The Exchanges

Q And thanks for taking the time out of the sprint to do this. I know it might sound redundant talking about the size of the market today, but it is astronomical. It is crazy. So could you talk about the size of the mag seven when they IPO versus today, the AI leaders and their size? It's a wild comparison.

A Yeah. Um, So the biggest of the Mag-Seven was Meta in 2012, IPO'd around a hundred billion dollars. All of the predecessors had IPO'd in the nineties, you know, in the 10 to fifty billion dollar range, which for that time was very large. But today you look, um, OpenAI's most recent round was 800 plus billion dollars. Um, SpaceX is most recent, you know, when they did the transaction with XAI was 1.25 trillion dollars. Um, Anthropix last round was, you know, high 300 of billions. Um, it's rumored to be significantly higher their next round. And so the fact that you now have Private companies that are breaking into pretty much the top 25 in the world before they even go public. It's an unprecedented time, but I think that's what makes AI so exciting, is that you have these private companies that are capturing so much value and so much growth while staying private, and I think that that is like nothing we've ever seen before.

AI assessment note: “So the biggest of the Mag-Seven was Meta in 2012, IPO'd around a hundred billion”

Q In our interview with Thomas, he was talking about the bank teller and how that actually evolved. It didn't really actually decrease any jobs. There were just more jobs spread out. Um, so that's always a hot topic. So in talking about market dynamics, you have this framework of sellers of shortage versus buyers of shortage. Can you break that down? Who's winning and, and what's kind of the explanation?

A Yeah. So we try to simplify sellers and buyers A shortage in the concept of you have semiconductor companies, you have power, you have memory, infrastructure, everything where, um, you have a fixed amount of capacity and you have greater demand, and that greater demand is driving price increases, margin expansion, and when price is the main lever of your revenue growth and you have fixed costs, your operating profit actually go up multiples of what your price increases or your revenue is growing at, and so we've seen companies that have been able to increase their earnings Three, four, five, six, seven X, just in a matter of a year or two. Companies that would have been plus or minus a certain range for, you know, 1020 years. And so, um, these shortages, which are critical to AI, right, whether it's memory, hard drives, like their shortage to kind of the development of AI, have driven significant earnings power, and the market is rewarding that set of earnings. What the market is not rewarding, and not rewarding as much, is the, the companies that are buying That shortage. And so that would be the companies that are putting the capex into the ground. And so if you look at the multiples of Microsoft, of Amazon, of Meta, they've actually all compressed in the last couple of years because their capex is going up. And if you just think about it, if memory pricing goes up a hundred …

AI assessment note: “we try to simplify sellers and buyers A shortage in the concept of”

Q lots of data. I've had a couple conversations with leaders at COTU previously, Michael Barton, Thomas LaFont, and we've walked through all the different stages of the tech cycles. So, We walked through internet, mobile cloud. We're now in the AI era. What is going on here? I feel like every six weeks we need an update at this point. It's just moving so fast. So what are you seeing?

A Yeah. I mean, honestly, what you said is, is the reality that AI is big and everyone can make these grand statements about how big it is. And we've also tried to do that a little bit in our slides about trying to size the TAM. But the most exciting part is the pace of the innovation. And you can look at that across. You know, how quickly companies have reached, you know, ten billion dollars, 30,000,000,050 billion dollars of ARR, which, you know, we talk about quite a bit, how OpenAI reached, you know, almost a billion users in the fastest time in history. Like all of the curves are so much steeper and faster of adoption, whether it's at the consumer level, the enterprise level, revenue level, like however you think about it. And so to us, we always think about rate of change. Like that's how you really define technology and how quickly it's, you know, catching on and how big it can be. And a lot of times you see quick rate of change and then flattening, right? A lot of times when we look at apps that used to be popular, there was an app that might've done something, it gets to like ten million users, and then it all of a sudden flatlines. But the fact that you're hitting these metrics, whether it's, you know, top-down metrics or user metrics that just continue to go and continue to grow at such hyper rates, and sometimes they actually are seeing a, almost like a little bit of …

AI assessment note: “most exciting part is the pace of the innovation... all of the curves are so much steeper”

Q liquidity, access your funds anytime. Companies like Scale AI, DoorDash, Service Titan, HIMS, Anthropic, Flexport, Robinhood, and Plaid trust and use Brex. Start today at brex.com slash sorcery. That's B-R-E-X dot com slash sorcery. Turing is training the next generation of AI with tasks that require real expertise and real world judgment. You've talked about how agents have become one of the biggest unlocks. So what is the difference there?

A So if you think about your initial chat GPT interaction, you would ask it a question. It would think for some amount of time, depending on the complexity of the question, and then it would come back to it with an answer, or not even a lot of times an answer. Maybe it comes back with an intermediate step of, okay, I searched all of X, Y, Z. I did this. Am I on the right track? Do you want me to go more in detail? Do you want me to do less? Yeah. And it was a lot of human in the loop. An agent now, when you give it a problem, like the, the, the, the amazing part of Opus Um, when it launched at the end of last year was that an agent can now spawn its own agents, and that can increase the depth of work, the time of work, and the quality of the work. And so all of those things, that was a huge model unlock. Agents, agents launching agents, I think is one of the most underappreciated unlocks that has happened, because you can give an agent a project now, and you can just go away, and you can come back, and it has done If not all of it, really, really far along. And the amazing thing is you can also, when you instruct it and prompt it initially, you can say, go do this, launch as many agents as you want, and then launch another 10 agents to check all of these agents work. So not only like, it's so collaborative and iterative now, where the human is almost out of the loop. And if you t…

AI assessment note: “An agent now, when you give it a problem... an agent can now spawn its own agents”

Q As the CIO of KOTU's public sector, I would be remiss not to ask you about risks. What are the biggest risks that you're watching?

A Um, look, the, the biggest risk is that there is some technology on the other side that materially changes where some of these shortages exist today. Um, and that risk exists only in the grand scheme of, okay, the types of stocks that would be impacted by that. If there was, let's say, if let's say that the deep seek moment that happened, uh, you know, last year, another moment like that happens where someone figures out a model that can do all the calculations with less power, less Semiconductors, less memory, all of that. Um, if that happens, it is probably good for the long run of AI, because that means that the adoption is going to happen faster, more people are going to use it. If AI becomes materially cheaper, there's the whole Javon's paradox argument, which is like, yeah, the cheaper it becomes, the more people will find creative ways to use it and do other things, right? Humanoids are still kind of in the back burner, but maybe if it gets cheap enough, like, there'll be way more dollars that get put behind solving that problem, and all of a sudden, like, Other things get accelerated, and so I think that, like, that is a risk that I'm looking at as, you know, someone who has a mark-to-market portfolio every day that I'm thinking about, um, and then, you know, there's, could there be some regulation that comes out of nowhere, um, that maybe changes the, the, or dampens t…

AI assessment note: “the biggest risk is that there is some technology on the other side”

Q Because of those behavioral changes, how is that affecting the architecture of AI?

A I think that we're now moving to a world where the chatbots, you know, we joke around internally, like the chatbots have amnesia, right? Because every time you, you might give it some files, you give it some medical records, it does a great job of going through and analyzing it. You go back the next day and you give it the same record, or you have to, you ask a question about those records, it doesn't remember. Um, I don't know if you ever watched that movie, 51st Dates? Yeah. Where basically he has to re-explain every day, you know, what the situation is. It's kind of, it's very similar to that. But what's changing is now you're getting persistent memory with all of the, the next evolution of, um, accelerators. And so I think that you're seeing memory content is increasing. The agent is not going to have amnesia anymore. The agent is going to be able to remember. And now because the agents are launching agents and they're doing very complex tasks, but also simple tasks, they need offload from CPUs too. Like historically CPUs is, have done something called serial processing, which is like you just give it a set of instructions and it does one by one by one by one critical instructions. But one in a row. Um, and I think that as an agent has some element, which is like, okay, make a reservation. That might just be very serial tasks. That might have something which is like, okay, …

AI assessment note: “creating a massive need for more CPUs and more memory”