Good afternoon, everybody. Welcome to the Goldman Sachs Communacopia + Technology Conference lunch session. My name is Jim Schneider. I am the Semiconductor Analyst here at Goldman Sachs. It is my pleasure to welcome Broadcom and CEO Hock Tan to the stage today. Welcome, Hock. Really thrilled to have you here. Thank you. Thank you for inviting me. Hock, before we dive into talking about your AI business, I want to get some of your broader thoughts on AI and its adoption across the world here. Given your deep view into the enterprise and all aspects, I think it is worth asking how you characterize the state of AI adoption across the average enterprise customer today. That is a very interesting question, Jim, and I guess I will start off with Broadcom as a company. We, like most large enterprises out there today, have been looking closely, reviewing how we can adopt generative AI through use of those frontier models and possibly even those open weight models that are floating around these days. We have been looking at it, and we have been doing that in earnest now for one year. Part of the benefit we have is we have quite good full access to Mythos Preview when it became available, and we look at how does it help our businesses. Where we have been looking at very closely is obviously we are a technology company in how we create products, how we design products, hardware, as well as software. We have engaged 15,000 of our 30,000 engineers in the company in looking at both in hardware and software. A lot of insights, a lot of learnings. The biggest issue is finding use case that generates what you call a return on investment that is clearly meaningful and significant. That's not easy to do, and I suspect a lot of companies out there face the same issue. The second thing is two frontier models. Any model created as applications, it's just a tool. They're not meant to literally replace your people, not at least until you realize it's a tool and how would you apply it. We find out that different users, depending on who they are, could use the tool very differently, very effectively, very efficiently, as well as others who are having time really getting anywhere. It's like us. If I use ChatGPT, and frankly, depending on how I ask, how I prompt ChatGPT, I get different sets of answers, and some more useful than others. Over time, you get better and better at it. There's a limit. What we really find is to find great use cases. My thinking is simply one immediate use case, and that's probably from our perspective, is creating new products. Creating or help us create new technology. What we're trying to ask for is not for it to create it out of thin air as much as create it through our engineers. What we find valuable out of it is they create it with very high quality. The quality is very critical, and that's what those AI tools enable us to do, and on a more accelerated fashion. For a semiconductor chip, that makes a difference because our engineers can still design the chip, but if they have a bug and have to respin, that's six months and $30 million for leading-edge silicon. OpenAI, Anthropic tools, Gemini tools could help us eliminate the need for that respin. So there's a return on investment, so we find that as purposeful. But in terms of just simply productivity improvements, the return on investment is debatable. So it's really about quality, accelerated time to deployment or creation of leading-edge technology. You don't go back and redo what's in the past. The other application, of course, is cybersecurity through Mythos to scan, to basically secure our infrastructure better. Finally, it's something I haven't used yet. We haven't used that, but we can imagine you can use something like a Fable 5 in the minds of critical thinkers in specific functional areas to really replicate the software stack that we're using in specific areas. That definitely is big, and that can generate a big return. But that is something that will require a lot more investment, and we haven't reached that part of the thing yet. The sum total of it is still early stage, but we do find it's important, most of all, the users. The good users, and good users are defined as critical thinkers. More in simple terms, the architects. Senior people who are able to understand the context. Understand the environment of where we are, the workflows, and being able to prompt the models to do it. I can imagine the day when they start using agents, which we are starting to last three months, that becomes even more important. To sum it up, it helps very smart people become smarter. Yeah. That is why I will not say the converse. As the models get more efficient, more powerful, you talked about token cost coming down quite dramatically. What is your personal opinion about where the greatest value is going to accrue across the AI stack, model layer, application layer, someplace else? Well, I am a believer in the fact that intelligence, and the more there is of it, is unique. It is like doing semiconductors. When we design a switch, a network switch, that is the best of its generation at a timely basis. The world beats a path to our door. There is no substitute for the best. Second best does not get anything. So I look at this the same way. For AI, as long as there is continual improvement, I would almost say exponential improvement in the LLMs that keep coming out from the frontier model guys, the value is going to accrue over time to the guys who do the best frontier models. Look at it this way. Take the case of a very good frontier model. 1 GW could generate $30 billion ARR, as you know. The cost to run 1 GW as you all know, is about $10 billion a year. So you have $20 billion that accrue to the model guys and the application that goes with it, and $10 billion fought over by the power guys, the cloud provider guys, even the chip and memory guys. The value comes, I think, in creating the best models, and through it, the best products to monetize the applications. Understand. Okay, now clearly the pace and scale of this spending and infrastructure build-out is exceeding a lot of people's expectations. In your custom silicon business, you've consistently talked about working with a very limited number of large customers who have their own models, can support very large spending programs over time. What do you think your customers are trying to achieve when they work with you? Is it supply diversification, cost optimization, some performance advantage by marrying software with custom silicon? Maybe talk a little bit about or quantify even the advantages they can get by working with you. I think it's really very direct and straightforward. When we design, we develop, I would say, what I call a custom accelerator. We call it XPU, as you all have heard, for a customer. It's really very effective when we actually co-design it, work closely with that particular customer who has to be, in our view, to be a customer, a frontier model developer. Because what happens is we are closely embedded our engineering teams with the team of a customer who help create those frontier models. It starts with the frontier models. They understand what they're trying to create. They understand what breakthroughs they are making for the algorithms to create workloads that run their models much more effectively, more productively, more efficiently. They turn the specs over to us. We work with them to create a chip all the way through to the physical part of the chip that enables the chip, the transistors, to be able to handle the kind of workloads that model does. So that co-design, co-working together, collaboration, tight collaboration really makes the difference. If you do that tight collaboration, you have seen a working example of what happened recently, and it's news out there, with OpenAI. We work closely with the OpenAI team, and within a year, we created an XPU called the Jalapeño, which performs as well, if not better, than the best state-of-the-art general-purpose accelerator out there. Yeah. Fair enough. Now, I want to get to the heart of the forecast you provided at your earnings call last week. You basically laid out the path to doubling your AI revenue for each of the next two years, $115 billion in 2027, $230 billion in 2028. You said those are supply constrained estimates. I want to take a minute to sort of unpack some of that. In terms of the supply constraint or the constraint part of it, you mentioned the ability to deploy physical data centers as one of those, land, power, shell. What is the timeframe that you think is the pinch point for data center availability on the infrastructure side? What is the biggest limiter inside that as you see it? Grid power, gas turbines, construction. Then how would you throw this wild card about political backlash into the equation? Now, what we are seeing now, the positive side to this phenomenon happening is you cannot build up easily an AI data center without a lot of lead time. It gives us pretty good visibility out there, what needs to be done. But equally, that lead time leads to a fair degree of sometimes uncertainty on preciseness of the timing. One of the biggest, before I even say that, now, for creating the chips, and even that we do, whether it is accelerators, AI accelerators, or network chips, connectivity chips. We understand what it takes to the supply chain for chips. We understand the need for wafers, leading-edge wafers, and memory, as well as even substrates. That is pretty well locked in. We can pretty well lock that in for sure in 2027, and the process of locking in 2028 gives us within some limits. That is more within our control. What is interestingly is we find out that is not enough when you want to put together AI data centers. You need power. You need a site. You need a building. You need a facility to house your chips for it to run. That is the other part of the constraint. I am not saying supply chips and memory are not constraints. They are constraints. But at least there are constraints we can work with and more deterministic. What is less deterministic is a power site that is power ready. No, that is power available, but not ready. If you want to make it power ready, say, in 2028, got to start construction on the sites now. That is not us, our customers, has to start construction on the site now to make it available and ready by 2028. That is two years out, almost two years, a year and a half out. A lot of it is not just equipment, whether they are transformer, gas turbines. That obviously are the pieces that go in it. It is also construction. Construction, including local permitting, though some of it exists, but also the rate of construction. You are talking about those of you trying to renovate or build your own houses. You know how construction is, right? That is what at large scale this is happening. There is a lot of other factors that come in that make it, I describe it as perhaps constraint. Because if I turn around and give you guys a forecast, which we do, we want to be sure that is a forecast that is very clear line of sight. That is what translate to the numbers you mentioned. It's a judged number based on not just supply of memory wafers, but availability and readiness of power sites. Fair enough. Okay, in terms of customers, Google is one of your long-standing strategic partners for custom silicon. Thomas Kurian was just up here a couple of hours ago outlining their strategy. Back in April, you disclosed a long-term agreement with Google to supply TPUs and networking products through 2031. Can you maybe talk about how this long-term agreement deepens the partnership between you and Google over that timeframe? Google and us have a very strong relationship. Been doing this for over 10 years, ever since the first version of the TPU. We've been doing every version since then. We will continue to do so under this LTA. Because if anything else, it expands and strengthen the relationship, very technical relationship, strategic relationship we've had with Google now for 10 years. We see that continuing. One of the things that we believe we offer to Google is the strength, the breadth of our technology, especially in semiconductor design. Because with every new generation of GPU or XPU, in our case, we are pushing, our engineers are pushing the limits of semiconductor technology. One example is, as you know, in the past, 10 years ago, you get better performance out of a CPU in chips by moving to a next generation process node. Each time you move to the new generation process node every 18 months or so, you double the transistor count, you double performance. Well, that's Moore's law. Now, Moore's law has come to an end. Moving down process node, you cannot achieve double transistor performance anymore. You barely eke out 5% improvement, but you get power. So that's one positive thing. But building the kind of accelerators we are building today, and these are monster multipliers, very complex chips. What we are doing with our customers is we're trying to, what I call, substitute Moore's law with something else. The simple way to look at it is Moore's law is one semiconductor chip is limited by the radical law of physics on optics, 800 sq mm. That's the largest size die you can make. So we can only cram so many transistors into one. The transistors are obviously mostly multipliers used in AI accelerator. What we are doing now is we are putting two die together in one chip. So effectively, you now have 1,600 sq mm, double. That is a way of getting the double the performance. The current generation we are working on has four dies. You can just imagine. The one after that probably will go to maybe no limit, eight die. Eight die in one chip. So your chip size is equivalent of, say, 800 times 8, over 6,000 sq mm running as one single chip. The amount of technology, the complexity of technology needed to be created to enable that to run, it is technology that is not just created at a wafer fab or with the memory guys. It is us designing it. It is very challenging, very interesting, and that is the relationship we find very exciting with every one of our customers. Because we are able to keep enabling them to create more and more high-performance chip, which enables them to drive as a key part of their stack much higher, better, exponentially better and better LLMs, which drives towards the, what I said before, more and more value out of super intelligence. My view of it is with the ability to create those kind of chips and the imagination and intelligence of the R&D engineers, the data scientists writing those algorithms, and finding breakthroughs towards super intelligence, we do not see a limit at all yet towards the exponential improvement of those frontier models. Yeah. As you know, there has been some investor concern about the potential for some of your customers, including Google, to insource more elements of their custom silicon strategy. Maybe help us understand what are some of the deep technical and/or commercial moats that you have that gives you confidence that that custom silicon business will grow with that customer? Well, I have expressed some of that and it continuing to grow. The example of it is literally we have more IP out there in semiconductors than probably most of our peers out there across the board, whether it is in I/Os through very high speed interconnects, SerDes, whether it is in creating much more density of multipliers within every square millimeter, as well as putting multiple die in one chip. You have to have dies communicate, chip communicate very well, very fast with each other, and being able to do it very low power, on and on. Then advanced packaging, how do you package a die in one chip? All these are, I consider, deep moats that our peers have to come forward with to develop it, to challenge us. We have the advantage of being earlier than most other guys. And most of all, it's no different than what we have been going through as a semiconductor company for the last 30 years, where we have picked different areas, whether it's switching, whether it's networking from a different point of view, Ethernet and computing. We have been able to essentially out-engineer and compete against our competitors. It's no different here, whether it's another peer semiconductor company or more interestingly enough, even our customer themselves. Yeah. Okay. You mentioned OpenAI before, so maybe mention them again. You previously signed a supply agreement before 10 GW of custom silicon capacity last October. Then at the end of June, as you referenced, you announced you delivered your very first ASIC sample called Jalapeño within nine months, I believe, which I think is pretty fast. So maybe discuss the partnership with OpenAI and what enabled you to deliver on that accelerated timeline. Well, the biggest thing is like what we do with our customers who are frontier model developers. They have a strong team in developing the models. They also have a team that translates what the model wants to specifications, to architecture of what the compute hardware needs. And we provide that creation of that chip with architecture from the customer side as well as from our side to create a micro architectures and the physical design. When you have that very tight, almost seamless flow, it's not that much a challenge to be able to create chips of that caliber that quickly. So the key is the tight engagement with the customer. So needless to say, relationship is very strong and continues to be very strong because we are working on the next and two generations to the Jalapeño. Yeah. Okay. Then in June, you also announced a special purpose vehicle in collaboration with Apollo and Blackstone, to provide initial $35 billion in financing for over 20 GW of compute capacity to various AI labs, starting with 1 GW for Anthropic this year and I think 5 GW in 2027. Maybe help us understand the parameters behind that agreement, and how broad it is. What's enabling Anthropic to sort of expand their compute capacity and maybe how broad could this agreement get? Then it comes back to a vision of this, of what we see in generative AI. One we look at is we have only six customers today at this point. Very concentrated group, but I believe these six customers, all each of them are creating their own frontier models, LLMs, and they each continue to invest to try to create a state-of-the-art leading edge. That is what makes this all very interesting because that is the creation of this phenomenon of super intelligent. Because from the models comes the products, applications as you call them, from which you monetize and from which enterprise consumers benefit. With this focus in mind, we also look at it very practically. We want to support all these six guys because they do the best, and we cannot tell who is going to win. By the way, they may coexist in where they land in their models as it is a continual state of evolution, I would almost say revolution in the way they exponentially improve. Of the six, two of them, two of the strongest guys do not have the cash flows to be able to generate the compute needed to train and create inference, productize their models. We create the chips for them. We step forward to enable them to have the compute needed to support their models. We do that by not just taking cash from our books over there. We do it by creating a platform, we call it XPV, where it would enable them to be financed. We do not finance directly. I will be direct. This is not circular financing. When a customer sees the demand that drives up their own demand and they do not have to use us necessary, they can get it somebody else, that is not circular finance. We are using financing to create demand. Demand is there. What we are doing here is we harness a lot of financial, bunch of financial partners and banks who are willing to come in and essentially take the risk on financing Anthropic and OpenAI, at least partially. On the balance of it, which is what we did with Apollo and what we will do on the XPV going forward, we will provide backstop in the form of residual guarantees on what is the equipment that is secured against that. There is combination of financial partners who will come in. Why they come in? Big part of it is they believe in the demand, but two is rates, right? They get better rates grading through this, and our partners, OpenAI and Anthropic, have the capability to handle that as they grow, as they generate more and more revenues to be able to take on this additional enabling this financing capacity. Yeah. Okay. Last question on this topic. We've seen a lot of rapid developments in China, specifically around open source, open weight models and a lot of activity from players in the region, whether it be Moonshot, DeepSeek or others. How do you see the role of open source or open weight AI models in the market over time? From where you stand, do you think any of those customers could actually become customers of Broadcom in the future? That's an interesting question. You're going to set me off, by the way about this. As I said in all this question you asked, AI chip believes in reaching generative AI, in reaching super intelligence. We're not there yet. We are not at AGI yet, but we can get there. The frontier models, the guys who do frontier models are pushing us towards doing that. As I said too, the better the models, the better appropriate use case, the better results outcomes you get out of it. But equally, to get to those models and get to those outcomes, those are just tools. You need to train up and you need to have users that know how to do it. Obviously I'm a believer in closed frontier models. That's not to say open weight models, as you put it, can't coexist. They could, b ut the value chain, I think over time, over the next five years, I believe, will reside with the leading-edge frontier models as opposed to the open weight models. Even though they may coexist, because I hear all the time as I go sell products to enterprises, how they don't like the fact token cost is getting expensive. When they start using it, and using it for very complex application, that maybe open source will give them a cheaper source. But my thinking and my experience as a technology company is, no, the best is what matters at the end, and value accrues to the best. I give you an example of the math, the economics we see today. Just today. Now, I agree things might change a year, two years down the road, but things have changed a lot since then between open weight models and closed models, roughly. You can get these sources of data from OpenRouter, Vercel. The total amount of compute tokens consumed and generated, the cost of generating those tokens to be consumed globally, and this data is obviously for inference. Training is not easily monitored, so it's not that, just inference. It's around $200 billion a year now. Give or take, round numbers. That's cost of generating those tokens, producing those tokens and infrastructure, $200 billion. Ask yourselves, what's the revenue you could attribute being created and earned by those model guys? I give you that. You can count and look at where the models are easier. Around $150 billion. Interesting, isn't it? This generative AI today, I know it's early stage, things are still growing. But we're generating $200 billion of token cost to generate $150 billion. But look below the surface, and you can split it up. On that tokens generated, half of it is generated through a frontier model, closed systems. The other half, roughly, in fact, more than half, generated through players offering open source, open weight models. That's about it. You look at the revenue, 75%, at least, is coming from frontier models. That's $120 billion. They spend roughly $100 billion. Not so bad. It's still growing. Let's look at open weight models. For spending $100 billion, the other $100 billion, they generate revenue $30 billion. You think that's a sustainable model? We don't know. That proves one thing. The value goes to where intelligence continue to improve. Fair enough. Just one minute left, but I wanted to ask you about capital allocation for a moment. From your financial forecast, you're going to generate a tremendous amount of free cash flow over the next couple of years, given the numbers you've laid out. Can you maybe update us on the board thinking around capital allocation in light of all that? Well, give me three months. We usually do the capital allocation annually in the December board meeting. Your question is very right on, in a sense. Because I can imagine, given our forecast, and we already start seeing that in 2026. Because we'll end 2026 with a record amount of cash sitting on our books, end of this quarter, end of October. Just simply because our revenue is growing so fast, driven a lot by AI. Go to next year, we're talking as our guidance forecast, $115 million for AI revenue alone. You add on it our non-AI revenue and software, you're talking about pretty high dollar number, which 45, mid-40s free cash flow. So we're going to generate a lot of cash. Your point exactly. More so in 2028. The question we have to look at as management and the board is, one, we could give it out as increased dividends. But how much of that will be increased dividends? Still a long way off. The other likelihood is we'll do a stock buyback. These are probably the only two choices we have. To probably bring down debt makes sense, but our total debt of about $56 billion was taken on when interest rates were much lower. So it's almost like, why would I want to prepay low cost interest or debt? Which really leaves two choice of capital allocation, increase dividend and probably do stock buyback. Very good. I think with that we're out of time. Thank you very much, Hock, for being with us. We really appreciate it. Thank you.