RSI Is Closer Than People Think, Per Tae Kim

0:00We have Tae Kim in the waiting room. Let's bring him in to the TVP on UltraDome. Tae, how you doing? Hey guys, doing great.

0:08Tae Kim:What's going on? So tell me, last time you were on the show, you bottom-ticked it? What's going on? I think I made the bullish call on CPUs, memory, and NVIDIA. NVIDIA is up like 5-10%, but the CPU names have still doubled, even after this big drawdown. and the hbm names are up 100 so i i'm hoping that you know it's the same thing again i come on here it's not go up again yeah ideally we could have like an emergency reserve of take appearances so if the market is ever down a strategic reserve we call you up you jump on that's good and then it was funny because it was literally the exact bottle and you know it went exponential after that okay so where are we right now with the level of fud the level of uh downward pressure on the ai trade broadly the chips the semi-trade like reset for us on like where sentiment is and then we can work through the different pieces of counter examples so i think sentiment's very negative and we kind of had this huge up parabolic up move the last few months and likely a lot of retail and hedge funds piled in and we were seeing this unwind now i think the first a big part of it was the iran war getting worse uh every time we had the the first ceasefire negotiations stocks started taking off right right after that and then when we had the actual ceasefire we had a follow-through and then as soon as trump started bombing iran again you know chip stocks have kind of plummeted in the last two three weeks uh and then now we we're seeing just you Back to the old pattern of media and the viral hot takes spreading a lot of FUD.

1:53I think we saw earlier this month, I think Reuters quoted Zuckerberg about agentic AI. They took it out of context. And then every media person was with the hot take that this meta was seeing bad returns and they're going to cut a CapEx. And then we had leaks right after that saying that it looks like meta is going to raise CapEx. So we're seeing a lot of this hot take FUD. Yesterday, I think we had a flurry of stuff that scared people. The Wall Street Journal vendor financing article that we'll see what happens with that. We had CMXT IPO in China, and everyone freaked out over that. We had the information article on ASML.

2:36We could go through each one. And then the Kimmy thing, it's obviously a big thing. Yeah, we'll definitely get there. And I want to talk about open source and NVIDIA's strategy there, obviously. Starting with the Mark Zuckerberg news in Reuters, this was July 2nd. Meta's Zuckerberg says AI agent tech progressing slower than expected. Zuckerberg added that the company's reorganization that included major job cuts was not as clean as it could have been. Zuckerberg and other meta executives have been seeking to moderate some of the organizational changes introduced this year. And they said that the trajectory of agentic development over the last four months hasn't really accelerated in the way we expected.

3:17The company's bets on new structure haven't come to fruition yet. And so people were sort of reading this as maybe meta is going to pull back. But then it felt like the response was extremely quick with Boz going on a podcast and Alex Wang sharing a whole bunch of progress across a few different models and data points. and then a semi-analysis wrote a whole bull case for MSL talking about how they have compute and also they have more of like the internal structural alignment to sort of properly YOLO in the AI era if I'm boiling it down as brutally as possible. Just because with Google, there's always this debate between, oh, do you sell the TPUs or do you sell the cloud, do you have vended in the product, whereas Mark Zuckerberg is able to sort of like go all in on this new idea.

4:09And so maybe there's more glimmers of hope there. But what else have you been tracking downstream of Meta's ambitions? Well, I mean, they've been very upfront that they're investing heavily in AI. Alexander Wang is tweeting multiple times every few weeks that they're going full force. They're going to redo open source AI models. I think he said that at the YC event over the weekend. And it's, I mean, if you actually look at, and then Reuters came out, I think, with an article saying that they're actually going to raise capex dramatically uh this year and next year so all that kind of fear that that quote about a gigantic ai from the town hall um that kind of like spooked the market for a few days it kind of it was completely false um the stuff like this yeah it feels like it's a comms error because the the language that's been coming out of meta has been a little bit like ai is going to replace our employees and And it feels like it'd be much better for them to come to the market with a message of we're going on the offensive.

5:13Like we're a hyperscaler. To be fair, that was the internal town hall. They didn't mean to leak it. And Reuters leaked that one quote and put out the headline before the article. Yeah. It's interesting.

5:24Tae Kim:Did Meta basically go through an eight-year period where internal town halls didn't instantly leak? I think everything leaked always. I think everything's been... I know, but there was a period where the attention of the media was way, way, way less on what Meta was doing internally relative to the 2010s. And all that attention just went to the labs, right? Yeah, yeah. No, that makes sense. Yeah, I guess the question is like, the question that I keep coming back to is like, where is their revenue ramp? Where is their AI revenue going to ramp? And when? right because as you say ads like the ads like the ai has yeah and that's they're accelerating always that's always that's always been my view too but when you're when you're continuing to ramp capex yeah with and saying like we're going all in on agentic and we're building a harness and we're also going to do open source and it's like well what is the strategy sure like yeah What is going to take you to a billion dollars of like pure AI product revenue or just API revenue?

6:36Tae Kim:And then to five and ten. And what's going to allow you to like justify the spend other than I think the market would love it. They just said, yeah, we actually need all these GPUs because we can actually be 10 times. We're already good at ads. We can be 10 times better. And that's where we're going to get the ROI on all of this CapEx. Well, they're definitely getting ROI on that. The market is worried about all those extra capex on the... They're going for the Frontier AI model race again. And they had to reset. A lot of people left and now Wang hired a ton of people and we'll see what happens over the next.

7:11It's going to take time. It's going to take six to 12 months before we see any more progress. It's just hard when you don't... The model that came out a few weeks ago was a lot better than people expected. It wasn't the frontier, but it was much better than what people expected. Yeah. So how have you been processing the NVIDIA letter around open source and all the back and forth, all the people jumping on, the companies that have been staying back? How do you work through that? It's been very impressive. They basically united the entire tech industry against Anthropic in the last three, four days.

7:4918 trillion in market cap has signed on last time i checked across i mean uh google for a little time amazon signed on eventually yeah okay they signed on yesterday they tweeted out interesting i think apple is still the holdout which is kind of strange because they're the one that would most benefit from open source open weight models being you know more available i would think but i would I don't know what Apple... But, I mean, they pretty much got the whole tech industry to kind of corner Anthropic in their position. Yeah. OpenAI signed on. Yeah. What did you think of Jory? Obviously, NVIDIA is afraid.

8:28Tae Kim:Yeah, I don't know how... I don't know if they're really... I don't read it as being, like, cornered by any means. Right? Well, Jensen is on the record that he said, I think to Bloomberg that there was a rising sentiment that something was going to happen on the regulation front. Oh. White House or whatever. So this was kind of a deep dive. Yeah. This was last week. You had at least four people in the admin say, we're not against open weights. We're against distillation. And at least I was reading into that of some type of regulatory action around open weights and then positioning it as we're targeting.

9:11Tae Kim:This is like we're hitting. I already wrote yesterday. Yeah. About pushback and restrictions and he's doing it under the safety umbrella. But definitely Microsoft and Biddy are worried that the White House or Congress is going to do something on this front. Yeah. That's why they took it. It seems very reasonable that he would have no problem with like Gemma or Lama or any of the open source from like American hyperscalers where if you find out that they're distilling, you just walk across the street and sue them. and also these big companies have huge, huge, I mean, they have safety teams, but also just like huge incentives to not have a safety incident happen on their watch because you're trying to like catch up to the frontier and then all of a sudden you have a safety incident.

9:56that's going to be really bad for your overall brand. And you have a different business to protect, whether it's social networking or Google search, if all of a sudden the Gemma model winds up being a thorn in someone's side for a cybersecurity reason or a bio reason, that would be really, really bad. But a foreign company that is just like hurling it over here can kind of just be like, you guys deal with the consequences potentially. So I think that's what Dario is worried about. What about the overall idea of where it feels like we're sort of replaying the deep seek moment, open source is going to reduce cost, and so that's a reason to pull back on the AI trade overall.

10:39How have you processed that? It's almost a perfect catalog. People are worried about Kimmy, but when you actually read the technical paper and their blog posts, this is not a tiny efficient model. This is 2.8 trillion parameters. it's going to require a ton of compute to serve. I mean, we saw it the first day they put it out, but their servers got slammed. Even in the blog post, they say it's best run on a kind of a server with 64 GPUs. So big, super clusters that are networked well, and that perfectly runs great on NVIDIA. And if you remember, you know, during the whole DeepSeek thing about a year or so ago, the market freaked out that, you know, DeepSeek was so efficient that will lead to a compute glut.

11:22But DeepSeq was an example of the reasoning model that actually, it was the opposite. It created a ton of demand. And I think the same thing has been happening with Kimi, where when you have more capable models that come out, people find uses for them. And right now, just like last year when reasoning models took off, Agentic AI and Agents are taking off right now. And the market is kind of like not realizing that because right now, just like last year when reasoning models were taking off. Right now, agendic AI is taking off. And the next six, nine months are going to be bigger than anyone believes.

11:57And Sam is on the record. Sam is on the record over the weekend. At the YC thing, again, I don't know why people don't listen to me. It's on YouTube. The next six months, it's going to be much more dramatically better for AI than the last two years in terms of advances in capabilities. And I heard you say RSI before. I think it's going to be RSI. People inside OpenAI and definitely Anthropic, Anthropic put a blog post on this. RSI, I think, is a lot closer than people think. And if RSI actually happens in the next three, six, nine months, that's going to soak up insane amount of meat. I mean, we have this exponential ramp for reasoning, exponential ramp for EGENTIK.

12:40And if RSI actually happens, and I think it sounds like both Frontier Labs think it's going to happen very soon, that's going to soak up an unbelievable amount of compute as the AI models use more compute to self-develop and improve. And I think that's one thing that people are missing, that both Anthropic and OpenAI are kind of winking that, oh, it's happening. Anytime I tweet something on RSI, all these Frontier AI researchers like my tweet. So I think that's good.

13:08Tae Kim:What is your sort of framework around compute hoarding? Because certainly it has been happening. When you look at, you know, like going back to the meta example, right? They're not selling compute yet. They're maybe curious about it or exploring some deals. They have all this compute and they're betting on their own ability to create the capability that will have enough demand to justify that. Do you just think there's so much demand overall that even if there's hoarding, it just will leak out and it's okay? Well, there's so much demand overall. I mean, the SK Hynix executives said during their IPO run that their customers are asking five to six times more than they're able to serve.

14:00And they're going to double capacity over the next five years, they said. And their customers, and I'm going to assume it sounded like Jensen, are asking for five to six times more than they're able to build. So there's overwhelming demand. You guys were at the Advanced AI A &D event. Lisa Su raised her agentic CPU forecast. Just three months ago, it was$120 billion for 2030. Three months later, they raised it to$220 billion. She doesn't do that. You have that just on the record. I've got that ready. I can do whatever. I mean, like, CEOs.

14:38Tae Kim:Well, I just love this chart because he called it perfectly. He actually did. It's crazy. CEOs don't, you know, raise their tans by, like, these multiples in a few months. If they're not seeing insane demand coming in. Especially not public CEOs who are serious business leaders who've been running, like, non-meme stocks for decades. So everyone's freaking out that this is the dot-com bubble all over again. But what if these hyperscaler GPU cloud businesses are amazing businesses? Like Morgan Stanley says, if you do inference, it's 60 to 80 % profit margins. These are amazingly profitable businesses as long as we keep growing the next few years.

15:21And again, just like last year, we're on this exponential run right now over the next two quarters. and the market isn't seeing that. Everyone's freaking out that, oh no, we're spending too much. And even Sam Altman podcast came out today and another podcast. Sam is out there. He said that he regretted pulling back on the compute purchases. They made a mistake by not putting the pedal to the metal because now things are taking off again. So like Amazon, the CEO in April, If everyone read his annual letter Andy Jesse wrote, he talks about how free cash flow works. We're not betting$200 billion on a hunch.

16:04We see the demand. We know it's going to be insanely profitable and free cash flow positive in the medium to long term. So that's why you're investing$200 billion now. And in a year or two, we're going to see insane amounts of free cash flow, the thing that people are worried about right now. It takes time to build out these data centers and fabs. And you bet now. Hold on, hold on, hold on. In a couple years. Like, if you see free cash flow, that assumes that, like, the revenues have to catch up and then the CapEx can't grow more exponentially. And so that means you have to see some sort of plateauing.

16:42Maybe it's at the end of the chart. Maybe it's this 20, 30 range. But there is a different world of just, like, continued growth forever. And then we sort of run out of money. The pushback I have there, that's a static view, right? If they don't grow revenue for the next three years, yes, you can't do that. But Azure is growing 40%. Google Cloud is going 80%. Amazon is growing double digits. So if revenue is growing 40 % to 80 % this year or next year and the year after, that's more revenue you have. That's more operating cash flow you have to invest. right? So that's what people are missing. And if this stuff, if the data center you're building now, you're spending all this now, generates unbelievable free cash flow in 12 to 18 months because this agentic AI is actually aging and re-architecting all the workflows inside companies and you need to do the agentic AI coding agents to make your product better.

17:46Because if you don't If you don't iterate a hundred different iterations of your product in R &D, if you don't do AI, just like AT &T is doing at the Genetic Advancing AI at AMD, he's talking about they're putting a hundred Gen AI models into production. They're burning a trillion tokens a month, and then that's growing double digit. The reason why they're doing that is because by using authentic AI, you're providing a better customer service, you have a better product R &D, and you're helping your companies make better products and services. And if you don't incorporate AI into your company, Verizon, your other company is going to incorporate AI and then disrupt you, and then you lose all your revenue.

18:28So everyone's wearing the ROI. ROI is important, but you also need return on revenue because if you don't use AI, your rival is going to use AI to beat you in the market. Yeah, yeah. No, I think the diffusion story is still, even though we got like sort of jitters by the token maxing thing, just the actual usage of AI across companies is still pretty limited in terms of the amount of people that are using it, the time that those people are using it. Like there definitely is a San Francisco bubble of startups where everyone is using AI a lot. But if you just walk into a normal business, a lot of people are like, yeah, I got to check that out, which is a crazy place to be.

19:08Let me give you some context here.

19:10Tae Kim:Yeah, Ara Karazian, Jared Sleeper over on X was saying enterprise adoption disparity remains enormous. and he cited Aura saying, usage would 100x if every company adopted AI to the degree of the most advanced companies. So there's like this small group of companies that are - People forget in the ramp data, like adopting AI can mean like having a ChatGPT pro account for someone, which is like not exactly the same as like using codecs and like coding agents and stuff. Like it's important. I think that, you know, if I have someone on my team, I want them to be able to go and do a deep research report.

19:46but that's like table stakes. The question is like, are you actually speeding up anything that's repetitive in your job? And that diffusion is just starting to take hold. So the total market size in terms of IT and knowledge management in corporations, it's about$6 trillion, right? A year. The two main frontier AI model companies, OpenAI and Anthropic, I'm going to say, I think this is roughly accurate, are doing$120 billion combined in ARR. Yeah. Why can't that go to 200, 300, 400 billion in the next year or two? I mean, they're growing at exponential rates. And we're taking off. And if the market is$6 trillion, right, why can't they grow to 200, 300, 400 billion in the next couple of years?

20:32I mean, it's like just do a little logic and rational deduction. This is definitely possible. And it's happening right now. And it's accelerating. and people aren't, you know, they're just taking, you know, these big headlines where we had this, you know,$50 billion for Financial Times, and we find out it's over 30 years. It's like on the homepage. Wait, wait, yeah, yeah. Okay, I wanted to ask you about this. NVIDIA revealed this tenant for$50 billion data center that will use its chips. Explain what is actually going on here. So the Financial Times, you know, put on their homepage today, that NVIDIA is going to backstop a lease for a data center in Texas for$50 billion.

21:13And I saw that. I was like, oh, my gosh. Oh, that doesn't sound good. No, it literally sounds like they're buying their own chips. It sounds like the most bad thing you could do. Then they actually read the article, like halfway down the article. It's like a 15-year lease, and it's only$50 billion if they renew the lease after 15 years. So it's like over 30 years if they renew it. Then if you think about that, you're like, wait a minute, 50 billion divided by 30 if they renew it. Okay, yeah. NVIDIA's 15-year lease commitment for the Texas site is worth basically 20 billion. And renewal options would take the total value to 50 billion over 30 years according to HUT8.

21:55Tae Kim:Okay, but what do you think they're, what are their plans for the site? Is this, they are going to have some, like, what do you expect them? So my point is this is a billion, you know, whatever, a billion or$2 billion a year, right? It's a non-story, but it's a big headline, sensational headline on the homepage. Yeah, and also it's not like you're taking a$2 billion loss every year. It's you are the tenant and then you are also renting that out. So hopefully you're making profit. It's a rounding error. It's like, you know, they're doing 320 billion run rate a year now. That's going to go to 400, 500 billion next year.

22:36And we're talking about something that might be a billion, you know, like this is not a story. But this is how people run with the sensationalized headlines and people panic and freak out. I think they just wanted to say the biggest number. That's exactly the point. And we're going to see what happens with this Wall Street Journal article. Both OpenAI and NVIDIA are not commenting so far. We'll see. But take us through the rumor. We have to wait. Rumor will time. Well, it's not a rumor. It's the Wall Street Journal and other people reporting that. Yeah. NVIDIA is in talks with OpenAI to backstop SoftBank up to 250 billion.

23:15We don't know the details. And I don't want to speculate and comment. But let's actually see the details before we – I think the market had a really big negative reaction yesterday to this story. Because everyone, I mean, Jim Cramer was telling his audience, like, sell everything at the open today because AI and data centers, dot com, you know, it was insane. It's just let's see the actual deal and the metrics and the numbers before we kind of can freak out. Yeah, yeah, yeah, yeah, that makes sense.

23:46Tae Kim:so yeah honestly when you say freak out and sell everything sell your dollars sell your house sell your house sell your stocks then i'll freak out but until then tay i feel i feel okay i mean i just see the fundamentals i see the ceo of aand expanding her tam you know dramatically over the last three months i see rsi under horizon like every ai researcher is like oh my god this is going to happen. We have to get there sooner. And then I see, you know, the obvious use case of agentic AI, where you have to re-architect your workflows internally. Every company has to do this. So everything is taking off.

24:27You see, like, you see when the president of Korea came to San Francisco area last week, you know, they had like a day in the Valley. Instantly, NVIDIA CEO Jensen Huang, Broadcom CEO, Huck Tan, Dario, Sam Altman are there, right? Do a little logic deduction. Why are they there like crazy? Because they need HBM memory and they're dying to have it. So if you think about that, that means there's insane demand and HBM memory is in shortage. There's tremendous demand for it, right? Talk about the, oh, sorry. Talk about the NVIDIA CUDA mode. it feels like a big piece of AMD's advanced AI event was maybe the CUDA moat isn't as much of an issue anymore in the age of agentic AI.

25:21You can have an AI agent write you the software that you need to use any chip, and that creates less pricing power for NVIDIA. But there's another world where you're not really, like, NVIDIA doesn't necessarily need a moat because everything's just growing so fast that they're still growing. But how have you interpreted the processing of, like, the potential death of the CUDA moat? So, AMD is on it. Kimmy wrote, like, a couple paragraphs in their blog post about how they created a GPU kernel and all that. So everyone, you know, gets scared or whatever. It's, like, we'll see what it's like in real life.

25:59you know this is just you know amd is incentivized to say oh kudo is not a problem anymore kuda's been a tremendous moat and i think it continues to be a moat and the reason why is uh it's super reliable all the bugs have been optimized and fixed and that comes from hitting the software you know millions of times and millions of times right like you don't know if you know if you use clock code or kimmy that what they figure out using their training data is going work in the real world right they could talk about one little piece that does well let's see how it actually works yeah but nvidia's big mode is uh its scale its co-design of actually working through the networking the cpu the gpu and how everything works together and the other big thing is their balance sheet and their ability to uh get supply commitments from you know i i think i said this before optical startups are upset because NVIDIA secured all the supply for all the optical components.

27:02Same thing with TSMC wafers. Same thing with HPM memory. So NVIDIA is using their size and Gorilla and being able to prepay and get components that are in shortage. So they've become the dominant, over the next year or two, you're going to see NVIDIA able to add tons of revenue because they were able to lock up all the supply components. That's another thing that people don't really talk about is their supply chain and their ability to work with partners and secure component inventory. Is there still energy FUD that we would run into an energy bottleneck before we run into a chip bottleneck? So Jensen said this last week on the Bloomberg interview that there are a lot of bottlenecks, including data center shell, power, and all those things, components, energy, whatever.

27:49So all those things. It sounds really bad, right? And then right after that, he said, I think the chip industry has enough supply to double their revenue every year. Basically implying NVIDIA has enough supply for energy and all that stuff. No one is pricing that in. So everyone talks about bottlenecks. NVIDIA CEO just basically told you on Friday that they have enough supply chain and all the bottleneck stuff to double revenue every year. And no one, you know, NVIDIA's revenue estimates for next year are a lot lower than double. I'll tell you that.