The Exchanges

Q And, and, and just tell us, give us that quick thumbnail on semi-analysis today. Like what is the business?

A Yeah. So today we are a semiconductor research firm, AI research firm. We, Service companies, our biggest customers are all hyperscalers, uh, the largest semiconductor companies, uh, private equity, as well as hedge funds, and we, uh, sell data around where every data center in the world is, how, what the power is in each quarter, what, how the build outs are going. Um, we sell data around, uh, fabs. We track all 1500 fabs in the world. For your purposes, only 50 of them matter, but like, you know, all 1500 fabs around the world. Uh, same thing with, uh, the supply chain of, like, whether it be cables, or servers, Or boards or transformer substation equipment. We try and track all of this on a very, uh, number driven basis as well as forecasting. Um, and then we do consulting around those areas.

AI assessment note: “today we are a semiconductor research firm, AI research firm.”

Q We'll see, but I'm thinking very similarly to you. But it is, and I think there's, you know, the market is seasonal. Yes. I think there's real, real concerns around inflation and rates. What was CPI this morning?

A It was a 4.2. I think we added core came in at like .2 versus .3. So a little bit better. Um, but you know, clearly we're, we're above four again and, um, and, and there's short term pressure on, you know, core PCE, et cetera. Um, and we have some unknown unknowns, but the market, I mean, if I had told you the fact pattern for this year, that we're going to be in a war with Iran, that, you know, oil was going to be at a hundred bucks, that, CPI was going to be creeping back up. The internet was going to be down 15%. Software is going to be down eight percent. You would have said, I want nothing to do with that market, right?

AI assessment note: “It was a 4.2. I think we added core came in at like .2”

Q Can you, can you think of an entity, like a comparable entity in the past? I, I'm, I'm having a hard time, like, imagining what that, what Brad just described actually is.

A I don't think there is a good comparison on this bill, because when you, when you just think about the amount of capital that's required, it's bigger than anyone, right? So it, this, this required a very, very novel set of partners, To come together, both with a, a big vision, uh, a, a, a large opportunity to get access to capital, and candidly, probably a little bit of a willingness of, we're gonna figure this out as we, we try to grow it, because we're, we're, this is not, this is beyond what we've done before, and I, the only analogy, and it's not, it's not a good one in terms of how I can think about it, is, is maybe with Global Foundries and Mubadala, relative to starting to see that traditional fabs, Needed to get extra capital, and that was at a much smaller scale. Uh, now, you know, Satya talking about spending eighty billion dollars, uh, of CapEx, you know, even, even for Microsoft, it's a giant number. Uh, and at these numbers, no, no comp, no one company can do it.

AI assessment note: “I don't think there is a good comparison on this bill”

Q You said if they stand still there, they would, they would have competition. Like what, what. Would be their area of vulnerability or, or what would have to play out in the market for other alternatives to take more share of the workloads?

A Yeah. So, so the main thing for NVIDIA is, you know, hey, this workload is this big, right? It's, it's, it's well over a hundred billion dollars of spend. Um, for the biggest customers, they have multiple customers that are spending billions of dollars. I can hire enough engineers to figure out what, how to run my model on other hardware, right? Now, maybe I can't figure out how to train on other hardware, but I can figure out how to run it for inference on other hardware. So NVIDIA's moat in, in inference is actually A lot smaller on software, um, but it's a lot bigger on, hey, they just have the best hardware. Now, what is, what does the best hardware mean? It means capital cost, and it means operation cost, and then it means performance, right? Performance, TCL. Yes. Um, and Nvidia's whole moat here is if they stand still, the performance CCO doesn't grow. Um, but interestingly, they are, right? Like with Blackwell, not only is it way, way, way faster, anywhere from 10 to 15 times on really large models for inference, because they've optimized it for very large language models, they've also decided, hey, we're gonna cut our margin, too, somewhat, because I'm competing with, uh, Amazon's, you know, chip, and TPU, and, and AMD, and all these things. They've decided to cut their margin, too. So, so between all these things, they've, they've decided, That they need to push perfo…

AI assessment note: “NVIDIA's moat in, in inference is actually A lot smaller on software”

Q Um, it's not just like- Broaden out beyond Broadcom. I'm talking NVIDIA and everybody else. Like, you know, we've had these two massive years. Right, of tailwinds behind this sector. Is 2025 a year of consolidation? Do you think it's another year that the sector does well? Just kind of

A Yeah, I think, I think the plans for hyperscalers are pretty, uh, firm on, they're, they're gonna spend a crap load more next year, right? And therefore, the ecosystem of networking players, of ASIC vendors, of, uh, systems vendors is gonna do well, whether it be NVIDIA or Marvell or Broadcom or AMD or, you know, generally, you know, some, some better than others. The real question that people should be looking out to is, twenty-twenty-six, um, do, does the spend continue, right? We are not, the growth rate for NVIDIA is gonna be stupendous next year. Right, and that's gonna drag the entire component supply chain up. It's gonna bring so many people with them, but twenty-twenty-six is like where the reckoning comes, right? Um, but, you know, will, will people keep spending like this? And it's, it's all points to where, will the models continue to get better? Because if they don't continue to get better, um, in my opinion, will get better faster, in fact, next year, then there will be a big, you know, sort of clearing event, right? Um, but that's not next year, right? Um, you know, the other aspect I would say is there is consolidation In the NeoCloud market, right? There are 80 NeoClouds that we're tracking, that we talked to, um, that we see how many GPUs they have, right? The problem is, nowadays, if you look at rental prices for H-one hundreds, they're tanking, right? Not jus…

AI assessment note: “they're gonna spend a crap load more next year, right? And therefore, the ecosystem”

Q a development sandbox world to a optimization world. Is that likely to happen? Is there an equivalency here or not? And if you could touch on why the, the, the back end is so steep and cheap, like, you know, just, you, you go a model, you know, behind, or you, you, like, the, the token, the price you can save by just backing up a little bit is nutty.

A Yeah. Yeah. So, um, today, right? Like, oh, one is stupendously expensive. You drop down to four. Oh, it's a lot cheaper. You jumped down to four Oh Mini. It's so cheap. Why? Because now I'm competing with four Oh Mini, I'm competing against Lama and I'm competing against Deep Seek. I'm competing against Mistral. I'm competing against Alibaba and I'm competing against tons of companies. I think, and in addition, right, there is also the problem of inferencing a small model is quite easy, right? I can run Lama-seventy-b on one AMD GPU. I can run Lama-seventy-b on one NVIDIA GPU, and soon enough there will be on one set of Amazon's Neutronium, right? I can sort of run this model on a single chip. This is a very easy, no, I won't say very easy problem, still hard, but it's quite a bit easier problem than running this complex reasoning or this very large model, right? And so there is, there is that difference, right? There's also the fact that, hey, there's Literally, 15 different companies out there offering API inferences, inference APIs on Lama and Alibaba and Deep Seek and Mistral, like these different models, right?

AI assessment note: “Why? Because now I'm competing with four Oh Mini, I'm competing against Lama”

Q I think a lot of the, um, people in our industry have kind of been on a pro-nuclear bandwagon past five years. Um, and, and that appears now to be something that's kind of tipped Across the US. Take us back to prior to that starting to happen. What were the forces that were pushing for decommissioning?

A So, the forces for, ah, the main force was, ah, the, the, the projections by, ah, the, the Energy Commission and other energy advisors were showing that there wasn't a need for Diablo. That was the main driver. So, there were a number of things that were going on in 2016, and what was important was that we were being used for a bridging strategy to get to have those other renewable projects come online to replace us in 2024 and 20 25. So, here we are. Unit one's current license expires November second of 2024. And it'll be really interesting to look at what's going on November third and fourth with our grid had we taken unit one off. So we're right there, right? Uh, unit two's, uh, current license expires in August of 2025. But we've put in for a 20 year, uh, uh, license renewal application with the federal regulator, the Nuclear Regulatory Commission. Even though the state's only asked us for five, We've done that to make sure we maximize the optionality for the state and seeing the progress of those other projects coming online to replace Diablo.

AI assessment note: “the main force was, ah, the, the, the projections by, ah, the, the Energy Commission”

Q Oh yeah, yeah. 15 trillion tokens used to train, uh, you know, uh, the model. I know at Grok you guys are deploying Lama three. I think you deployed it the same day that it came out. Um, so how important is this? How important a development is it in the world of models?

A Well, really, you know, Zuck came out and threw down for the entire world of folks that are building models. And it's, it's really disruptive because, um, when you look at the rankings, you have a model that's much smaller, so much easier to run on all different types of hardware, uh, and much faster. And so those two things are like catnip for developers. And for us, we saw within the first 48 hours, it become the most power, most popular model that we run on Grok. And so really replacing what? Replacing Mixtrel eight by seven for us, which was, you know, generally considered the best open source model at that point. And what the capabilities have happened beyond us sort of running it, the developers that use it, the use cases we've seen it in are incredible. And people are doing a direct replacement with, with open AI across the board. They come to us. So they come to, you know, all the different providers and they replace out, uh, open AI. And they don't really see any performance impact or any reasoning impact or, which is.

AI assessment note: “it's really disruptive because, um, when you look at the rankings”

Q Hey, Ern, a quick question for you. Obviously, there are a lot of software companies in the Valley. Based on what you've been through, if you were advising them on what metrics matter most, What's at the top of the list for them to pay attention to internally?

A The, there's probably a few that are just like literally the underlying business model economics. Um, I, I like, I'm a big believer in gross margin. Um, uh, your, your gross margin will ultimately determine your operating margin. Um, there's almost no way that, that you can, you can kind of make those two, um, you know, kind of get out of sync. Um, and so if you're, if you're subsidizing something or, you know, or, or in just such a commodity business and you're, you know, 40, 50, 60% gross margin, like there's just like, No way you're gonna have an operating margin that looks like a software company. Understanding your gross margin, managing to gross margin, I think is super important. Um, uh, you know, all forms of LTV CAC are, are probably good. I don't know what the latest, you know, make, you know, everybody, every two years is a new term in the industry that, that is used, but like something that just shows that you can acquire a customer profitably and whether the payback is a year or two years or three years, almost doesn't matter as much as, um, as, uh, do, do, you know, ultimately generate a, A long, a long-term, you know, sticky customer. Um, I think cashflow is, is super important. Um, I'm, I've, I've definitely like, I've, I've gotten religion on cashflow. Um, and I think companies, uh, getting the cashflow sooner, um, is a, is a really good move. I think it, it pu…

AI assessment note: “I'm a big believer in gross margin... all forms of LTV CAC... cashflow is, is super important”

Q And you went so far as to suggest This might allow them to pass the best proprietary models coming out of the US, um, yet this year, maybe by Q four of this year. So why don't we dig in there? What is your theory of the case? Why is China, you know, doing so well in open source? And should US model companies like OpenAI and Anthropic be concerned?

A Yeah. So, uh, let, let's kind of tie it into, I think three important things that we see happening. The first one being, you know, the Chinese and, and the president addressed this at the AI summit, right? He addressed, um, the, the point around using copyrighted work and he says, you know, he used a great example. If you read a book and you use it, you're not violating the copyright there. And so he addressed that concern. And that was one of the major things that, you know, a lot, a lot of people didn't talk about, but I think it's important for the model makers. And so the Chinese just have been able to work around that because of, you know, their position on IP. And what we're really seeing here, and I think, you know, Bill teed it up even better off of my tweet, which is, um, you know, they're able to compound. So what you're seeing very quickly is both the open source nature, the open weights nature, uh, allow them to basically compound on each other. So instead of working in silos and instead of having to create giant training clusters, Um, separately. They can basically take each other's work, build on top of it, almost consider it like a remix of someone's model. K-II, sort of a well-known remix of what, what Deep Seek had done, and now we're starting to see that happen really fast, and we're seeing two dimensions of it going quickly. One, we're seeing the leading edge…

AI assessment note: “both the open source nature, the open weights nature, uh, allow them to basically compound”

Q everybody in the world to use that model. And do you believe they have a shot at out-competing, being the upper right, you know, by the end of the year? And if so, do you think that will be the outcome? Like these companies that are using Quen on Grok, do you think they would prefer to use OpenAI so long as it was equally capable and equally priced performance?

A So two things that we see is brand and, you know, the US domiciled or, you know, someone that they can kind of point at that wins. And so if that shows up, it will win because if you're a company and, you know, at some point you have to, you know, get your teams to sign off on what is it that you're using? What are the risks associated with it? And like, you know, who is liable if something goes wrong? And so I sort of feel like, Um, with OpenAI's release and, you know, Meta charges back, or even if some of these startups emerge that, you know, we can point at, I think we'll see a huge shift back towards those models versus, versus the Chinese ones.

AI assessment note: “I think we'll see a huge shift back towards those models versus, versus the Chinese ones.”

Q Yeah. No, no, which is going back to the point that we made on open source, but if you know in the back of your mind that there's something that's 90% as good, but it's 90% cheaper, how does that factor in? Because we've also never had that, that factor as we're going through this growth curve.

A I mean, since, almost since we started the pod, you know, I've, I've routinely highlighted that the steepness of that price curve on, you know, As you, as it kind of becomes, you know, less, less cutting edge is something I've never seen before. I've never, ever seen it. And I'm sure that a lot of, um, people sit around and say, well, it's okay if I'm losing money here because, you know, six months from now, I'll just use the older model. And we also talked in the past about how in the internet age, all the startups began with Oracle and Sun. And eventually they all moved to Linux and MySQL. And so there was a, there was a, we got to win it all cost phase. And then there was a phase where you started worrying about cost and optimization. And so one day, one day we'll likely, you know, make that, make that move. And, and, and I, a few of the companies I've talked to that are running inference at scale, they are already starting to think that way. Like they're looking, they're looking at it, you know, from that way. From that lens.

AI assessment note: “it's okay if I'm losing money here because, you know, six months from now”

Q Totally. Well, in, in the spirit of storytelling, Michael, do you remember when you first met Gurley?

A Yeah, I do remember when I first met Gurley, uh, this was, this was in the, in the nineties, and Bill had written this research report that was super thick, and I'm reading this report, I'm like, how the bleep, bleep, bleep, does this guy know more about our business than we do? It's like, what? We must be totally screwing up here, and, and, uh, he had, Uncovered a whole bunch of analysis and thoughts about our business, and we were so busy, uh, kind of distracted by growth that we had missed a few things, and Bill shined a massive light on that, and it was super helpful. So I, I, I, uh, became a fan instantly of his work and have been a fan ever since.

AI assessment note: “Yeah, I do remember when I first met Gurley, uh, this was”

Q Of bringing in employees that make radically different amounts of money than the other employee base. How, how do you think that will be? How difficult will that be to manage?

A I think it'll be a challenge culturally, for sure. Um, you know, uh, he could have a long line outside of his door, uh, with people, uh, you know, wanting this or complaining about that, and that could be a distraction. So, you know, I think people, uh, generally have a sense of fairness, right? And they, they want to be treated fairly. Um, relative to others and relative to the, the opportunities that they have out there in, in the overall market. And if they feel that they're not being treated fairly, uh, that's going to be a problem. So I don't know how that, I don't know how that gets sorted out. I do think the math could work for them, given, given everything you guys just talked about. Uh, and, and obviously if you, uh, reduce this down to a race to super intelligence or something along those lines, the size of the prize is, is tremendous. And they do have an incredible business that is aided by these advancements, uh, in a, in a big way. And there aren't a whole ton of companies that can go do this.

AI assessment note: “I think it'll be a challenge culturally, for sure.”

Q capitalist from birth is going to better align us with the policies that allow the company country to continue to grow. Um, and I think growth is a critical element to making sure that we get our deficit to GDP back in a, uh, you know, in, in a manageable place. Michael, I know you care a lot about that issue. Any, any other thoughts on, on that particular point?

A Yeah, I mean, government's obviously been spending too much and, uh, there's been some, Some renewed attention and focus on that. That's a good thing. Uh, it gets priced into, to the currency, right? And we see it in all the effects, uh, You know, whether it's inflation or the value of the currency. And, uh, you can't really, uh, escape that. I think the, the, the, the spending has to come under control. Uh, now maybe we get this incredible, uh, productivity lift. I'm sure we're going to talk about that as we get to the AI fund portion here. Um, but We shouldn't be spending, uh, so much more than, than we're taking in as, as, as a government. We, you know, we, I've sort of stepped back from the hysterics and you say, we don't have a loan to value problem as a country. Uh, we, we, we have a spending problem.

AI assessment note: “we don't have a loan to value problem as a country. Uh, we, we, we have a spending problem.”

Q It's truly an Apollo scale project. I, I, I guess as you think about, again, ARM, why don't you explain to us again, what is, what is ARM actually delivering into Stargate? Um, I know you're, you know, you're embedded in the GB 200, but maybe just share with everybody else the role you play.

A Well, maybe at the highest level, the way to think about it is you've got a, a giant, as you said, Apollo, Manhattan Project, whatever terms you want to use about Well, I guess it's probably the largest infrastructure to build out in the history of the world. And every data center, uh, whether it's running general purpose compute, whether it's running inference or running training, needs a base CPU to run everything, end quote. And that's our role. And whether that is going to be what we, uh, are part of today, which is GB 200, and we're super happy to be partnering with NVIDIA on that product. Uh, or other areas that we haven't talked about yet in terms of productization, there's lots of opportunity for ARM, because the, the base CPU will be ARM. And I think therein lies a huge opportunity. You know, one of the things that people don't always appreciate with, let's take, again, GB 200 running in an AI data center, all of the other work that needs to take place, whether it's the hypervisors, virtual machines, anything that the end quote Normal CPU has to do in a data center has to be run by something. And that's what, that's what grace does. So then when you baseline that relative to, okay, GB 200 is where we are today, the opportunity going forward in terms of these large data centers doing, um, some level of mixed inference and training, uh, reasoning, uh, reinforcement train…

AI assessment note: “needs a base CPU to run everything, end quote. And that's our role.”

Q and he talked about how he just built layer after layer after layer of the stack, you know, over the course of the last decade and a half. But when he said that, Sonny, I know you had a reaction to it, right? Even though you know it's not just a GPU company, When he really broke it down, it seemed like, you know, he did break new territory here.

A Yeah. Like what was great to hear from him and really, you know, positive for, you know, folks thinking about where Nvidia lives in the stack right now is he kind of got into details and then the sub details below CUDA. And he really started going into what they're doing very particularly on mathematical operations to accelerate their partners and how they work really closely with their partners. You know, all the, the cloud service providers. To basically build these functions so that they can further accelerate workloads. The other little nuance that I picked up in there, he didn't focus purely on LLMs. He talked in that particular area about how they're doing that for a lot of traditional models and even newer models are being deployed for AI. And I think just really showed how they are partnering much closer on the software layer than the hardware layer alone.

AI assessment note: “what was great to hear from him and really, you know, positive”

Q mix? And he said, of course. Right. But again, I think conventional wisdom is all around the size of clusters and the size of training. And if, if models don't keep getting bigger than their relevance will dissipate, but he's basically saying every single workload is going to benefit from acceleration, right? It's going to be an inference workload and the number of inference interactions is going to explode higher.

A Yeah. One, one technical detail, which is you need bigger clusters if you're training bigger models. But if you're running bigger models, you don't need bigger clusters. It can be distributed. It can be distributed. Right. And so I think what we're going to see here is that the larger clusters will continue to get deployed. And as Bill said, they'll get deployed for folks, maybe a limited number of folks that need to deploy it for a hundred billion dollar runs or even bigger than that. But you'll see inference clusters be large, but not as large as a training clusters and be a lot more distributed because you don't need it to be all in the same place. And I think that's what'll be really interesting.

AI assessment note: “if you're running bigger models, you don't need bigger clusters. It can be distributed.”

Q And make the argument, why, why are they faster? Why are they cheaper in your mind? But yet, notwithstanding that fact, NVIDIA is going to do, let's call it, 50 or sixty billion of inference this year. Um, and these companies are, you know, still just getting started, right? Why is their inference business? Is it just because of installed base?

A Yeah, I think it's a combination of install base. And I think it's because that inference market is growing so incredibly fast. I think if you're making this decision, even 18 months ago, it would be a really difficult decision to buy any of those three companies because your primary workload was training. And the, you know, the first part of this pod, we talked about how they have such a strong tie in integration to getting training done properly. I think when it comes to inference, you can see all the non NVIDIA folks can get the models up and running right away. There is no Tie into CUDA that's required to go faster. That's required to get the models running, right? Obviously none of the three companies run CUDA. And so that moat doesn't exist around inference.

AI assessment note: “I think it's a combination of install base.”

Q because that was all now in the P&L, you had to expense that like cash, which Warren Buffett and many others had been pushing for since 97, 98. But in order to get around that, Everybody started adjusting what they, you know, their, their earnings, EBITDA, they started adjusting the stock-based compensation expense out of EBITDA, ok, so as though the expense no longer counted. Um, did you start that?

A I don't know that I started it, but we did. So we did do the proforming, but the reason that we did the proforming wasn't because we didn't view that every single dollar of RSUs that we were handing out had real, very significant value. It was more that we were going through this transition period where Some of the compensation we had handed out historically was, were stock options, and now we were moving to a world of, of RSUs. So there was a multi-year period where some of the stock-based compensation was option-based, some of the stock-based compensation was RSU-based. So that would have created even more confusion around, you know, what, how to, how to think about core earnings. Um, and so we did pro forma them and, you know, obviously almost all companies pro forma them. Um, I think the big difference is, is that we as, as a company, we always viewed one dollar of, of an RSU as fully being worth one dollar. In fact, if you really are bullish on your company, which we were, We viewed that a dollar of RSUs was actually worth more than a dollar because we certainly had a view, hopefully, that if we did our jobs, that, that the stock was going to go up. And so that the, it's, it was our most valuable form of currency that we use to pay, uh, employees. Um, and we certainly thought about that all along the way. We never, it never occurred to us That, that it shouldn't count or t…

AI assessment note: “I don't know that I started it, but we did.”

Q You just did a super deep piece on Tranium. Why don't you do the Amazon version of what you just did with Google?

A Yeah, so, so funnily enough, Amazon's chip is the Amazon, I, I call it the Amazon's basics TPU, right? And the reason I call it that is because, yes, it uses more silicon. Yes, it uses more memory. Yes, the network is, like, somewhat comparable to TPUs, right? It's a six, it's a four by four by four Taurus. Um, they just do it in a less efficient way in terms of, you know, hey, they're spending a lot more on active cables, right? Right, uh, because they're working with, uh, Marvell and Alchip on their own chips versus working with Broadcom, the leader in networking, who then can use passive cables, right, for, for, uh, because their surities are so strong. Like, there's other, there's other things here. Their surity speed is lower. Um, they spend more silicon area. Like, there's all these things about the, the Tranium that are, you know, you could look at it and be like, wow, this would suck if it was a merchant silicon thing, but it doesn't because it's, it's, it's, Amazon's not paying Broadcom margins, right? They're paying lower margins. Um, they're, they're not paying the margins on the HBM. They're paying, you know, they're paying lower margins in general, right? Uh, paying the margins to Marvell on HBM. Um, you know, there's all these different things they do to crush the price down to where their, their Amazon Basics TPU, the Tranium II, right, is very, very cost-effecti…

AI assessment note: “Amazon's chip is the Amazon, I, I call it the Amazon's basics TPU”

Q all that compute in-house Right? To train a model and then to run the model. I suspect if there's an upside surprise, if we went around the table, I'd say this is the place that's getting the least amount of attention and could have the biggest upside surprise. Any, any thoughts, Clark, on what you think is being overlooked or areas that you think are misunderstood about the business today?

A I, I would say, I would say what the last few weeks have proven is that Elon, um, Their team can stand up all this compute. Actually, if you just, you know, went back one and a half years, you know, they were behind in the race to stand up compute. They were, you know, they, they didn't have that many H-One hundreds. They brought in Colossus. Then they brought in Colossus II at a scale much larger than anyone else. And now, you know, as we gear for Vera Rubin, you know, from, you know, a lot of my conversations, it looks like they've, you know, secured maybe up to 20% of It's extraordinary. Especially in the early days of, you know, when, you know, these, these chips are very scarce, that, that they're going to have a, ah, a, a lead on all of this, because, you know, people think that they can stand up this compute better. So I think they'll, you know, what, what the last few weeks have actually shown is that Elon, you know, Elon will take, you know, take a shot at hitting the frontier, but if it, you know, if for whatever reason, Um, they, they have over procured some capacity. This is a very scarce asset that they have shown that they can monetize at actually, you know, best in class margins and payback periods.

AI assessment note: “what the last few weeks have proven is that Elon, um, Their team can stand”

Q healthy operating margins at Azure. So I guess the question is for Microsoft, how do you compete in this world? That is, uh, where people are levering up, taking lower margins. While balancing that profit and, and, and risk. And do you see any of those competitors doing deals that cause you to scratch your head and say, oh, we're just setting ourselves up for another boom and bust cycle?

A I mean, I'd say at some level, the good news for us has been competing even as a hyperscaler every day, you know, there's a lot of competition, right, between us and Amazon and Google on all of these, right? I mean, it's sort of one of those interesting things, which is everything is a commodity, right? Compute, storage. I remember everybody saying, wow, how can there be a margin? Except at scale, nothing is a commodity. Um, and so therefore, yes, so we have to have a cost structure, our supply chain efficiency, Our software efficiencies all have to kind of continue to compound in order to make sure that there's margins, but scale. And to your .1 of the things that I really love about the OpenAI partnership is it's gotten us to scale, right? This is a scale game. When you have the biggest workload there is running on your cloud, that means not only are we going to learn faster on what it means to operate with scale, that means your cost structure is going to come down faster than anything else. And guess what? That'll make us price competitive. And so I feel pretty confident about our ability to, you know, have margins and, and that this is where the portfolio helps. I've always said, you know, you know, I've been forced into giving the Azure numbers, right? Because at some level I've never thought of allocating compute. I mean, my capital allocation is for the cloud from wheth…

AI assessment note: “When you have the biggest workload... that means your cost structure is going to come down”

Q way that I haven't seen in a long time. You've been around a long time. You're very close with President Trump at this stage. Help us understand, like, what is the nature of industry-government relationships? We saw that dinner last week with all the CEOs. You know, you spent a lot of time. Is it unique? Have you seen anything like this in your career over the last 30 years?

A It was, it was hard to go to D.C. in the past, as you know. Getting an appointment is almost impossible. President Trump has a open door to leaders who wants to come in and help them understand the future. This is an administration that believes in growth. Fundamentally, President Trump wants America to grow. If we can grow economically, we will be strong militarily. If we could be, if we could grow economically, we will be secure. I've never met somebody who is secure who's poor. Being, being rich as a nation is an essential part of national security, and he knows that. He also wants America to win the AI, the AI race. This is going to be a very long-term race, and, um, and he understands that this is a pivotal time. He wants the technology industry to run. He wants everybody in the world to be built on American technology. These are sensible, logical things. You know, the opposite is strange to me. If I take everything and I just reversed it, we want our country not to grow. And because we don't want our country to grow, we don't need any energy because we know we need energy to grow. And so let's not have any energy. And in fact, we don't want our technology industry to lead. He understands that our technology industry is our national treasure.

AI assessment note: “President Trump has a open door to leaders who wants to come in”

Q And you went so far as to suggest This might allow them to pass the best proprietary models coming out of the US, um, yet this year, maybe by Q four of this year. So why don't we dig in there? What is your theory of the case? Why is China, you know, doing so well in open source? And should US model companies like OpenAI and Anthropic be concerned?

A Yeah. So, uh, let, let's kind of tie it into, I think three important things that we see happening. The first one being, you know, the Chinese and, and the president addressed this at the AI summit, right? He addressed, um, the, the point around using copyrighted work and he says, you know, he used a great example. If you read a book and you use it, you're not violating the copyright there. And so he addressed that concern. And that was one of the major things that, you know, a lot, a lot of people didn't talk about, but I think it's important for the model makers. And so the Chinese just have been able to work around that because of, you know, their position on IP. And what we're really seeing here, and I think, you know, Bill teed it up even better off of my tweet, which is, um, you know, they're able to compound. So what you're seeing very quickly is both the open source nature, the open weights nature, uh, allow them to basically compound on each other. So instead of working in silos and instead of having to create giant training clusters, Um, separately. They can basically take each other's work, build on top of it, almost consider it like a remix of someone's model. K-II, sort of a well-known remix of what, what Deep Seek had done, and now we're starting to see that happen really fast, and we're seeing two dimensions of it going quickly. One, we're seeing the leading edge…

AI assessment note: “both the open source nature, the open weights nature, allow them to basically compound”

Q so, simultaneous with seeing more IPOs, which is awesome, There, there has been a trend for companies to stay private longer. I think the Collison's used to hint maybe, and now they're more kind of maybe never. And, and some investors in the ecosystem are encouraging that behavior. What, what do you think is different about the people that choose to go out now that the window is quote open?

A I think, um, I mean, they each have different reasons. Some may have, uh, just view from a financing opportunity, the ability to tap the public market, both on the equity and the debt side to be simpler, right? As a public company, I think that's the big piece of it. Second, look, it could be a brand defining event for a company, right? For, for your product, for your employees, giving the level of transparency to your customers, that you're well-funded, that you have a fortress balance sheet, You know, all of that, you can with, withstand the regulatory scrutiny that comes, and even just the, um, the scrutiny from investors, right? That you have the discipline and, and with everything that comes public, people looking at your numbers, so all of those things, right? I happen to believe that all these companies should go public. Um, I also think, by the way, there's, there's a democratic element to it, where I think the wealth creation belongs to the public market. Um, I think you attract different types of investors, not just frankly a public market versus a private market, but also the retail investor. What can you learn from the retail investor, either positive or negative about your business, right?

AI assessment note: “they each have different reasons. Some may have, uh, just view from a financing opportunity”

Q course, all the other people that are going to be needed to, to network and, and, and do the things in the data center. But is there an idea that you guys are investing in an entity that unto itself will have power and maybe grow in, You know, and serve other customers, or is this really just about coordinating the activities of the people who are around the table?

A Yeah, I, I think what I can say, uh, today, Brad, it's, uh, much more of the latter than the former. Uh, you know, could there be opportunity for the former somewhere down the road? You know, potentially, but right now, it's what you just described. And, uh, the operational, uh, control will be, uh, from OpenAI. Uh, so they'll, they'll call the shots relative to all the things relative to the operation. Obviously, there's existing relationships with NVIDIA. There's existing relationships with Oracle. Existing relationships with Microsoft, uh, ourselves, but going forward, uh, they're, they're going to be in a very, very key, uh, key role on the operation, which I think If you kind of go back again to, to, to Sam and, and, and team spending a lot of time and energy over the past 12 to 18 months of seeking for ways to get, ah, opportunity and access to large resources to advance the, ah, training of these large models, it's kind of where he's kind of been with this. So I think that's a, it's, it's, it's not inconsistent with some of the, ah, the actions and behaviors you've seen, ah, over the last, ah, last number of months.

AI assessment note: “it's, uh, much more of the latter than the former.”

Q What are you hearing from the hyperscalers? I mean, they're all out there saying our capex is going up next year. We're building larger clusters. Um, You know, is that in fact happening? Like what's happening out there?

A Yeah, so I think when you look at the streets estimates for capex, they're all far too low. Um, you know, based on a few factors, right? Um, so when we, we, we track every data center in the world and it's, it's, it's insane how much, especially Microsoft and now, uh, Meta and Amazon and, and, and, uh, you know, and many others, right? But those guys specifically are spending on data center capacity. And as that power comes online, which you can track pretty easily if you Look at all of the different, uh, regulatory filings, and use satellite imagery, all these things that we do, you can see that, hey, they're gonna have this much data center capacity, right?

AI assessment note: “when you look at the streets estimates for capex, they're all far too low”

Q everything on their head and they said, well, if we're not doing that anymore, it's way better because we can just move on to inference, which is getting cheaper. And you won't have to spend all this capex. I'm curious. Those are two kind of views of the same coin, but what's your view on, on large LLM model scaling and training costs and where we're headed in the future?

A Yeah, I mean, you know, This, I mean, I'm a big believer in scaling laws, I'll sort of first say, and in fact, if anything, the bet we placed in 2019 was on scaling laws, and I stay on that, right? Which is, in other words, uh, don't bet against scaling laws, but at the same time, uh, let's also be grounded on a couple of different things. One is, um, these exponentials on scaling laws will become harder, uh, just because as the clusters become harder, uh, Everything. I mean, the distributed computing problem of doing large scale training becomes harder. Um, and, and, and so that's kind of one side of it. So there is, uh, but I would just still say, and I'll let the OpenAI folks speak for what they're doing, but they are, you know, continuing to, you know, pre-training, I think is not over. It sort of continues. But the exciting thing, which again, OpenAI has talked to open, I mean, uh, about, and Sam has talked about is what they've done with O. One. Right? So this chain of thought with auto grading and, uh, uh, is just a fantastic. In fact, you know, basically it is test time compute or inference time compute as another scaling law, right? So you have pre-training and then you have effectively this test time sampling that then creates the tokens that can go back into pre-training, creating even more powerful models that then are running on your inference, right? So therefore,…

AI assessment note: “I'm a big believer in scaling laws, I'll sort of first say”

Q with the fact that he said everything in the world today is becoming highly machine learned, right? Almost everything that we do, he said, almost every single application, Word, Excel, PowerPoint, Photoshop, AutoCAD, like it all will run on these modern systems. Sonny, do you buy that? Do you buy that? You know, when people go to replace, you know, compute, they're going to replace it on these modern systems.

A So when I was listening to it, I was buying it, but then when I, he said one thing that kept resonating in my mind, which he said, inference is going to be a billion times larger than training. And if you kind of double click into that, these old systems aren't going to be sufficient enough. Right. If you're going to have that much more demand, that much more workload, which I think we all agree, then how is it that these old systems, which are being decommissioned from training are going to be sufficient? So I think that's where that argument didn't hold, just didn't hold strong enough for me. If that grows as fast as he says, it is as fast as, you know, you guys have seen it in their numbers, then it's going to be a lot more net new inference related, uh, you know, deployments. And there, I don't think. That, that argument holds on the, the transfer from older hardware to newer hardware.

AI assessment note: “I was buying it, but then... that argument didn't hold, just didn't hold strong enough”