry is much more complicated.

• 16:03Speaker 00

  I think we've already seen AI's value creation exceed, so sort of there's like the whole, like, like the famous, like, oh, 300 billion problem or 200 billion problem now. It's 600 billion problem. I'm sure. So Koi is gonna put out like the 1.8 trillion dollar problem, right? Soon enough. But like, like, there is some like reality in that, of course, but, you know, ignores that like infrastructure spend today is accounting for five years of revenue, not like one. And the revenue looks like this, not like flat line. But I think, um, I think the main thing is that AI is already generating more value than the spend. It's that the value capture is broken, right? Like, I legitimately believe OpenAI is not even capturing 10% of the value they've created in the world already.

• 16:45Speaker 00

  Just by usage of chat, right? And I think the same applies to, you know, Anthropic and Cursor and whoever else you're looking at. I think, I think the value capture is really broken. Even like internally, I think like what we've been able to do with like four devs with, in terms of like automation, like our, our spend on like Gemini API. Is absurdly low and yet we go through every single permit and regulatory filing around every single data center with AI. And it's like, and we, we take satellite photos of every data center and we like, we're able to label our data set and then recognize what generators people are using, what like cooling towers and the construction progress and substations. All this stuff is like automated and it's only possible because of Gen AI.

• 17:26Speaker 00

  But and we do it with like very few developers and then like the value capture that I'm able to generate by selling this data, by consulting with it is so high, but the company is making it is like, like they, they get nothing out of it. Right? Like, I think this like, there is a value capture challenge here uh that far out exceeds the sort of creation, right? And, and, and as you get models like GPT five or open source models, like continuing to drive it down. It's like the value capture is just harder and harder and harder for these companies. Cause they're making, you know, 50% gross margin on inference if they're, you know, or, or less in many cases.

• 18:02Guido Appenzeller?

  And so many words, you're saying we're getting commoditized uh and, and therefore you can't capture the value. And thus you should temper your expectations of how much, how much uh you can spend GPUs?

• 18:11Speaker 00

  Well, no, I think, I think you can, I think there's still ways to like inflect hugely on value capture, right? But I mentioned the ads are a huge value capture.

• 18:20Erin Price-Wright?

  But that needs to happen before, before we see massive.

• 18:23Speaker 00

  Um, no, I think, I think the other thing is like, there's a lot of capital that's not been spent, right? Like the hyperscalers still can grow capex uh 20, 30% next year, right? From what they're doing this year. In addition, companies like CoreWeave and Oracle uh because they're tapping capital markets can raise way more than 20 to 30% capex. And then you go down the list further and it's like, oh, the largest infrastructure funds in the world, like Brookfield and Blackstone. Well, actually they're, they're turning all of their eyes to investing even more into infrastructure, AI infra. Um, and then you're like the sovereign wealth funds of the world, like the G42s or, you know, um, the Norway one or GIC uh in um Singapore.

• 19:04Speaker 00

  Like these people have barely started touching AI. And so I think there's a whole lot more capex that can come without it being necessarily like economically motivated day one. I'm also saying economically motivated capex can only grow so much, but there's so much other like where it's not clear from, you know, if you have a spreadsheet and you're basing it on real business, that you should actually spend this much. But people will because they believe, I believe, I think you believe, like infra, you know, people believe that this will be, you'll get profit out of it, but there's no like 100% certain like, you know, way to argue it. Yeah.

• 19:43Guido Appenzeller?

  How threatened is Nvidia by custom silicon?

• 19:47Speaker 00

  I think that's the biggest thing, right? Is when we look at orders from Google and from Amazon, right? Especially their and Meta, their custom silicon is not Microsoft. Their custom silicon kind of sucks. But the other three, they're really upping their orders massively over the last year. You know, Amazon is making millions of trinium. Google's making millions of TPUs. TPUs clearly are like 100% utilized, right?

• 20:15Guido Appenzeller?

  Yeah.

• 20:15Speaker 00

  Trinium's not there, but I think Amazon will figure out how to do that. And Anthropic will.

• 20:22Speaker 00

  So I think that's the biggest threat to Nvidia is that people figure out how to use custom silicon more broadly. And this sort of becomes the sort of like, if AI is concentrated, then custom silicon will do better. And that's not even talking about like OpenAI's silicon team and stuff, right? Like if AI is really concentrated, then they'll do better, custom silicon. But if it gets dispersed broadly because there's all these open source models from China and there's all these open source software libraries from Nvidia in China and it makes the deployment costs like rock bottom, then potentially.

• 20:58Guido Appenzeller?

  Hear me out here. If Google's TPU is able to compete with Nvidia, in theory, it could do it on the open market. Nvidia is worth more than Google these days. Shouldn't Google start selling their chips to everyone? I mean, in theory, they should be able to achieve a higher market cap.

• 21:12Speaker 00

  I absolutely think so. I think Google's even discussing it internally. I think it would require a big reorg of culture and a big reorg of like how Google Cloud works and how the TPU team works and how the Jack software team and XLA software teams work. I totally think they could. It would just take them like shaking themselves pretty hard to be able to do it.

• 21:39Speaker 00

  Yeah, but I totally think Google should sell TPUs externally, not just renting, but like physically.

• 21:45Guido Appenzeller?

  It's kind of funny if a side hobby in theory has a higher company value potential as your main product.

• 21:52Erin Price-Wright?

  Than your entire business, especially as you think about the degradation of search as a core business.

• 21:55Speaker 00

  Yeah, I think, but I think like if you were to ask like Sergey, right? Like, hey, do you think selling chips and racks is more valuable or cloud or Gemini? He'd be like, no, no, no, no, like Gemini is going to be worth way, way, way more. It's just not yet today, right? And so I think like, like today, you say Nvidia is the most, again, it's like a whole concentration thing, right? The world is super concentrated in terms of customers, then Nvidia will not be the most valuable company in the world, right? But if it gets dispersed more and more, which arguably we're starting to see with a lot of these open source models getting better and better and better, and with the ease of deploying them getting better, then you would see, I think you could argue Nvidia

• 22:40Speaker 00

  will remain the most valuable company in the world for a long period of time.

• 22:45Guido Appenzeller?

  Historically, no pun intended, software has eaten the world in most markets, right? I mean, like if you look at the early networking days, Cisco was the most valuable company on the planet, right? For a while. It's no longer, right? They're the guys that build services on top like, like Google or Amazon or Meta eventually eclipsed.

• 23:02Speaker 00

  Well, which is why Nvidia is like making all these software libraries, right? Like that's, that's, and they're trying to commoditize inference, right? Like, you guys don't, I think, even have an inference API provider investment, do you?

• 23:14Guido Appenzeller?

  Well, we have, we have all kinds of model providers.

• 23:17Speaker 00

  Model providers, but I'm talking about a pure API provider investment, I think. Right? Is that correct? I think I talked to one of the team members, maybe, maybe Rajco or someone about like why you guys don't, didn't invest in like, you know, like a Together or like a Fireworks. And sort of the argument was like, well, we think inference, just serving models. Alone without making them will sort of be commoditized.

• 23:37Guido Appenzeller?

  Yeah.

• 23:38Speaker 00

  Right?

• 23:39Guido Appenzeller?

  We have some in the stable diffusion ecosystem, like with like.

• 23:41Speaker 00

  With FAL, yeah. Yeah, it's a little bit different dynamics there, I think.

• 23:46Guido Appenzeller?

  They they tend to be make much more component models than the the L M folks.

• 23:49Speaker 00

  Yeah, yeah. But but like you guys don't have one of these like, you know, base 10 or any of these like sort of like API investments because you think this is from someone on the infra team that you guys think it'll get commoditized because the software Nvidia is making because V L M and S D laying, which is like open source software coming out of Berkeley and now, you know, sort of has their own environments now. And supported by many like this being commoditized means that like API providers aren't necessarily worth a ton, right, is sort of your argument, maybe. I think that's that's relevant to this whole thing, which is, you know, why, right? Like, why would you do this?

• 24:25Guido Appenzeller?

  Shifting gears, what about the the silicon startups? What's your take on those? I mean, there's a ton of capital flowing into that, right? We've seen, I have not numbers, but probably billions being invested in chip startups.

• 24:38Speaker 00

  Yeah, for sure, for sure. I mean, like whether you're looking at like companies like, I think it's like pretty impressive that a few companies like Etched and Revo's and a number of other companies, you know, Maddox and others like have gotten the amount of funding they've had without even launching a chip, right? You know, in the past, like, yeah, silicon companies would make money or raise money, but they would at least launch a chip before they get a, you know, a big round, but like Etched and Revo's like have raised, you know, a lot of money without ever launching a chip publicly, which I think is, I mean, it speaks to, well, like, yes, silicon is super capital intensive if you're building a chip, especially an accelerator, which has so many moving pieces.

• 25:22Speaker 00

  And there's, there's like, there's like 10 different AI accelerator companies out there, right? Like that are newish in the last few years.

• 25:27Guido Appenzeller?

  I think there's a lot more.

• 25:29Speaker 00

  That are like, yeah, that's fair.

• 25:32Speaker 00

  And then, and then, and then there's the old guard, which continues to raise money, right? Like Grok and Serebras and and Samenova and Tens Tor and so on and so forth, right? Like or GraphCore getting bought out by SoftBank and SoftBank dumping money into this effort as well, right? There's, there's a lot of capital being invested to disperse, dispel sort of Nvidia's top dollar or top position.

• 25:55Speaker 00

  But it becomes challenging, right? It's like, how do you beat Nvidia, right? Like, the hyperscalers, I think, are like, kind of lucky in that they can, they can do mostly the same thing as Nvidia, right?

• 26:05Guido Appenzeller?

  They have captive customer, which is themselves.

• 26:07Speaker 00

  Right.

• 26:07Guido Appenzeller?

  It's a huge asset.

• 26:08Speaker 00

  And they can, they can just win on supply chain, right? Like, I'm using cheaper providers.

• 26:12Guido Appenzeller?

  It's a margin compression exercise, essentially.

• 26:13Speaker 00

  Yeah, yeah. And maybe, maybe for certain workloads, like meta for recommendation systems, they'll have a better, you know, they can specialize more. But for the most part, it's like, no, we're, we're targeting the same workloads. We can just simplify supply chain or, or in-house a lot of it and compress margin, and it'll be fine. But in the case of, you know, these, these other companies, it's like, well, they don't have a captive customer. So now you have to contend with, well, I'm using the same ecosystem. And either I can use some custom silicon provider who's going to take a margin anyways on top, and that's going to compress my, like, what I can sell for. Or I can try and in-house everything, but then it's like, this is really hard, right? Like, I'm going to do all the software design, I'm going to do all the silicon design,

• 26:55Speaker 00

  I'm going to build all this different IP, I'm going to manage the supply chain on chips, on racks, on everything, right? Ends up being a huge effort in terms of team size.

• 27:06Speaker 00

  All in the end, like, hey, I make a 75% gross margin as Nvidia. AMD sells their GPUs for 50% gross margin. And they have a hard time out engineering Nvidia, and they're great at engineering, right? Like, they're, they, but yet they still take more silicon area, more memory to achieve the same performance, and they have to sell for less so their margin gets compressed.

• 27:27Guido Appenzeller?

  That makes some sense. But, but, look, the, I think, historically, if you look at it, typically, new entrants in markets didn't win by marginally improving on something existing. That happens sometimes, but, but more likely, they, they jumped on some of, some kind of disruptive technology leap, right? Whereas like, we have a different approach, we have different technology. Is that possible here? I mean, to, to some degree, maybe it's over simplifying a little bit, but I think part of the reason why the transformer model won was because it runs so incredibly great on GPUs, right? Like, a recurrent neural network is similarly performant, it looks like, but it runs terribly on a, on a GPU. So, so did we sort of pick the model for an architecture? And now it's, it's hard to come up with an architecture that, you know, Really.

• 28:06Speaker 00

  Well, it's hardware, software co-design, right? Like, there's like, there's all this hype about neuromorphic computing, right? Like theoretically, it's amazing and super efficient. It's like, okay, great. Like, there's no ecosystem of hardware, there's no ecosystem of software. It would take like, you know, tens of thousands of people who are the best at AI today focusing on that to even prove out if it's worthwhile or not, right? On a hardware side, on a software side, on a model side. And so like, you look at like Grok, Cerebras, Samenova.

• 28:34Speaker 00

  They all like sort of over-indexed to the models that were leading at the time when they designed their chips. And so they made certain trade-offs, right? They put a lot more memory on chip. And Nvidia was like, well, we're not gonna do that.

• 28:44Guido Appenzeller?

  Faster at least, right?

• 28:45Speaker 00

  Well, more like, like if you compare the amount of memory of SRAM on Nvidia's chips, it's much, much lower.

• 28:51Guido Appenzeller?

  Yes, correct. They went SRAM instead of DRAM. But then they usually have less DRAM, so there's a trade-off there as well.

• 28:56Speaker 00

  Right, there's less DRAM, there's more SRAM. And because there's more SRAM on the chip, you have to have less compute on the chip. And so they ended up losing, right? Because the model sizes got too big and all this, right? And so you have this like super weird dynamic where they bet on something that was actually better, right? Like, I have no doubt that Cerebras would run certain types of models better than Nvidia or Grok. Or, hey, Dojo, right? Dojo runs certain, you know, in Tesla's Dojo, would run certain types of models way better than Nvidia's chips. Because they're optimized to that. But then it's like, oh, well, actually, even in vision tasks, you use vision transformers now. So it's like, okay, cool. Because model sizes grew and all these things. So it ends up being a, you know, catch 22 in that like you optimize for something.

• 29:40Speaker 00

  And so now like, today, you have this new age of AI accelerator companies are like, okay, we're gonna optimize for transformers. But the time they started designing, they're like, okay, transformers are dense models that are this big. What's the best, you know, the hidden dimension is 8K and your batch sizes are this big and your sequence lengths are this big, so let's just make a super large systolic array. So you can, you know, create the maximum efficiency and then turns out, oh, look at DeepSeek or, you know, go look at what the labs are doing. Actually, their, their shapes are much smaller. Actually, you need to do a bunch of small matrix multiplies, not massive, massive, massive, you know, singular matrix multiplies per layer. And then it ends up, you know, oh, well, that chip you're designing for that is actually not super effective for that. And so the software is evolving constantly because of what, because

• 30:21Speaker 00

  of what works best on Nvidia. And you see that with, you know, whether it be what DeepSeek's doing, or Alibaba's doing, or what the labs are doing internally. And you even see this like for Google, right? Like their open source Gemma models make different decisions because the shapes of a TPU are different than a GPU. And those, the GPU and the TPU are actually not that far apart, right? Like you would say, yes, they're very different, but like Blackwell and TPUs are very, very, they're converging on similar designs, actually. Whereas like to beat Nvidia, you can't just have the supply chain, you know, with, right? You don't have this captive customer. So now you need to do something, you know, that will give you 5X advantage, right? In hardware efficiency

• 31:03Speaker 00

  for a certain type of workload, and then pray the workload doesn't shift, right? Because Nvidia is also optimizing their architecture generation. They've added a lot of stuff to make their model, their chips way better for the existing models. But it's like they're taking, you know, large steps every year, every two years towards something. Whereas you have to like go way over there in left field and hope that models stay over there, right? Because you have to win by 5X because Nvidia is going to have supply chain efficiency over you. They're going to have time to market over you in terms of like a new process node or new memory, or, you know, whatever, whatever technology, right? Even AMD, right? They got to 2 nanometer before Nvidia. They had higher density HPM. They use 3D stacking. All these

• 31:48Speaker 00

  things on supply chain that should be better than Nvidia, and yet they still lose. They're still the software angle, right? Nvidia is fantastic. Yeah, and then their software as well, right? But it's like. Nvidia is going to have better networking than you. They're going to have better HBM. They're going to have better process node. They're going to come to market faster. They're going to be able to ramp faster. They're going to have better negotiations with, whether it's TSMC or SK Hynix and the memory and Silicon side or all the rack people or like copper cables, everything. They're going to have better cost efficiency. So you have to be like 5X better.

• 32:13Guido Appenzeller?

  But to be fair, if somebody had a viable competitor, which would even be marginally cost competitive, if my guess is many of the big consumers of GPUs would immediately shift some revenue there just to have a number two, right? Just to turn into two.

• 32:24Speaker 00

  That's AMD today, right? And Microsoft stopped. I mean, like they're.

• 32:28Guido Appenzeller?

  It's still pretty limited traction, though, right?

• 32:30Speaker 00

  Sure, but Meta continues to buy from them. And Microsoft did buy a bunch and then they stopped because it's like, well, yes, they're, you know, AMD is giving you all these advantages, but it ends up still not being better on a performance per watt basis. And they have a way bigger software team. They're somewhat competitive on like all these dynamics that I mentioned, right? So you can't just like do the same thing as Nvidia. You really, and do it better, right? Or try and execute better like AMD. Like you have to really leap forward in some other way. But that's the, the design cycle takes so long that models will shift, right? Because they're like, oh, what's the next generation? TP and GPU look like? Okay, let's optimize for that. And the research path is, you know, like, great. Like, yes, neuromorphic computing could be the most optimal thing for us to do, but no one's working on that

• 33:14Speaker 00

  because you have to advance in the tech tree. You've chosen, right? If you restart the tech tree, you're going to be like, well, this sucks. And so like, if it branches this way and you're over here, you're screwed. Because you have to be.

• 33:26Guido Appenzeller?

  There's a mode.

• 33:27Speaker 00

  Because the supply chain stuff means that 5X actually turns into a 2.5X. And then Nvidia can compress their margin a little bit if you're actually competitive. And then that 2.5X becomes like a 50% better. And then, yeah, so it's like, it ends up being way too difficult to, and the software stuff, right? Everything like takes your 5X and makes it like, oh, you're actually only 50% better.

• 33:46Erin Price-Wright?

  And defense supply chain for sure.

• 33:48Speaker 00

  Yeah, defense supply chain. And then like, they get, they get that, right? Like, so it's like, and Lutnik himself said, we had to do the, this for rare earth minerals. And it's like, interesting. China, there's like provinces in China that have like rules that say the H20 is not efficient enough to be deployed, which is like super bizarre because it's clearly the best AI chip China has. Huawei is still a little bit behind.

• 34:13Erin Price-Wright?

  Well, what's interesting is that, you know, efficiency is just not, is so much less of an issue in China than here because they just have the power infrastructure to be able to support. So even if they're running less powerful chips. You know, you would imagine that it doesn't really matter because China has just such an infinite supply, infinite supply of power that, you know, they'd sort of be okay with it. So it's interesting.

• 34:33Speaker 00

  Which is, it's a big challenge in America, right? Like, um, there have been, there have been companies that were like, they would, they like, you know, Jensen keeps saying he couldn't give away H20 in America for free, but I've literally like heard companies like say, like now, say like, yeah, no, I mean, I wouldn't because like, I only have this much power. Um, how am I going, you know, in data centers ready to go over the next year? If I bought an H20, I'd literally have less compute capacity and then I'd lose, right? Even if it was free, like, it doesn't make sense. Um, whereas China doesn't care, they can build these things. Um, they have the muscle. Um, I'm, I'm curious how this all shakes out. You know, China's posturing really hard. They even like put out something that was like, we're investigating to see if there's backdoors in the H20. It's like

• 35:16Speaker 00

  there's no backdoor in the H20, like chill. Um, you know, it's like, you know, GPU, GPUs are usually like firewall from the public internet anyways. Like you, you step through stuff before you get to the GPU clusters. Um, so like a backwall, backdoor wouldn't even matter. Um, I don't know. I think, I think it'll be, it'll be interesting to see because, um, China can definitely deploy way, way, way more power to AI the moment they decide to. Um, but there's these like, there's like competing interests, right? Like.

• 35:50Erin Price-Wright?

  Because they want Huawei to be better than Nvidia.

• 35:52Speaker 00

  Yeah. And then this is how Nvidia argued to the administration. They're like, if we don't do this. Actually, I think it's like a very like powerful argument that like, like for example, within Triton, which is a common ML library, anyway, like, like ByteDance has open sourced some stuff that plugs into this that is like super awesome. And there's like all these other libraries. It's not just models that China open sources. It's like software for Nvidia that Chinese companies open source.

• 36:18Speaker 00

  In a sense, like by Nvidia selling GPUs, Nvidia's argument again was like, they were able to stop Huawei from building up a software ecosystem and the Western ecosystem is better. But then on the flip side, it's like, again, if you believe the models deliver more economic value to society than the hardware, which I actually think they do, it's just there's a value capture problem today. Then you're giving China way more by giving them H20s and soon a version of Blackwell that's cut down like Trump said, right? Versus selling them the chips, right? The economic value derived from selling them the chips is not as large as being able to somehow sell them AI services.

• 36:56Erin Price-Wright?

  So is China gatekeeping power for AI?

• 37:01Speaker 00

  I don't think so. I think, again, there's a lot of like, what we see is that even with H20 being sold to China into China and future versions of the chip, H20E and other chips, we still see Chinese companies like Alibaba renting GPUs outside of China because the GPUs they can get outside of China are just so much better on a dollar spend per performance basis. Renting them or even like going through sort of like a Singaporean company that is effectively a Chinese company and building data centers and putting chips in them. So it's like, I don't think China's limiting the power per se. It's that it's, you can only, if you can spend, like Chinese companies are growing their capex

• 37:44Speaker 00

  way more than US companies on a percentage basis next year. The dollar, absolute dollar numbers, obviously the US companies are spending more still on AI. The percentage basis, Chinese companies are growing more next year. And you still have the problem of like, well, dollar spend to AI output in tokens or in whatever is going to be lower because these chips are worse. So power is not the gating factor. It's always capital, right? At least today, right? Now China can spend a lot more capital if they wanted to. They're subsidizing the semiconductor industry to the tune of like 150, 200 billion dollars a year through SOEs, through capex that's not generating revenue, et cetera. So it's not like they couldn't do this to the AI ecosystem, right?

• 38:27Speaker 00

  Given Meta's capex is like 60 billion, right? And Google's capex is like 80 billion, right? Like they could totally spend way more than that on a single effort. They just haven't decided to. And I just think for the US, our build outs are constrained by power, right? Google has a ton of TPUs sitting, waiting for data centers to be powered and ready, as does Meta with GPUs, right? We posted about how Meta is now building these effectively tents.

• 38:54Guido Appenzeller?

  Isn't this to some degree also coupled to their unwillingness to sell them to a broader ecosystem? I mean, if they want to be confined in their own data centers and they didn't ramp data center build out for their own hyperscalers quickly enough, then yes, that constrains them, right? If they were on the open market, what would we constrain?

• 39:13Speaker 00

  Yeah, for sure, because companies like CoreWeave, why is CoreWeave valuable? It's really because they build infrastructure really fast, right? And their software is nice, I think, but a lot of their customers are bare muddle, right? Just replace the GPUs whenever they're broken and network it properly.

• 39:29Guido Appenzeller?

  They grew more aggressively.

• 39:30Speaker 00

  And they'll go anywhere.

• 39:31Guido Appenzeller?

  Jensen likes them as well.

• 39:32Speaker 00

  Yeah, they'll go, yeah, that's very important as well, but they'll like go, because it unconcentrates the ecosystem, which is better for Nvidia.

• 39:41Guido Appenzeller?

  Have you worked it into it? I know exactly what was going through his mind.

• 39:44Speaker 00

  Yeah, so I think what's really important is that CoreWeave doesn't care, right? They're like, oh, crypto data center, I will convert it to AI data center, right? They bought a company for like $10 billion that's doing crypto mining, which is worth like $2 billion like a couple of years ago. And it's not because their Bitcoin mining business is growing, it's because they have powered data centers, right? Anywhere and everywhere, people are trying to build powered data centers. And companies like CoreWeave and Oracle are moving to that. Actually, today, Google just didn't bought 8% of a crypto mining company called Terro Wolf, right?

• 40:20Erin Price-Wright?

  Not because they're getting into crypto mining.

• 40:21Speaker 00

  No, because they need the data centers, right?

• 40:23Erin Price-Wright?

  They want the power.

• 40:24Speaker 00

  They need the power, right? And it's like, all the hyperscalers have said, screw off to my sustainability pledges, because they need power as fast as possible, right? They're doing things that are not, they take a little bit longer to move the ship, but like... Even if you didn't do it in your own self-built data centers, there's still a lot of challenges in the open market. There's a deficit, right?

• 40:50Speaker 00

  And that's constraining American chip build outs heavily.

• 40:56Speaker 00

  Yes, others could maybe do it a little bit faster, like CoreWeave or others, right? Oracle's got an open mind as well.

• 41:04Speaker 00

  But it's still constraining US build outs heavily. Even though the capital's been spent, right? The chips are, you know, 60 to 80% of the cost of the cluster, depending on what chips you're getting. So it's like, they've already bought the chips. They just can't put them anywhere because the data centers aren't ready. It applies to Google, applies to Microsoft, applies to Meta, applies to a lot of folks.

• 41:23Erin Price-Wright?

  I mean, it's really hard to build power infrastructure in the US, so.

• 41:26Speaker 00

  Power, grid interconnections, transmission, substations, all of this stuff. Like, like, electrical contractors, electricians in Texas, if you're willing to be a travel electrician, it's like, it's like oil pay, right? Like, it used to be that, like, if you're physically adept, you could go make, you know, 100 grand in West Texas. But like, who the fuck wants to do that? Now, it's like, well, you could, you could go like 200 miles away from Dallas, and what's still a reasonable town, and, and build a data center and work on the wiring within the data center and all this other stuff, the transmission stuff, and your pay is up like 2X now, versus what it was just a few years ago. This labor problem's a challenge too, and it's, it's, yeah. I think in China, they don't have any of these problems,

• 42:10Speaker 00

  but they just haven't spent the capital yet, but capital is an issue as well.

• 42:15Speaker 00

  Because of the scale of what's being spent, right? Like, like Nvidia's revenue this year is gonna be like over $200 billion, and next year, it's expect over $300 billion. Plus, Google's gonna spend like $50 billion on TPU data centers, right? And it's like, and Amazon's gonna spend tons and tons on Tranium data centers. It's like, the scale of dollars is quickly growing to nation-state level stuff. And, and what's more important is being able to decide to spend the dollars and what's cost-effective. And so, to some extent, China's still constrained by that. But they can smuggle chips in, they can build data centers outside of China, they can rent data centers outside of China,

• 42:54Speaker 00

  and be more, and have the most cost-effective, you know, Blackwell chips or whatever, right? ByteDance is, you know, either the biggest or the second biggest customer of Google Cloud for a reason, right? And they're getting, you know, over, you know, they're getting. Many, many blackwell from them, right? And the same with Oracle and the same with Microsoft and all these other companies are renting tons of chips to China anyways, because it's more cost effective to do that than build it yourself. So it's not like China has this mentality where we only have to, well, the government does, but the infrastructure companies don't, like Alibaba, Tencent, ByteDance, et cetera.

• 43:25Guido Appenzeller?

  So what's the endgame for data centers? I mean, like, we need more power, we need more cooling. Well, the empty, every, all data centers would be next to a nuclear reactor or lots of solar, you know, next to a deep level, like, deep seawater that we use for cooling or something like that, or what's?

• 43:43Speaker 00

  I think that, like, cooling is, like, the physical cooling of a data center are like, you know, there's this whole narrative about like, oh, AI uses so much power and it's like, not really, you know, farming alfalfa uses like 100x the water of AI data centers. Even by the end of the decade, it'll be the same. And it's like alfalfa is like worth very little. So it's like, I don't know, there's like, it's like, cooling is like, not that, you know, people have like experimented with like, you know, undersea data centers to reduce the cooling cost.

• 44:09Guido Appenzeller?

  That doesn't make sense.

• 44:10Speaker 00

  It's like 5, 10% savings, but then like, if you want to-

• 44:12Guido Appenzeller?

  It's easy to get the water out of the ocean then, then with the data center into the ocean.

• 44:15Speaker 00

  It's like, if you want to service it, like, you're screwed, right? So like, the same with power, it's like, we talk a lot about like, the power is not actually that expensive, it's just hard to build, right?

• 44:24Guido Appenzeller?

  And how to get to the right place.

• 44:25Speaker 00

  And how to get to the right space and convert it down to the voltages and all this stuff that chips need.

• 44:31Erin Price-Wright?

  So it's less the magnitude of power and more where it is and how it moves?

• 44:35Speaker 00

  Well, the magnitude too, right? Like, it's gonna be-

• 44:37Guido Appenzeller?

  In terms of total worldwide energy consumption, AI data centers is still-

• 44:42Speaker 00

  Yeah.

• 44:43Guido Appenzeller?

  It's a fraction of a percent.

• 44:44Speaker 00

  Yeah, yeah. Even by the end of the decade, you know, the US will be like 10% of our power will be AI data centers, which is still like-

• 44:52Guido Appenzeller?

  Of electricity.

• 44:53Speaker 00

  Of our electricity.

• 44:54Guido Appenzeller?

  And in terms of energy, that's even a smaller fraction, right?

• 44:56Speaker 00

  Oh, yeah, yeah, because you think about-

• 44:57Guido Appenzeller?

  But shifting to electric vehicles, you can probably make a bigger swing than, than, you know, with all the AI data centers we can build.

• 45:03Speaker 00

  But outside, like, it's like, in Europe, like, that number's not moving up that fast and like all these other countries, I think-

• 45:10Speaker 00

  We need to build a lot more power, but it's not like some, some crazy, crazy like amount. It's just like doing it properly is the hard thing. And, and again, like the cost of power, like you go look at like these deals people are signing, they're still signing like even though the price has skyrocketed uh from like a few cents a kilowatt hour for these massive, massive purchases to like 10. It's still, you know, when you think about the full TCL, the cluster, you know, the GPU costs and networking, all of this stuff far outstrips the power. And same with cooling.

• 45:41Guido Appenzeller?

  What percentage is power from like if you have a four year amortized GPU data center? What percentage will be power?

• 45:49Speaker 00

  80% of the cost of a GPU data center, if you're building Blackwell, is capital.

• 45:52Guido Appenzeller?

  Yeah.

• 45:52Speaker 00

  Right? It's the GPU purchases, it's the networking, it's the, it's the physical data center uh conversion, powered conversion equipment. All of this stuff is like 80% of the cost. And then 20% is going to be your land and your power and your cooling and your cooling towers and your backup power and your generators and all this stuff. It's like nothing, which is why it doesn't matter if you spend, you know, 10% or 50% more on that. Um, because at the end of the day, the expensive thing, right? Like this is why what Elon did would seem silly, right? They spent a lot more money on, you know, generators outside the data center and these mobile chillers to cool the water down for their liquid cooling instead of like the more cost effective option because it got the data center up three months faster.

• 46:36Speaker 00

  And so like that three months of additional training time is worth way, way, way more on a TCO basis, right? The performance you got out of the chips and the time to market and all this is way, way faster. And therefore it was the right decision, even though this part of the data center ballooned its cost, everything else is still there and you're still paying for the chips. And if they were sitting idle, it's not worth it.

• 46:57Erin Price-Wright?

  Just by like bypassing the grid, bypassing anything to do with interconnect, anything to do with public utilities.

• 47:03Speaker 00

  Exactly, exactly.

• 47:04Guido Appenzeller?

  What's your take on Intel? Where is Intel going?

• 47:07Speaker 00

  I think the world, well, the US needs Intel. National boundary. I think the world needs Intel because Samsung is doing worse than Intel on leading edge process development, in my opinion, based on even on various customers in the industry having done test chips at like Intel versus Samsung. They think, I think the industry generally agrees that Intel is further along, you know, the sort of the two nanometer class process technology than Samsung is. But both are way behind TSMC.

• 47:36Speaker 00

  And TSMC is a monopoly in some extent.

• 47:40Speaker 00

  The number one question always people ask is like, why is TSMC not making more money? Why are they only raising prices, you know, next year, you know, three to 10%, depending on what it is? It's like, TSMC is a monopoly, like they could raise a lot more, but they're, they're good Taiwanese people rather than like dirty American capitalists. If TSMC was owned or was managed by Americans, I think most of the ownership is actually American in terms of the stock, it's on the New York Stock Exchange and all this, like, you know, they would have raised prices a lot more.

• 48:08Speaker 00

  And so like, there is this like difficult, difficult

• 48:13Speaker 00

  thing to be done that like, hey, there's one island that controls all leading edge semiconductors and not just all leading edge, like the majority of trailing edge production as well. Something needs to be done. Intel is behind, but not like, not like absurdly so, right? Like if something were to happen to Taiwan, Intel would have the most advanced technology in the world, right? It's just, it's not economic. Can you keep Intel as one company if you want them to be competitive? I think the process of splitting it would take so much executive time and so much executive effort that you would have been bankrupt by then, right? And that's the big challenge, like I think Intel should be separate, right?

• 48:49Speaker 00

  But to properly split the company and for all the management time that's needed is like absurd. And instead, like what you need is like, you need Liputan, who's the CEO, who's CEO of Intel. You know, there's a lot of drama going around about him because he's, he's one of the greatest semiconductor investors ever, right? He's invested in so many different companies first. You know, he was on the board of like SMIC, which is China's TSMC effectively, which is like a big like drama or like the, some of the biggest tool companies in China is the first investor in them. Because, you know, it was a multipolar world there and he's making good investments. But like, you know, now like people are getting mad about that, but it's like, no, he recognizes he, the companies, like he understands the supply chain. He needs to not spend his time on splitting the company because then he never actually fixes the company, right?

• 49:34Speaker 00

  Intel's problem is that like, It takes them five to six years to go from design to shipping the product. In some cases, more. And when they tape out a chip, right, like, you know, you send the design to the fab, the fab brings back the chip. They go through 14 revisions in some cases, where it was like, the rest of the industry goes through like, one to three, right, revisions, if they're good. Of like, send the design in, get the chip back, test it, send the design in, right, for a public launch. And they'll launch a chip in three years.

• 50:04Guido Appenzeller?

  So, but if you look at Intel today, right, they still don't have a competitive entry on the AI side.

• 50:10Speaker 00

  And they won't.

• 50:11Guido Appenzeller?

  Right. Can you, so what, but what does it mean for their offering? I mean, they're still doing great on GPUs. They don't have a good AI chip product. Is it long-term sustainable positioning, right? I mean, as a standalone chip company?

• 50:24Speaker 00

  I mean, IBM still makes more money every launch off of mainframes. So it's not like, it's not like x86 is dead. It's like, you don't get the growth rates, but like, you could totally run this as a very profitable enterprise. And I think the same with PCs, right? There's, you know, there's some turmoil, there's some ARM entry, there's some AMD competition. Well, like, I think it's a very, it can be a very profitable business if it had like, one-third the people or half the people working on it. And so, like, Liputan to fix Intel needs to go into both the design company and lay off a shitload of people, but like, keep all the good people and make sure that they're designing fast and they're launching, from design conception to launches, two to three years, not five to six. And that's on the design side and make that profitable. And then on the fabs, you have to do the same thing.

• 51:09Speaker 00

  There's all these people, like, one of the heads of fab automation at Intel. I explicitly told Liputan, because, you know, we have a couple ex-Intel people who are actually good in the company, that worked on the fab side, and we're like, they were like, who's the worst people and friends? It's like, oh, this guy sucks. I explicitly told Liputan, he had never talked to the guy, because it was like, four layers down. The company has like absurd amounts of hierarchy. It's like, four layers down, he goes and talks to the guy and he's out, right? And it's like, like, he figures out like, who's bad, right? And who's good? And he has to go in and he's still like, hey, the vast majority of the team at Intel is the one who led the world in production. And process technology for 20 years.

• 51:47Guido Appenzeller?

  Yeah.

• 51:48Speaker 00

  Right. But there's a lot of like built up crap. So he has to go figure this out, right? He can't waste his time on like, oh, all this like structuring to split like. I I think it would be better if the company split. I just don't think he can spend the time to do that. And if the design side of the company is, you know, you're not really going to get into AI, you're not really gonna. You you have to make some money there, but the fabs I think could truly become a competitor. But they're gonna go bankrupt by the time anything, you know, can happen. So they have to figure out how to get capital. So he has to figure out how to get capital. He has to figure out how to clean up all the crap, uh, make the, um. You know, yields go up, right? Make the product ship way faster. Like, all of these things are.

• 52:28Guido Appenzeller?

  I think the goals are completely correct. I mean, I think I think the big challenge, just reflecting back on my time there, right? I think the big challenge is that right now, if you look at Intel, right, they have essentially software, so the the chip design making. And then there's, you know, the core manufacturing part, right? And they have three very different cultures. And it's very hard to get everything under one umbrella, right? And so I think that is the big challenge.

• 52:48Speaker 00

  I think you could, you should even run the company separately, right? Like, but like. You can't physically separate them entity wise because it's going to take so long to sever all these things. Because he doesn't have time, right? Like, Intel is literally going to go bankrupt if they don't have a big cash infusion. Or they lay off like half the company, right? Uh, which some could argue you need to lay off like 30% of company anyways, but. There's a lot of bad things that happen if that happens, right? And they need to spend a lot more on building the next generation fab, even if they fix the fab, and they don't have money for that, right? So there's like, there's like a lot more more important problems than like physically separating the company, even though. I think long term, yes, the fab has to be separate from the chips, design software, right? Like, or chip design, uh, part of the company, just like.

• 53:31Speaker 00

  That's gonna make each company much more accountable, uh, be able to service their customers better, et cetera. It's just. That's gonna take too long and they're gonna go bankrupt by then.

• 53:40Guido Appenzeller?

  Awesome.

• 53:41Speaker 00

  I I but I think, I think, I hope, I hope, I pray, someone does something, right? Like, you get a big capital infusion, I don't know. The big hyperscalers are like muscled into like, oh, okay, wait, if TSMC eventually grows their margin to 75% because of the monopoly. Plus, they intake all this stuff like co-package optics and power delivery and all this, like. All of a sudden, the cost is gonna spike. So we should actually just throw $5 billion at Intel each, right? Screw it. And and that could actually give Intel enough of a lifeline to potentially get to something and maybe be competitive. That's that's the hope.

• 54:12Erik Torenberg?

  Can we finish by finishing this game that we started when we gave Sam Altman advice? If Jensen was here, what advice would you have for him?

• 54:23Speaker 00

  If Jensen was here?

• 54:27Speaker 00

  You know, I think he has a massive, massive balance sheet, right? Jensen does. He's cash, free cash flows like, you know, ridiculous.

• 54:37Speaker 00

  The tax cut, the new Trump, you know, tax bill instituted something really incredible, which is that you can depreciate all of the GPU cluster cost in year one, which we put out like a note about how like the tax implications to like Meta are like $10 billion a year. And across each of the major hyperscalers, it's like massive. It's like, well, Nvidia is going to spend tons and tons of cash.

• 55:00Speaker 00

  They're going to spend like, you know, tens of billions of dollars of taxes. Why don't you get into the infrastructure game somehow? Now, this is obviously going to be like crazy because like now they're buying their own GPUs and putting them in data centers and doing stuff. And they're competing with their own customers, but they're already doing that anyways because their customers are trying to make chips.

• 55:19Speaker 00

  But they should like accelerate the data center ecosystem with investments, right? Because really, we think we can like have very high degree of like accuracy on what they're going to do next year in terms of revenue because it's just the number of data center watts that are being built, right? Like this is harder thing to shift up and down, right? Now, there's a little bit of share difference between how much is TPU versus GPU. But like, it's like you have to accelerate the infrastructure and you need to spend all of this capital that you're building, right? Like, okay, do you want to go the route of like doing buybacks and dividends? Like, great, like you're a loser if you do that, right? Like you can make more money by reinvesting and building a bigger company that's not just chips into the ecosystem or servers into the ecosystem,

• 56:01Speaker 00

  but actually like controlling the infrastructure end to end somehow. So I think there's something he could do there with this massive war chest. And there's a reason like Nvidia has done some buybacks and they've done some dividends and increasing. But the cash on their balance sheet keeps growing and they're going to have north of $100 billion of cash on their balance sheet by the end of this year, I think. So it's like, what are you going to do with that? I think there's something moving into the infrastructure layer much more that they could do if he really wants to be the king of the world, right? Which I think he does.

• 56:35Erik Torenberg?

  Sergey and Sundar.

• 56:40Speaker 00

  I think they should open up the kimono on TPUs, right? Like, start selling them, open up the software, open source a lot more of the XLA software, because there's open XLA and there's XLA, but the vast majority is closed source. Really, really open up the kimono on that and be a lot more aggressive, right? They're still pretty not aggressive on data centers. They're pretty not aggressive on a lot of elements of the company. The TPU team's next-gen designs are pretty not aggressive, partially because a lot of the TPU team has left to go to OpenAI, the best people that I knew. It was actually really annoying. I knew like four people or five people and they all went to OpenAI. And it's like, fuck, like, now I don't get as much. I met some other people, right?

• 57:22Speaker 00

  But it's like, you know, I think they could be a lot more aggressive in many ways across the company. They don't have to be, right? But they could. Because AI, you know, like this chat GPT, take gray, the shift of search queries, the monetizable ones especially to purchasing agents is gonna really screw Google long-term if they don't, you know, get their act together. I think they've gotten their act together on DeepMind. There's still some inefficiencies, but Sergey works within DeepMind a lot and they're driving hard. They're still a little bit behind, but like, I think like, physical infrastructure, TPU, and how much money they could make and how much they could take the wind out of everyone else's sails if they start selling TPUs externally and reorg around like, building

• 58:04Speaker 00

  data centers much faster so that they do have the most compute in the world, because they did. But now there's certain companies that are gonna surpass them, potentially, over the next few years if they don't really get their act together. So I think that's what I would say for them.

• 58:18Speaker 00

  Yeah, and also like, like, learn how to ship product.

• 58:23Erik Torenberg?

  Zuck.

• 58:25Speaker 00

  I think Zuck, you know, it remains to be seen what goes on with superintelligence, but like, they're trying to move super fast with the data centers, you know, like, screw it, we'll build tents instead of like, physical data centers, because we only need these for five years anyways. You know, the superintelligence moves, you could say whatever you want, but like, you know, trying to buy like, Thinkie for like, 30 billion or SSI for 30 billion didn't work out, so then they spent, you know, not even that much on hiring, not 30 billion on hiring all these people. So I think that he recognizes the urgency with the models, with the infrastructure. So I really think he needs to like... You know, if you read his website post about like AI, like I think, you know, he sees the vision,

• 59:07Speaker 00

  right? There's the wearables, there's integrating AI into that, there's being your AI assistant to do all this purchasing and stuff. I think he sees the vision, but I think he also needs to focus on like actually like releasing that faster, but also like the products that they do outside of their core IP, every time they launch something is kind of mid, right? You know, Meta Reality Labs is doing well, but I think they should like go more explicit, like have a chat GPT competitor, have a Claude, like Claude code competitor, like just start releasing way more products. Because they're really just focused on their individual gardens rather than like branching outside of it.

• 59:46Erik Torenberg?

  Do you think Apple should have that same sense of urgency? Or if Tim Cook was here, what would you tell him?

• 59:51Speaker 00

  The funny thing is like some of their best AI people are now like at superintelligence. They're building an AI accelerator, they're gonna, they're, they're, they have AI models, but they're just like way slower. They did mention on the last earnings call, they're gonna allocate more capital to this, but it's like, guys, Apple, like you guys are gonna lose the boat if you do not spend like fifty, hundred billion dollars on infrastructure. You don't think the cons theory will, will cut it?

• 1:00:14Speaker 00

  I think, I think like more and more you'll see people like, you know, great, Apple has this world garden, but like they can only do so much to protect it, right? IDFA, like they shut down ads to, or data sharing to Meta, but Meta made better models and now they have way more data and way more power over the user than they ever did before. Kind of, it was good that Meta kicked the crutch off of them, or Apple did. But the same applies to like AI, like, yes, they have access to the text and they have access to this, but like, I think other people are gonna be able to integrate user data and agents will be able to integrate all this user data and they'll start to lose control of what the user experience is as more and more gets disintegrated by AI being

• 1:00:55Speaker 00

  the interface rather than touch, rather than, you know, touchpad and keyboard. And I don't think they've truly realized what happens when the interface to computing is, is AI. Like they, they market it, but like that's gonna shift computing really heavily. They have great hardware and their hardware teams are working on awesome stuff, inform factors, but like. I just don't know if they like get what is actually going to happen to the world in the next five years truly well enough and they're not building fast enough for it.

• 1:01:24Erik Torenberg?

  What about Microsoft to that end?

• 1:01:27Speaker 00

  Microsoft has the same problem. I think they were super aggressive in 23 and 24. And then they pulled back heavily, right? Now like OpenAI is slipping through their grasps. There's that whole thing there. They cut back on data center investments heavily. They were going to be the largest infrastructure company in the world by like a factor of 2X, which would have been, you know, you could argue maybe that was like too much and maybe it wouldn't have been economical. But like they're losing grasp on OpenAI, their internal model efforts are failing spectacularly. Like they're on LLM arena right now and they're pretty decent there, but it's like, that's just like a sick authentic model. Like it's a code name, but like whatever, like

• 1:02:06Speaker 00

  MAI is like failing. Azure is like losing a lot of share to Oracle and CoreWeave and Google. And so on and so forth, right? Their internal chip effort is by far the worst of any hyperscaler. Like they're just like mis-executing. Like GitHub, how is GitHub not the highest ARR software code model?

• 1:02:28Guido Appenzeller?

  They only had the best IDE, the best source code repository, the best enterprise sales force, the best model company as a relationship, and they were the first to market, right? It's like they're really going for them.

• 1:02:39Speaker 00

  And like there's just nothing, right? It's like GitHub Copilot is failing. Microsoft Copilot is like still crap, right? Like

• 1:02:47Guido Appenzeller?

  Yeah, it's unusable.

• 1:02:48Speaker 00

  And it's like, what is going on? Like you need to shake the crap out of the company. Like I think they win a lot because they have the best business to business relationship with so many enterprises.

• 1:02:58Guido Appenzeller?

  Best sales force on the planet.

• 1:03:00Speaker 00

  Yeah, but like they end up like not having the actual product to sell them, which is like really scary. So they need to really work on product. Satya has done great on sales and stuff, but like, yeah.

• 1:03:11Erik Torenberg?

  If Elon was here, what advice would you give him?

• 1:03:14Speaker 00

  A lot of people at XAI are mad about the porn models, like porn stuff. It's fine, like you're gonna make a ton of money off of this. This is how you accelerate the revenue of that company. But like, he's losing a lot of talent and axing a lot of good projects. But Elon is a magnet to amazing talent and building stuff. So I won't bet against him, but it seems like since he left the administration and focused on stuff again. But I think, I don't know, I think he's focused on a lot of things. And I think like RoboTaxi, starting to look good actually again. Like, I haven't ridden one yet, but I have some friends who ridden one. It's like, looks pretty decent. He could like, not make these snap decisions, which often are the reason why he's amazing. But like, some of these snap decisions are hurting him.

• 1:03:52Erik Torenberg?

  Yeah.

• 1:03:52Speaker 00

  I'm not sure if I can give Elon that much great advice, because I think maybe it's just like, focus on like, the products again, right? More. But he is working on that stuff a lot, yeah.

• 1:04:03Erik Torenberg?

  I think that might be a good place to wrap.

• 1:04:05Speaker 00

  Awesome.

• 1:04:05Erik Torenberg?

  This is a great discussion. Dylan, thanks so much for joining us.

• 1:04:07Speaker 00

  Thank you for having me.

• 1:04:10Speaker 01

  Thanks for listening to the A16Z Podcast. If you enjoyed the episode, let us know by leaving a review at ratethispodcast.com/A16Z. We've got more great conversations coming your way. See you next time.

• 1:04:24Speaker 01

  As a reminder, the content here is for informational purposes only, should not be taken as legal, business, tax, or investment advice, or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any A16Z fund. Please note that A16Z and its affiliates may also maintain investments in the companies discussed in this podcast. For more details, including a link to our investments, please see A16Z.com/disclosures.