Why Broadcom Stock Is Worth $650

Key Takeaways

• Broadcom’s AVGO key business lines.
• XPUs, GPUs, TPUs: What’s the difference?
• Where the relationship with Alphabet GOOGL may be headed—and what it could mean for Broadcom’s future.
• Who Broadcom’s competitors are today and what threat they pose.
• The potential impact of artificial intelligence regulation.
• Morningstar’s forecasts for Broadcom’s revenue, earnings, and profit margins.

In this bonus episode of The Morning Filter podcast, co-host Dave Sekera sits down with senior analyst William Kerwin to talk about Broadcom. They unpack the company’s three key business buckets—XPUs, networking chips, and software—and break down the company’s relationships with major customers like Alphabet, Anthropic, OpenAI, and Meta Platforms META. They also discuss the mix shift of AI spending over time, what that shift will mean for Broadcom’s business, and how Broadcom stacks up against its competitors.

Tune in to hear Morningstar’s forecasts for Broadcom’s revenue, earnings, cash flow, and margins, why we think Broadcom will maintain its wide Morningstar Economic Moat Rating despite the changing AI landscape, key assumptions behind our $650 fair value estimate, and why we believe the stock is underappreciated in the market.

Have an idea for a bonus episode? Send it to themorningfilter@morningstar.com.

Transcript

David Sekera: Hello, I’m Dave Sekera, Morningstar’s chief US market strategist and the co-host of the weekly podcast, The Morning Filter. Today, I’m joined by Will Kerwin, senior equity analyst on our technology team who covers semiconductors and hardware stocks. Today is Thursday, Sept. 17, and we’re conducting a bonus episode of our podcast to conduct a deep dive into our analysis and valuation of Broadcom.

Will, thank you very much for joining us today. I’m really looking forward to this. Although before we get into talking about Broadcom AVGO, I actually have to commend you on a couple of your prior calls, a number of them that we’ve actually highlighted on The Morning Filter in the past.

For example, I think it was early 2025. We talked a lot about Marvell MRVL and your call there. I mean, I think that was probably one of the most differentiated calls we had on the technology team at that point in time. Probably one of our most differentiated calls versus the market, really across our entire coverage. Certainly has worked out well. I don’t know how much that stock is up. Since then it’s up well over 200%, and I think it’s even still undervalued today.

Kerwin: Yeah. Well, thank you. It’s great to be here, and certainly Marvell has been a rocket ship since last year.

Sekera: Yeah. Well, and then of course, the past couple of months we’ve also been talking about in the hardware space, a lot of these commodity-oriented technology hardware companies were up a couple hundred percent, maybe even a thousand percent in some cases in the first half of the year. You really stuck to your guns on those, talked about why we thought that they’re overvalued. They’re getting well into 1-star territory, some of the most overvalued stocks, and you weren’t willing to fold; you kept our fair values where they were even though the stocks kept going up. And really, since June, a lot of those stocks have fallen 30%, 40%. I think some of them are even down 50%. And even from there, it still looks like they have further to fall.

Again, I really like to see that balance. It’s not that you’re always a bear on everything and you’re always not a bull on everything. It’s really nice seeing that blend and really being able to help investors that way. But let’s get into Broadcom.

Kerwin: Absolutely.

Broadcom’s AI, Networking, and Software Businesses

Sekera: Just to start, I think Broadcom is one of those names that five, six, seven years ago was not necessarily a household name. A lot of people maybe don’t know exactly what the company does. It’s an amalgamation of a couple of different businesses that have merged together. Can you give us that 30,000-foot view to start off with? Who are they? What do they do? Maybe talk a little bit about some of the more important product lines.

Kerwin: Yeah, absolutely. Honestly, what they do today is going to look vastly different than what the mix was five years ago, before they were a trillion-dollar market cap, before they were a household name. I think the most important business line here is the AI chip business.

Now, they are a competitor to Nvidia NVDA and Nvidia’s GPUs, but they are not exactly like-for-like. What we call what Broadcom does are XPUs. Effectively, you can think of these as a customized version of a GPU. The purpose is the same, whether that be for AI training or AI inference, but rather than being what we think of as a generalized solution for that—which Nvidia does extremely well; they can sell them to any customer in the world—Broadcom helps co-design chips for AI processing tailored to a specific customer, to a specific model, to a specific application.

Think of a large application like Google search algorithms or recommendation algorithms for Meta Platforms META. They can have enough use cases for those that it becomes economical to invest in their own customized chip. Effectively, what you get—and I’m sure we’ll dive into this later on today—you can get co-investment from the customer there, and then, in turn, they get a lower marginal cost per chip for those applications. That’s kind of the first big bucket.

What Broadcom used to be more concentrated in, that is now a little bit smaller in the face of the growth of those XPUs, is networking chips. So, networking being the connectivity between servers in a data center, connectivity for wireless or wired connections in an enterprise, in a campus environment, what have you; they make the chips that really make all of that networking gear run. They’re by far the best of it. They do it very well. It’s a large business, but a little bit smaller than that XPU business.

And then outside of chips, they also have a pretty sizable software business as well. Now they had software before this, but it greatly expanded with the acquisition of VMware back in 2023. Most of that business now is for what we call virtualization software, which, at a high level, effectively helps servers and storage in a data center run smoothly together. Those are the three big buckets I would say: AI chips, non-AI chips, and then software.

Sekera: Just roughly speaking, if I’m thinking about their total revenue, what percentage of revenue comes from each of those major business lines?

Kerwin: Just in the past couple of quarters, the AI chip portion has exceeded half of total revenue. We see that growing from here, being by far the biggest driver. You can think of the non-AI chips being less than 20%, and then the balance, call it about 30% for software.

Why XPUs Are Becoming More Important in AI

Sekera: OK. And now we’ve talked about XPUs, and just to try and understand what XPUs exactly are and how they kind of fit in the chain. You have Nvidia’s GPUs, we’ve got XPUs, Google has their TPUs, you’ve got the old-fashioned CPUs. Can you maybe just walk me through: What are each of these different products used for? And you kind of mentioned about being a competitor but not a direct competitor. I mean, can you use XPUs without GPUs, or XPUs used on top of GPUs in collaboration with them? I mean, just explain that entire process of the compute for AI.

Kerwin: Sure. In short, it’s all of the above. Predominantly, all of their XPU customers will also be buying GPUs from the likes of Nvidia. So, it’s an “and,” not an “or” type situation. Now you mentioned the Google TPU. That is actually, in our view, a version of an XPU, and Broadcom is the co-design partner with Google on the TPU. When we think about XPUs versus GPUs, TPUs fit into that XPU portion of the bucket, even though there are some specifications that can make it its own third thing that we don’t need to get into right now.

But effectively, you will choose an XPU when you have a large-scale application because it takes billions of dollars of co-investment from the likes of the customer. So, Google for the TPU; Broadcom helps design for Meta’s custom chips as well. Those customers have to invest billions of dollars, and yes, they get a lower marginal cost per chip. These chips are still not cheap. They’re still expensive. You need the justification to invest in an XPU. It’s being able to buy such a significant volume that those cost savings on a per-chip basis help to offset that co-investment that you put in upfront.

Right now, we see these reserved for the largest of large-scale customers and, really specifically, within those customers, these large-scale applications. Think of a large language model running inference like a ChatGPT, a Google Gemini. Those are obvious use cases for an XPU. A smaller enterprise running their own inference probably doesn’t make as much sense given that level of investment. So, predominantly large customers and predominantly large applications within those customers. And I would just add: predominantly for inference rather than for training. We typically see GPUs, those generalized versions, as more flexible, which benefits training.

Sekera: Actually, I’d like to get into that a little bit more and try and understand, maybe explain, what is the difference between training and inference? And I guess even taking a step back from there, thinking about what the use case has been for GPUs, XPUs, TPUs for the past couple of years, and where do we see that going forward?

Kerwin: Absolutely. Training, at a high level, is building out a new model. When you see, whether you’re using Claude or whether you’re using ChatGPT, you usually have a choice of what type of model: GPT-4, GPT-5. Now they have GPT-6 Astra from OpenAI. To train those models, you’re essentially running a huge amount of compute processing across tens, if not hundreds of thousands of GPUs, typically. And effectively, what that creates are what we call model weights.

And you can think of these as kind of probabilities for the next stream of text that’s going to come out of that chatbot. Now, inference is once you have the trained model, it’s ready to go, it’s ready to be used. Anytime I type a query into Claude or into ChatGPT, and I get a response back, that’s an inference response. Training the model kind of happens in the background. When it gets released to the public and you use it, that’s inference.

GPUs are more flexible; they’re more expensive than XPUs, but they’re better for training because there’s a lot of malleability required in training effectively. They can shift what they’re working on during training more effectively than an XPU can. Whereas once you have those model weights, as we call them, an XPU can run very, very efficiently knowing those model weights and running inference. And so, in short, what you get with an XPU is you get a lower marginal cost, you typically get a greater amount of power efficiency, and we see power becoming more and more of a constraint in building out AI architectures. And then you also get generally higher performance because it gets tailored to a specific software language for a customer or even a specific model.

Sekera: I’m just trying to think through how would you expect that shift over time between how much compute power is being used for training versus how much is being used for inference? Sounds like the XPUs are much more weighted toward a mix shift toward inference.

Is that something that we’re looking at going forward, or is there just still so much training that needs to be done that there’s still going to be a lot of use case for the GPU, and the XPU also has that use case for the inference? I don’t know. I’m just trying to understand. When I’m thinking about the broader picture of how much compute power there is today, how much is coming online in the next couple of years, how much will be coming on by 2030, how is that shift going to occur from those training to the inference models?

Kerwin: We think it’s going to shift toward inference, but we talk about this happening over a long period of time. Even still today, as much as inference has been getting a little bit more press in the media, training is still the vast majority of the spending that’s going on. And it’s concentrated in a handful or two of extremely large customers: OpenAI, Anthropic, Google, Meta, et cetera. But we see inference growing in that mix over time.

Now, that doesn’t mean the GPUs are going to go away. When I talk about XPUs being good for inference, but really only for those largest of large-scale customers, if you think about inference eventually coming down to a smaller scale in an enterprise—so, a business that wants to take a smaller version of a model and incorporate it into their own systems—they’ll probably use GPUs for that rather than XPUs.

But a large-scale model training from one of those large customers that I named probably uses GPUs for training, but then they will use XPUs for their large-scale, like a chatbot or some other type of large-scale software with a lot of use cases. We see the mix going toward inference, but we think it’ll take some time. We also actually expect a mix shift of more XPUs gaining share over GPUs.

Now, this doesn’t mean that we think XPUs are going to eat Nvidia’s lunch, take over the entirety of the market. If you look at 2025, by our estimates, it was about a 10% split of total AI compute spending toward XPUs, with 90% GPUs, mostly from Nvidia. We forecast that to rise up to about 25% or maybe even 30% in the next five years. And we think that’s a pretty solid steady state for it to sit at. We still expect GPUs to be the vast majority of the market and the right option for most customers, but we think that more of these large customers will adopt higher volumes of XPUs to save money, to save energy, and to get better performance.

The Future of Broadcom’s Relationship With Google

Sekera: OK. Well, maybe we can shift gears here a little bit. You talked a little bit about who some of the largest customers are. My understanding from the research I’ve read from you is that there’s a little bit of concern right now with some multisourcing, that Google has now been utilizing a couple of other manufacturers to produce some of their chips, but they’re still using Broadcom for the vast majority. So, maybe a little history between Google and Broadcom, what that relationship has been, where you see it today, and maybe where you expect it’s going over the next couple of years.

Kerwin: Yes. Google is Broadcom’s anchor customer for its XPU business, and the TPU was actually the first-ever XPU. And they’ve been working with Google on this for 10 years. Before AI was a household term, before any of us had heard of ChatGPT, Google was using the TPU for internal workloads. If you think of Google Search, that really is kind of the initial AI use case in the algorithm to provide good ads, to provide good search results. The TPU has been used internally at Google for some time, and that’s where Broadcom really built up this expertise to then apply it to some of these other customers.

Google has been the majority, and I would say the vast majority, of Broadcom’s XPU revenues up until this point, but we see that shifting a little bit into the future. There’s a combination of Google adopting some secondary sources. We still think Broadcom will be the primary source for the TPU, but also Broadcom, for its part, is expanding its own customer base; going from extreme concentration with Google to adopting new large customers like Meta, OpenAI, and Anthropic.

How Broadcom Stacks Up Against Competitors

Sekera: Got it. OK. When I’m thinking about who the competitors are here to Broadcom, how do you see that competitive landscape over the next couple of years? Is there really anyone out there that can compete with Broadcom today? Are there other people that you’re concerned about coming in over the next couple of years and being able to do the same sort of design on these types of products?

Kerwin: It is a field with competition, but Broadcom, from a scale and from a technology perspective, we think clears the rest of the pack. The two that I would name are Marvell, a name that you referenced at the top of the show. They have a small business, but they have designs with Microsoft MSFT and with Amazon Web Services, or AWS. And then MediaTek as well, which is kind of the one in focus for Broadcom investors because they are the secondary source now in Google for the TPU. Now, I’m sure we’ll get to this a little bit later on. There have been some concerns that maybe MediaTek will eat Broadcom’s lunch and take over its primary source. We don’t see that to be the case, but those would be the big three. There are some smaller kind of design consultancy shops out of Taiwan, but really it’s Broadcom at the top and then Marvell and MediaTek.

Sekera: OK. Just thinking about technology in general, it’s certainly one of the fastest adapting fields out there. How can we be as confident as we can be that you don’t have any kind of new technology, new algorithms, shifts in software that you wouldn’t need this type of XPU product really to be able to run that inference going forward?

Kerwin: It’s certainly a fast-evolving environment, but actually what we are seeing is that the shift is moving toward more XPUs rather than less. And so what we’re seeing is more heterogeneity in computing, and effectively you could interpret that as being good for the flexible GPUs, and that can be the case. But what we see is these large customers, that are kind of the sweet spot for an XPU provider, are building out more large use cases that could portend maybe they don’t just have one XPU, but maybe they have a couple different XPUs for multiple different use cases.

We actually see the mix shift moving toward XPUs, again, not reaching a majority of the market and it’s far from our expectation, but what we see is this heterogeneity and more of these massive scale applications actually being beneficial for the mix shift toward XPUs.

Why Morningstar Thinks Broadcom’s Moat Is Durable

Sekera: Which kind of leads to the next question. Of course, we wouldn’t be Morningstar without talking about the

. In this case, our moat committee has awarded the company with a wide economic moat, meaning that we think they have long-term durable competitive advantages, so they’ll be able to outearn their cost of capital for at least the next 20 years. Again, in a field like technology where things move so fast, what’s really the primary moat source here, and how do you think about that over that long of a time period?

Kerwin: Well, with chips like this at kind of the cutting edge, we think predominantly about intangible assets. And really this breaks down to the design prowess that Broadcom has in designing these XPUs and its networking chips for that part. I mean, I may have glossed over the networking chips a little bit at the start, but those have a big place in AI as well. Broadcom’s AI chip business, XPUs, are by far the biggest portion, but these networking chips, connecting XPUs and GPUs together, also vital and also important.

The design prowess there is really what makes the moat for Broadcom, and we would call it by far and away the primary moat source. And really, this comes down to the best performance, whether that’s in an XPU or a networking chip, being first to market for the next generations of these, so, for the next fastest XPU, the next speed of networking that you achieve. What we’ve actually seen is Broadcom supports this with a massive research and development budget. Again, its scale dwarfs those of its competitors, so it’s able to outspend and still be extremely profitable. And what we’ve seen in networking, at least, is actually that its technological lead has expanded over time. We look at that as: When you get to a new speed of networking, who is first to market? And Broadcom has had a lead for some time, and we’ve actually seen that expand in terms of when its competitors bring out a new product.

So, intangible assets are the primary source; really, the design of these chips is extremely hard to replicate. They’re at the cutting edge, really second only to Nvidia in our view. And then I would add on the software business, nothing to scoff at in terms of scale, and 30% of overall revenue. We see pretty significant switching costs there too. We awarded a moat to VMware preacquisition by Broadcom, and really this is just that designing a data center around this software stack, it becomes incredibly difficult to rip out. We think intangible assets are primary, but for that software business, also switching costs.

Why the AI Buildout May Have Further To Run

Sekera: OK. I’m going to want to talk about your forecast specifically for Broadcom in just a minute, but before we even get there, I think one of the hardest things maybe to conceptualize for investors today in artificial intelligence is trying to understand—we know how much compute power there is out there today. There’s pretty good visibility as far as the data centers that are being built, how many gigawatts of compute power are probably coming online over the next year in 2027 and, to some degree, even into 2028. But again, maybe walk us through Morningstar’s technology team’s kind of view on how much compute power really needs to be built out over the next couple of years, and how is that compute power actually going to end up being utilized at the end of the day?

Kerwin: Absolutely. And really, when you think about it in terms of a company like Broadcom, we think about the capital spending dollars by the companies that are building out this compute power. Now, at a very high level, directionally, what we see is an acceleration occurring that we think will continue through 2028. There’s a massive amount of capex spending this year, put it around $750 billion by a select group of five or six companies leading the charge. And we think that’s going pretty materially above $1 trillion in 2027. Another year of enormous growth.

We think you’re going to see another year of enormous growth again in 2028. And that’s informed by a few factors. That’s informed by the orders that are coming in for players across the supply chain. And really we think it’s fueled by supply constraints that exist today, where the buildout would be occurring faster if there were more chips available, more networking equipment available, more power available, if it was faster to build a data center. We see this almost parabolic trajectory through 2028. After that, we think growth can continue, but we forecast a pretty material deceleration from these levels. I mean, for Broadcom, we’ll get into their forecast, but enormous growth over the next two years, and we see that tapering pretty significantly down in 2029 and into the end of the decade.

Now, we think that growth will continue. There’s certainly some fears around what if there’s a correction in AI spending, a decline in one year. For us, we look at a few factors. Those supply constraints that are keeping things more or less rational in our view, and then the monetization of AI. We think that’s kind of the end all be all metrics. What we look for is: Are software companies able to use this to expand their own addressable markets, to expand their own profitability? We look at public cloud businesses from AWS, Microsoft, and Google.

We see their revenue rates accelerating, actually, as they start to integrate more and more AI. And then we have only some breadcrumbs here, but you look at the monetization of a company like Anthropic that’s preparing to go public, and the stories that have come out so far before releasing their S-1 make it appear like the path to pretty strong profitability is underway at the very least, if not, we’re there yet. Those are kind of the things that give us confidence that this is a durable growth trajectory and not a flash in the pan.

Could AI Regulation Slow AI Spending?

Sekera: Well, if we’re on the topic of Anthropic, OK, we have to also talk about, in the past week or so, there’s some commentary that’s come out of Anthropic about potentially slowing the pace of development of some of the AI models out there. It sounds like some of the other platforms might be willing to go along with that as well, that they need to bring some sort of regulatory body, maybe in-house, in order to monitor the development and the use cases here. What’s kind of Morningstar’s thoughts, as far as one, is that something that potentially we think might happen? And two, if so, is that something that can slow down the development and the buildout from the AI buildout boom?

Kerwin: Well, first off, I’ll say that I think just this week, Hock Tan, the CEO of Broadcom, affirmed Broadcom’s guidance through 2028 in response to this letter. Clearly, management doesn’t see the pace of the buildout slowing down as a result of this. Now, I think we tend to agree that this is a little bit more bluster than substance, and it appears to be self-serving in my view, in that they’re trying to get a favorable regulatory environment even before there’s any sense of material or regulation involved. And if you think about it from the point of view of Anthropic or Sam Altman’s OpenAI, who agreed with the letter pretty immediately after it was published, these are the top two players in the US AI scene right now.

And so there can also be perhaps a cynical view, but perhaps a realistic view that these two companies want to ensure that they’re durably on top amid potential competition coming in, whether that’s in the US or from China, and maybe regulation is the way to slow down would-be competitors and maintain their own position. In short, it’s something to watch certainly, but we don’t see this slowing the buildout at least over the next two to three years.

Broadcom Forecast: Revenue, Earnings, and Fair Value

Sekera: All right. Well, you hit on the topic of them giving guidance for 2028. I think maybe just to start off with talking about your own forecasts, how do you forecast a company like this, especially with multiple different business lines, thinking about—we’re in fiscal 2027 with the company today—what their fiscal 2028 guidance is. Maybe walk us through short term. How do you model out the top line for those individual businesses this year, next year? And then maybe the follow-up to that is going to be, well, then how do we even think about it three to five years out?

Kerwin: Yes. The guidance was to double their AI revenue in fiscal 2027 to about $120 billion and then double it again in fiscal 2028 to the tune of $240 billion. We model roughly in line with that. We actually see upside to that. We thought it was important that management noted that they have supply guaranteed for those levels of demand, and this is an era where many revenue guidances across the AI infrastructure supply chain have been supply constrained. That’s really the key lever to look for.

But I think the function for Broadcom is kind of like three variables. We look at the growth rate of overall AI spending, and the capex of this leading group of six customers is a really good proxy for that. Then we think about the share shift within those customers—XPU, GPU. And again, I talked about our expectations that XPUs are going to outgrow, outperform GPUs over time and take some share.

And then what does Broadcom’s share look like within the XPU paradigm? We think there will be some multisourcing here. So, actually, from a near 100% share of XPUs two or three years ago, we see Broadcom’s share actually going down, but the broader XPU share gains lifting up the company, and we still expect Broadcom to actually outperform the growth of even Nvidia over the next five years.

But I talked about the tapering, right? So, doubling for the next two years, which is almost common at this point in the AI buildout, but I don’t think we should forget that that is incredible guidance. After that, we have a pretty strong tapering down toward 25% growth in 2029 and then dipping into the midteens in the ensuing next two years for Broadcom. I think it’s important to note that we don’t think the pace that we’re on currently is durable over a five-year period, but we think that it’s durable over the next two to three years and that we think still we can grow from there.

Sekera: OK. The next topic, of course, is going to be operating margins. Now, you’ve talked a bit about how there’s going to be the mix shift within the company, that you have such huge growth in the XPU business will dwarf some of their other business lines. How do you think about operating margins as far as that mix shift? How do you incorporate that into your model? What kind of margins do you see? You also talked earlier about how they share some of the development costs with their own customers. How does that flow through the income statement?

Kerwin: Broadcom is already at pretty excellent operating margins. They’re mid-60% range on a non-GAAP level for operating margin, which is frankly superb for a hardware company, for a chip company.

We don’t see actually a ton of expansion from here. I think this is an important note. When we think about XPUs, yes, there’s this co-development expense, but I talked about those coming at a lower marginal cost to the customers. Part of that is because Broadcom takes a lower gross margin on it than Nvidia takes on their GPUs. There’s this interesting dynamic where these XPUs are actually gross margin dilutive, but operating margin neutral to accretive. And the way that happens is they get this contra operating expense, which comes in the form of the co-investment from customers. We actually forecast gross margins to go down as the mix toward XPUs increases in our model, but we have operating margins holding roughly flat, actually expanding a little bit as they scale with that contra opex from these customers.

We think the profitability of Broadcom, again, best in class for a chip provider, is extremely great, and it leads to free cash flow, which I’m sure we’ll talk about as well as that undergirds our model. But in short, not a ton of operating leverage even on this massive growth because of some of those margin dynamics of the mix shift to XPUs, but very attractive margins nonetheless.

Sekera: All right. When I’m thinking about earnings and earnings growth here, it sounds like that earnings growth is going to be pretty much in line with that operating margin growth. What are you looking for earnings this year and next year? What kind of market multiples is the stock trading at versus what your long-term intrinsic valuation is?

Kerwin: Yeah, exactly. We forecast pretty close to 50% compound annual growth for earnings through the next five years, which, again, the numbers that we get used to for a stock like this are simply astonishing. And I think the multiple is a really important question here because the market is trading well below our $650 fair value estimate, and we don’t think our valuation is that crazy if you believe in the growth that’s to come through 2028 that management is guided for.

If you look at our 2028 model for earnings, it implies, at our valuation, the stock being at about a 25 times earnings multiple, which we think is fair given we still expect double-digit growth thereafter, and it’s, excuse me, an extremely high-quality business. The earnings growth, yes, moving in line with operating income growth, we don’t see a huge amount of operating leverage over the already really pristine levels, and we think it can grow into this valuation in that the multiple we imply today is not that unreasonable when you think of those AI sales quadrupling over the next two years.

Sekera: And when you say grow into that multiple, you mean just growing into a 2028 multiple?

Kerwin: Yes.

Sekera: We’re not even talking all the way out to a 2030.

Kerwin: No. Yeah, just to 2028. Yeah.

Sekera: What’s the market pricing in, or what is the market missing? If I look at the multiple it’s trading at on our 2027 earnings estimate, that seems somewhat reasonable, but then there’s such huge growth in 2028 that the multiple looks really attractive at that level. Is it just a matter of the market’s just not giving them credit for any growth past 2027? And if so, when would you expect that the market would be able to get enough confidence in that growth in 2028 to start rerating that stock higher?

Kerwin: I think there are two key debates about the stock right now. One you’ve already referenced, which is this multisourcing dynamic in Google. We’ve seen a lot of reports, and I would say a lot of rumors around what the share shift is going to look like now that MediaTek has qualified as a secondary supplier to Google.

Even as recently as two months ago, there were reports that Google was kind of going to split into these two different XPU lines. One would be extremely high volume and high complexity. One would be lower volume, and that MediaTek won the higher-volume one, and that Broadcom was going to go from almost a dominant share in Google to a minority share. That has not happened. Broadcom affirmed, with Google’s approval, that they’re in the high-volume, high-complexity chip on their last earnings call. We think that’s been debunked, but there’s still some pessimism in the market that MediaTek is going to eat away at the share there and Broadcom will be worse off for it.

I think the other one, and maybe the more important one at this point, is fears about the durability of AI growth. I think the market has adjusted to expecting pretty high growth in 2027, and we don’t think our capex expectations above a trillion are really anywhere outside of consensus at this point. But I think when you think about 2028 and an acceleration off of 2027, which is already kind of mind-blowingly high, the doubling again from Broadcom, I think the market is pricing in a good amount of skepticism there. And then as you get even longer term, which we like to think about at Morningstar, I think the market is kind of pricing in a significant amount of uncertainty that this AI growth can endure at these levels and not go down after ’27 or after ’28. And so, we talked about our levels of confidence, the improving monetization of AI models, the supply constraints that are bringing some rationality into this buildout. It gives us confidence that through 2030, we think growth can continue, but I think the market is pricing in a significant amount of doubt into that confidence.

Sekera: When would you think that the market would start getting that kind of clarity? When you think about what Broadcom does, the suppliers that they have to work with—whether it’s like Taiwan Semi TSM in order to be able to get the type of machinery that they need there to build these types of chips—how long is that product cycle? Would it be maybe first quarter, second quarter of 2027 that they’ll actually be working with their suppliers to build out what their production is going to be in 2028? Is that the type of time frame that we’re looking at, and maybe that’s when the market starts getting that better clarity that yes, they will be able to hit these numbers that they’ve guided to for 2028?

Kerwin: I think it comes down to printing the numbers that they’ve guided to because if they hit their ’28 numbers, we don’t see that being priced in today nearly enough. Now, I think there are a couple of hard catalysts we could look toward. One is when earnings for Q4 get reported toward the end of January in 2027; we expect all of those large AI spenders to guide to capex for 2027.

Now, we expect that number to be above one trillion. We think that’s largely in consensus, but I think once that number actually gets vocalized and printed, that could be a hard catalyst for suppliers like an Nvidia, like a Broadcom, that this spending is real and it’s happening. But I also think there could be another one in later 2027 when the market starts to bake in 2028 expectations. And if the growth rates continue to be high, if potentially there’s some upside even to what expectations are for 2027, then you could see that 2028 discount that we see being applied start to come away and some positive reaction from shares.

Now, I think it’s important to note on OpenAI and Anthropic—these two huge customers that we expect are going to help Broadcom diversify away from Google—even as recently as last quarter, they’re effectively 0% of sales. Broadcom has this large business that does not include these two customers. As soon as 2027, Broadcom expects Anthropic to become its largest customer and eclipse Google, which implies Anthropic being $60 billion in revenue next year from practically zero in 2026.

Sekera: And that’s not included in your model today?

Kerwin: No, that is included in our model today. But we think that they can guide to this all they want. Once you start seeing a $10 billion step-up every quarter as Anthropic builds in, and then OpenAI is a little bit more of a 2028 story, and we think they’ll be pretty rapidly coming close to Anthropic in terms of customer scale. We think those are going to be hard to ignore, but they’re not being printed yet. And so, right now it’s a lot of guidance, and do you trust management? We do. We actually put out a report in March before this guidance that we expected Anthropic to become their largest customer. We think that’s going to bear out, and we think the market is going to need to see it to believe it.

Sekera: OK. So, in your mind, is that probably the biggest downside risk case in our valuation today?

Kerwin: I think AI, when I think about Broadcom, I think of it as close as you can get to a pure play on the XPU trade. I think the durability of AI spending is the biggest risk long term. I feel confident in what 2027 looks like. We’re becoming increasingly confident in that continued acceleration through 2028. The market is clearly skeptical that this can endure through 2030. I think that’s the biggest risk. If you do see a correction in spending or if capex disappoints when it gets reported in late January or early February of next year, those are the downside risks I think about. To me, a lot of that risk is already priced into the stock at this point. And we see a lot of downside being priced in, not a whole lot of upside being priced in.

Sekera: Yeah. And actually what I was getting to is Anthropic coming on as a customer in and of itself. If they’re not getting any revenue from them today, you’ve got enough sources out there that give you enough confidence that they are coming on as a customer over the next couple quarters.

Kerwin: Yes. I think one of the worries that maybe the market has, or investors may have, is that maybe they are coming on, but maybe that keeps kind of pushing out to the right, and maybe we don’t get $60 billion next year. Maybe that is only $30 billion and some of it gets pushed out. And we don’t expect that to happen. We think the ramp is pretty locked in for 2027, but I would agree that’s definitely something on investors’ minds.

Sekera: OK. And then, just wrapping up the forecast discussion here, free cash flow. I mean, the type of margins that you’re talking about here are so large, this company’s just got to be printing cash. So, what are they doing with that cash? How do you expect that to change the shareholder value or the long-term intrinsic value of the company over time?

Kerwin: Yeah, free cash flow. I mean, CEO Hock Tan has said the primary goal of this company is to print free cash flow. So, you think about the AI chip business, the non-AI chip business, software; there aren’t a whole lot of synergies between those three buckets, and yet they’re able to use all three of those to generate a massive amount of free cash flow. We’re talking about free cash flow margins of about 50% for a company that’s above $100 billion in annual revenue. It’s just massive.

Now, historically, they’ve done a lot of acquisitions. VMware, the largest in history, a few years ago, to the tune of around $70 billion. They took on a lot of debt for that acquisition, but given the scale and ramp-up of their cash flow right now, we actually don’t think that debt load looks anything hard to manage for them. Really, we’re focusing on organic investment, which they do. They have an R&D budget in the mid-teens of billions every year. And then after that, it’s really shareholder returns. And so they buy back a lot of stock. We think that’s quite accretive to shareholders when the stock is trading well below its intrinsic value in our view, and we expect acquisitions to continue in the future.

Now, we think those will probably tilt a little bit more toward software because we think the company likes to remain diversified and the organic growth is coming predominantly from the chip side of things. And they’re already so good at it. We don’t necessarily see what kind of acquisition could happen on the chip side. Maybe some bolt-on software acquisitions, but again, with the scale of the free cash flow that they’re generating and the amount at which we expect it to grow to the tune of pretty in line with earnings at about 50% annually over the next five years, we think a lot of that is going to go back to shareholders, which we really like.

Key Risks Broadcom Investors Should Watch

Sekera: All right. Just to wrap things up, I kind of want to get back to what really are the risks to an investment in Broadcom today? We talked a little bit about how part of our expectations are bringing on new business from Anthropic, but what else, as an investor, should I be watching? What should I be listening for that could be red flag warnings that if I’m involved in this stock and I hear these types of things that maybe I need to go back and reevaluate my investment thesis and some of my assumptions?

Kerwin: Well, it again, all centers around AI. I think start with the pace of the overall AI buildout. Are numbers disappointing in 2027? Are they disappointing in 2028 when those get printed? And then think about the concentrated customer base. I mean, it feels like two different tales when we hear about the reports coming out of Anthropic and coming out of OpenAI. Now we think Anthropic is the largest customer between those two for Broadcom, but we expect OpenAI to be a pretty significant customer too. If that business looks a lot worse than Anthropic’s, if and when they IPO, that could be negative toward that company specifically, their ability to monetize, and maybe their future investment going forward. I’d look at OpenAI’s financials. I’d also look at Anthropic’s financials, which seem to be doing better, at least based on current reports.

And then I would look at the growth rates for Broadcom versus Nvidia. Is this share shift that we expect actually happening? Are we seeing Broadcom’s XPU business outperform Nvidia’s GPU growth? Or are we seeing potentially an architecture shift that you alluded to earlier going the other way than we expect, and GPUs are becoming more in favor? I think it goes back to overall AI spending and the GPU/XPU mix shift. And then for these large customers, are they becoming profitable in AI, and will that allow this spending on Broadcom’s chips to endure?

Sekera: Got it. Will, thank you so much. This has been a great conversation. I think this is really going to help a lot of investors really kind of understand our investment thesis here, especially some of the differentiated points that you brought across as far as why we have such a view as far as where the stock really could potentially be over time versus where it’s trading today. Again, thank you for this. I really appreciate it.

Well, that’s it for today’s bonus episode of The Morning Filter. If you have ideas for future bonus episodes, please let us know. Send them to themorningfilter@morningstar.com. And if you have an interest, you can always read more about Broadcom or any of the stocks that we cover at Morningstar.com or whichever Morningstar platform you use for more details. We hope you’ll join us every Monday for The Morning Filter podcast.

It drops at 9 a.m. Eastern, 8 a.m. Central. Every week, Susan Dziubinski and I review what investors should be watching for in the week ahead, and new research from Morningstar investors shouldn’t miss. Then, we always wrap it up with a couple of picks, pans, or swap ideas based on our fundamental valuations. In the meantime, please like this episode and subscribe. Have a great week.