🤖 AI capex outruns app-layer revenue - Nvidia's terminal-value problem

“I think the revenue scale will continue for like the next two or three years probably for Nvidia. But the real question is what is the terminal value? ... People ultimately realized that the value was going to go to other parts of the stack, the application layer.”

“The real question ultimately will be, does the initial cost of the infrastructure exceed the ultimate value that's going to be realized on the application layer? ... you could look at Cisco during the early days of the Internet build out, and everyone thought Cisco was the picks and shovels of the Internet, and they were going to make all the values go to Cisco. So we're kind of in that same phase right now with Nvidia”

“As it turns out, all that fiber eventually got used. The Internet went from, you know, dial up to broadband ... So I think that the history of these things is that the applications eventually get written. They get developed if you build the infrastructure to power them ... we're just at the beginning of a wave that's probably going to last at least a decade”

“what he's basically saying is it's still more artificial than intelligent. And everybody needs to take a deep breath and understand that there's just going to be a lot more work before you get to this omnipresent agent that just replaces and destroys everything and thinks on its own.”

“And so AI is a bunch of, no offense, probabilities and statistics today... the great leaps in productivity, Friedberg, that you're talking about are probably a three to five year phenomenon, not a three to five month”

“I think the world needs a lot more AI infrastructure, a lot more than it's currently planning to build and with a different cost structure.”

“But I just think they're going to be able to unleash a whole lot of new applications. And if they're metering the usage of the models and providing the best dev tools, I think there is a business model there.”

“I think that's why these tools companies or the tools products that Google, Microsoft, Amazon, and a few others are building are actually incredible businesses because so many enterprises and so many vertical application builders are going to be able to leverage them to rewrite their entire business functions.”

“And then their average growth rate from 1998 to today was negative 1.4% a year. So it went from about 200 billion to about 130 odd billion. And why? It's not that Intel was a worse company, but it's that everything else caught up. And the economic value went to things that sat above them in the stack. Then it went to Cisco for a while, right? Then after Cisco, it went to the browser companies for a little bit. Then it went to the app companies. Then it went to the device companies. Then it went to the mobile companies. So you see this natural tendency for value to push up the stack over time. And in AI, we've done step one, which is now you've given all this value to Nvidia, and now we're going to see it being reallocated.”

“So I think that they're going to have less maneuvering capability than Cisco had in the future. And obviously, there's this deep concentration risk, which is going to be deeply challenging.”

“So while this is incredible for NVIDIA, the chicken is coming home to roost because if you do not start seeing revenue flow to the bottom line of these companies that are spending $26 billion a quarter, the market cap of NVIDIA is not what the market cap of NVIDIA should be.”

“So I do wonder if there is a slow reckoning underway right now in technology that's, as we talked about in the show, a function of both kind of an economic slowdown.”

“No, no, that's happening in startups right now. I'm seeing startups with four developers do what just a couple of years ago, they would need eight to do there.”

“It is true that we've spent probably $750 billion to a trillion dollars collectively now on things with the AI label, but it is also true that that's probably generated less than $10 billion of revenue.”

“The thing to keep in mind is that I still haven't seen enterprises building production level products using AI.”

“we're in the AI dial-up era, to use an analogy here. Hey, you can see the promise, but it kind of sucks in terms of reliability.”

“all of these AI tools can add a 10 or 15% lift to an individual person if you are in a traditional organization. I don't think these things are the panacea that they're marketed as being. These are not creating 10x engineers.”

“I agree with Chamath that it's just too soon for AI to be responsible for this. I mean, the AI productivity gains are just starting, and we're not really seeing job elimination yet or job replacement.”

“Well, I think we're all converging on the same realization, which is you can't spend a trillion dollars and only have toy apps to show for it.”

“So if you fast forward another 24, 36 months, I do think that there's a great reason to be optimistic that there's going to be extraordinary ROI based on the infrastructure that's being built.”

“I think that's what's likely to happen. I'm much more bullish than you guys about this investment that's being made.”

“But I think we're going to see a tipping point next year where a lot of labor arbitrage occurs.”

“No, I think it's definitely this mini AI bubble deflating a bit.”

“So, I suspect what happens is that the overall market grows, but the number of companies grows by an even larger order of magnitude. And in all of that, the reality is that it's going to be very hard for these big folks to sort of see this value captured that makes any of this investment worthwhile. So, I think that you're going to have to have some sort of reset in terms of the capex that's happened here.”

“So it doesn't matter how many tens of billions you're spending each quarter right now. You have to get there. You have to make sure you don't miss the boat.”

“let's rationalize the capital to when are we expecting revenue? How do we get revenue sooner to have that as a good productive cycle? How do we be not trying to just spend like drunken sailors, which is easy to do, but to be targeting business outcomes?”

“if you just look at the competition set that they're up against, they're losing 5 billion a year, they're making three and a half, they put this thing at 150 billion, it's 40 times, 50 times revenue. To fill in that valuation on a price to sales basis, it's kind of crazy.”

“And ROI has gone vertical since they ramped their capex on GPUs. It actually just started to level out in this latest quarter. So the ROI on AI has been very positive thus far. Just a fact”

“And I think it really showcases the kind of the impressive impact that AI is having on workplace productivity”

“And so I just see this world where it's all getting commoditized quite quickly. And so I'm just like sort of scratching my head, like where is the market value here?”

“I still think you could probably have the entirety of the model providers make 10 times more revenue than they do today because we're just literally in the first percent of the total TAM.”

“Yeah, I'm really excited to talk about it because I think I've been trying to figure out how they justify their valuation over the long run. And I'm not a public market person, but I am fascinated with Nvidia. And their Cloud GPU business is definitely a majority of the revenue. So I think a lot of what we're seeing is them trying to grow into that and trying to expand in case the music stopped. Now, I don't actually think the music is going to stop.”

“I think that value and money are correlated in some industries, like real estate or heavy industry, but I don't think that it's very correlated in tech.”

“So I don't think it's over done and blown. I think this is going to be, as you guys have said before in the podcast, bigger than Industrial Revolution, but it's also true that you'd need to have a good ROIC.”

“So basically, they would need to make an incremental $27 billion of operating profit a year on the $75 billion for this to meet their ROIC performance. That doesn't seem crazy because that's just under 20% of their annual operating profit.”

“I think that I was completely wrong on a couple of my earlier thoughts. One of the thought that I had for a long time was, it just seemed like all these base models were asymptoting. So I was not convinced where all this CapEx would go in a productive way.”

“But if agents become a reality, one, the ROI on AI and Blackwell is going to be very high, and two, what will, it's not good for human employment, but what will be kind of, I would say, the rate-limiting factor is just compute.”

“I mean, the amount of money that's been poured into AI has not yet been justified by the amount of net on the return, on the actual investment, obviously.”

“I think it's going to hold up. I mean, look, I just think that this investment in CapEx and the data center build outs is so strategic right now. I mean, I think that the hyperscalers, it's only good for their businesses, but I think they see it as so strategic to their survival that they just have to do it no matter what. In other words, I think they're totally inelastic on this.”

“What I would say is that what they are encountering is the trough of disillusionment... And I think what has happened now is people have spent billions and billions of dollars. I think you can see it in the revenue traction of the AI companies. But I think where we are today is that there are some real technical complexities that have not been solved.”

“So we went from basic LLMs, Totally. You know, which don't require that many tokens to give you an answer, to the reasoning model, where you can spend a thousand times more tokens just getting one answer to a question. And now the agents are going to be even more token intensive than that. So the amount of compute required to serve all these new applications is going to be massive, which is why I think the CapEx build out actually makes sense.”

“Microsoft said that in their Q1, they processed 100 trillion tokens, 50 trillion alone in March. And so the tokens are really going basically vertical, which is probably because of these reasoning engines, which are much more sophisticated and require more compute power.”

“I think there's a billion people in the developed world who could spend, let's call it a hundred a month, $1,200 a year, it doesn't take a genius to figure out this is a trillion dollars in revenue based on market cap. That's probably $10 trillion in market cap.”

“Is this stuff going to hit fast enough to actually justify the amount of investment, or are you going to see a crater here?”

“a billion business users in the world, it's going to put you at a billion dollars in revenue, which would be like a $5 to $10 trillion TAM just for the business side”

“You can calculate something called return on invested capital. It's actually gone up since they started making these AI investments.”