All right, good afternoon. We have a standing room only session. Good afternoon. Welcome. I'm Vivek Arya. I lead the BofA Semiconductor, SemiCap Equipment research team, and I'm really delighted and honored to have our keynote speaker, Hock Tan, President and CEO of Broadcom. Under Hock's leadership, Broadcom has become the most profitable semiconductor company on the planet. 50% plus free cash flow margins, very diversified portfolio across wireless, across wired, AI, networking, storage, infrastructure software. I'm really delighted to get the opportunity to chat with you, Hock. Thank you.
Happy-
Welcome to our conference.
Happy to be here.
Thank you. Maybe, Hock, just let's start with the state of the union. You know, a lot of mixed messages, from the macro, environment. Consumer, a bit softer, but, some of the investments in data center, somewhat stronger. How do you see things right now versus what you thought at the start of the year?
Well, our thinking since the start of the year hasn't changed. Now, first and foremost, we are largely in enterprises infrastructure. We do very little consumer, with the exception of our wireless business. Other than that, it's all in infrastructure for data, for hyperscale, telcos, and enterprises, very much so. What we figured is 2020, 2021, 2022 were very strong growth years. We all see that. It was one of those up cycles in the semiconductor industry, is one way to look at it, but largely to some extent, triggered by COVID. It was extraordinarily strong. We displayed a big part of that too.
We always thought 2023 would be when things level off. Frankly, late in the beginning of fiscal year, which was November last year, if you asked me, my thought would be 2023 could be the point, the tipping point, when our industry, the IT industry, especially for infrastructure, started rolling over. We enter what I call a recession in the industry. We haven't seen it yet. We have seen 2020, fiscal 2023, kind of, moderated a lot, slowed down. When we reported our results, Q2 was still moving quite okay in the semiconductor space, you know, high single digits, close to 10%. Q3 and probably Q4 would definitely slow down year-on-year growth. As I said, it's a slowdown in the growth rate. It's not a rollover.
I get a sense that we won't see that rollover in 2023, in terms of demand, in terms of what we will see out there, because we are shipping, as I've said before, constantly, even 2 years back to today, we're only shipping to what, as much as we can determine, our end customers will consume. I guess the bottom line to what, this is indicating to me is consumption, deployment of our technology is still at a high level. May not be growing as fast as 2022 or 2021, but it's holding up.
Hock, what do you think has helped Broadcom to grow this year, or is helping Broadcom to grow this year, versus, decline in the broader semiconductor industry, including many of your peers, right, who are exposed to the enterprise environment?
You know, I think the simplest answer, if you look at that whole thing, is. You can extend it to the broader perspective of our entire semiconductor industry, where it's known as a cyclical industry. Maybe I'm trying to prove to the world it really should not be a cyclical industry. When you think about it, underlying fundamental demand, as measured by, to the extent you can measure it, conceptually, consumption of the customers, and I'm talking particularly infrastructure, obviously, no position talk on consumer, since our exposure is limited, is that it's always growing for years towards, with certain exceptions, towards the northeast corner. Maybe not as fast as the market, as the industry matures, still growing. Where does the cycles come from?
We used to believe, I used to believe, that's because supply comes in discrete, large chunks. When a tipping point happens, where demand exceeds supply in a particularly extreme way, you know, it's, you start to trigger a cycle. What happens is, when there's a shortage of supply, customers' behavior, they behave in a manner that says, "I'm rushing for whatever products I can lay my hands on," so they place excessive demand out there, excessive purchase orders. If you ship to what your customer, those panicked customers are doing... It's great! When the cycle is over, when supply catches up, guess what? You'll have all that inventory you have shipped out to digest, and then you go plunging down.
If you never ship to one, day one, rather ship to what date you think they should actually consume, then I guess you will actually trace that steady growth rate. I believe that's what we're seeing in 2023. We like to believe in 2021 and 2022, we were shipping very closely to consumption, and we're doing the same in 2023.
How do you put those checks in place? You know, because, I mean, you have, you know, some of the largest, hardware customer, some of the largest customers. If they come and say, "I need, you know, 20 of this widget versus 10," how do you say, "No, we should really be giving you 10"? How do you make that judgment?
I guess there's no hard and fast rule. It's really about being very closely engaged with your customer in a more architectural manner, in a more strategic, deeper manner than, basically, shipping just to what procurement tells you they want, supply. In a sense, it's about relationships, but it's also about the kind of products you're shipping. If you're shipping to a core product, it's probably easier to get that level of engagement than if you're shipping building block products, then you tend to be kind of in the periphery.
Got it. Let's, you know, we will, you know, dig deeper into your AI portfolio.
You know what's coming.
Before we, you know, get to the specific aspects of the portfolio, when you look at the overall spending environment, Hock, that doesn't seem to be growing to the same extent that AI is growing. Do you see that push and pull, that customers are just putting so much more emphasis there, so something else has to suffer disproportionately also?
You know, I don't know the answer to that because, obviously, it's not clear in this early stage of the AI, generative AI, especially explode, a surge that's going on, that we're seeing clear trade-offs. That could be, you're right, but it's not clear to us. If they are pushing out, what you call traditional IT spending, just so that they can invest in AI. I'm referring in particular to basically the hyperscale. Enterprises, not sure at all. I don't think enterprises are really jumping in and investing in AI in a big, in the way the hyperscalers are doing.
Got it. Hock Tan, you've been in the industry for such a long time. You have seen many of these hit or hype cycles, you know, come and go. What's your gut? Do you think generative AI is a long-term, successful investment, or do you think it's too early to make that call today?
I don't know. It's too early. What you're saying, how sustainable it is?
That's right.
Well, AI is something that's been going on for a long time, right? Businesses, companies have been steadily investing in, we used to call it, what? Analytics, then became AI, then, machine learning, now generative AI. It's all you know, putting a label on what is essentially the ability to try to do a better job of figuring out what's going on in our businesses, and in the world around us. Generative AI seems to have kicked things up a notch because it's literally trying to scale out, trying to scale out in a massive manner, large databases, to get us to experience that's totally different from what it used to be. It's been a more, I call it's an evolution.
This evolution will continue, and when you ask, "Is this a sustain?" My rather straightforward, this ingenuous answer is, "Yes, of course it will, and it'll keep growing more and more." Will it go up at the scale we're seeing today? Can't tell.
Right.
I don't know.
Got it. What was the genesis of, you know, your Broadcom's involvement in the AI? You have had, right, very strong expertise in compute offload ASICs for a while. Did that investment start because a customer came and engaged with you, or did you know, have that, you know, vision or foresight that, "Look, AI is gonna be a big thing?" How did your involvement start in AI?
There wasn't any vision here, I assure you. No. The vision is very simple. We have a vision. We're not totally blind, guys. The simple vision of our business is, we have 22 different product divisions, 17 of them, before you fall asleep, I'm just putting in context, in semiconductors. You know our business model. Each of these business division are doing products in a specific vertical in technology, in the semiconductor industry, very narrow vertical in some situation, where we have been, where we've been at it for many, many years, and where we are, we have a leadership position. If we're not in a leadership position, technology, we don't play in that game, and we're 17 of it where we are. Within that 17, I assure you, compute is not there.
Any computing, whether it's ARM CPU, x86 or GPU, which I consider as computing, we're not in that game. There are other players there who have been doing it for much longer, and that ties into our business model. We do what we're very good at doing and where we foresee, by investing and continually driving it hard, to be continuing in the lead on a roadmap in any particular areas we are in. You know the areas we are in. In particular, networking, switching, routing, among other areas, and other, which you know about. This is not an area we're in. One of the areas in the 22 division that we feel very good about, is basically creating Intel IP cores for in silicon. Intellectual IP cores for connectivity, for networking.
I like to use the word connectivity, of all types, at very leading-edge stages of the technology. We use silicon, and we're very good at implementing leading-edge silicon. As part of it, we are very good at converting and designing highest performance type analog components into digital, which is then used across the rest of the company, whether it's in broadband, RF, networking, storage, the highest technology. This division that does ASICs, was approached many years ago by one hyperscaler to try to develop an AI engine, a very dedicated engine, an AI engine focused on the models of that particular hyperscale. Very dedicated, very tuned, optimized for it. After 5 generations, at least 6 years, we are where we are.
Took a lot of work, took some level of investment, but it's, our IP really relates to our ability to manifest in silicon what is a very complicated AI training chip and inference chip. That's how we got into it. It's through having the technology and the skills to create silicon, not an architectural definition play that says, or vision that says, "Hey, AI is the place to be in." We don't do computing typically. It's not an area, not to say we can't do it. It's not an area where we want to come in and compete against players who have been focused on doing it much longer and who probably out-invest us in those space.
Got it. How broad, Hock. We know that, you know, you have one large customer, right, in that compute offload business. What are your plans to diversify that business, right, with other hyperscalers? Do you see other hyperscalers also having the same intention to have, you know, that kind of large internal solution to compete with the GPU alternative, as an example?
You know, hyperscalers, big companies, always have this, after a while, gets into this aspirational intention to consider, "Hey, I could do a chip better focused on my model than what's out there in merchant." That typically happens. Until they get into it, and the water is freezing and boiling, it's not that easy. You do not only do one chip, one generation, that you feel the need to do it. Once you're in it, you got to do the next generation, and the next, and the next. It never stops, because if you stop, then, what are you gonna do? Keep using the older generation. There's merchant silicon out there who is probably continuing to progress, your competitors will be using the better product than you are. Careful about going in.
It doesn't stop a lot of large enterprise hyperscalers from doing it. Having said that, you all have seen it, there's always a news. In truth, it's a much, much less a phenomenon than we all think it is. Because many of these guys know it needs a lot of investment, it needs a lot of focus, and that's not their business model. At the end of the day, you want to compete against a semiconductor, focused semiconductor business with lots of investment and who is doing it for the entire industry, it's hard. Having said that, many of the hyperscale have a lot of resources, very smart people, a lot of need, so they will always like to try. It doesn't mean they will succeed.
Got it. When they come to you and say, you know, "We want to engage with you," how do you know, put some sense into that relationship? What is it that once, you know, once you have a 1 or 2-year visibility, then you're like, "Okay, whatever you want to do?
We go by financial return, right? My model is all financial. They're willing to pay for my engineering, I might consider it, but at the end of the day, that's not... I'm not, I'm never going to outguess, out-argue to a customer that his model is not a smart model. If he's willing to pay, happy to do it.
Makes sense. When you look at the long-term kind of mix between a general purpose versus an ASIC solution, do you think it stays where it is, or do you think that the trend is a lot more towards general purpose over time?
I don't know about AI. Let me tell you, historically, on the products we have done, Broadcom, as you notice, I have 17 semiconductor divisions. Effectively, only 1 is in ASICs, custom silicon. There are 16 others, and they're all merchant silicon. Where do you think I'm gonna likely come out in my experience? Merchant silicon tends to prevail long term. Long term might take a while, because some of my customers are very rich, and they are very ambitious, and they're willing to do it. I'm happy to oblige, because I have a division that focus on doing the best they can for their customers, and it's still consistent with the model we have.
Got it. The complementary part of those accelerators is networking. I think, you know, you mentioned the goal to kind of almost quadruple your switching sales tied to generative AI, to almost $800 million this year. What's helping that? Because that's, you know, a very significant growth rate, right? It's, in fact, a higher growth rate than we even see in the accelerator businesses for some of the other players.
Well, yeah, because what is happening is. I mean, for the last several years, as more and more enterprises, and that includes cloud guys, put in more resources, and then some of the cloud guys have been running AI networks, clusters of AI that need to be networked together. We have been supplying products, networking switches, fabric, that enables clusters of AI engines to work together at larger and larger scale as the models for basically models for predicting, for analytics, keeps growing bigger and bigger. We've been broadening it out. But for years, high-performance computing, which is another way of calling AI, by the way, used to be called HPC, high-performance computing, then we all call it AI, then we call it something else, machine learning, and now it's back to AI, generative AI.
Whatever it is, the scale is growing. What is particularly clear is AI networks, AI workloads are different. High-performance computing workloads, it's what they call elephant workloads. It means, unlike traditional IT infrastructure, when x86 run in your data centers, in your company, the number of workloads is much simpler, much less. AI workloads tend to be large. Good reason they're called elephant workloads. Huge workloads. When you want to then run AI models through clusters of AI compute engines, because each of those engine is now no longer big enough, high performance enough to run those large database, to create those small to get any good outcomes from those large models, you need to network them.
You find you need a different kind of network, because you need those workloads to run faster, they're larger workloads, and to get full output in terms of performance from your GPU or TPU, you really want the network to be able to match the performance of the GPU that keeps improving. You don't have to, but only means you run your GPU slower, and you're happy with lot smaller models. Last year, generative AI showed up with OpenAI and ChatGPT, where they talk about taking huge database, large language models, LLMs, as we call it now, and load it all into this massive computer, computing machine, to be able to literally create analytics out of a huge database, especially in training. You need to do it with lots.
You hear about lots and lots of AI engines, which have to work in a synchronized manner. You have to network them. Then we suddenly see a demand, starting last year, in particular, of a different actually, almost for us, seeing it, almost for the first time. We've known about it existing in some hyperscales. We now see it for first time, that we need to create a different kind of product to optimize this kind of workloads. Since then, if nothing else, this has gone even more extreme. When we hear about trying to network from used to be few to 4,000 nodes, which is GPU, to 8, 16, and now we're talking about 32,000 by next, end of this year, early next year. To do that, you need a whole new scale.
You need a whole new, not just switching, you need a fabric. That's what pushing us now to step up, to create, almost, to innovate on, more advanced technologies to handle that. Unfortunate thing is, we have most of all the building blocks of that technology in our, within our, the company. A lot of it, of what we do, what we're seeing here. When you think about AI, the theories behind AI has been around for a long time. By the way, an ability to network and run jobs completion in a training chip, or in a cluster of training chips simultaneously, and exchange parameters, exchanges very efficiently, so that by the time when you finish your exchange, you just finish the next job completion.
There's a law to that, the law of physics that makes this happen, that's calls for a different kind of network that is one step higher performance than traditional data center networking. We have the technology. We're out there engaging with all the hyperscalers in helping them create, in a much better way, this, their AI networks to essentially optimize or utilize their GPU engines or TPU engines much more efficiently, effectively than what they have been able to do just using regular networking. That's a whole new step up, and it's been a great driver, as we announced, and as you pointed out, to that business that relates to AI.
Right. Over time, what is the right way to think about growth in switching in that? Before I get to that question, how do you differentiate, Hock, between a switch that is a regular switch versus a switch that's for generative AI? Is it a certain port speed? Is it a certain SKU? How do you, how do you make that distinction?
Unless, until we come out with the new products. Last year, we came out quickly with a product in partnership with Arista Networks, called the 7800R Arista, and that's Arista markets as an AI switch. That's clear because it's not a Tomahawk sitting in there, it's a Jericho. Why? A Jericho, which is used for routing in telco networks for the last 10 years, older generation, has a lot of the technology and architecture that is perfect for AI networks today. Just need to am up the speed, but you have that. It's like recycle the technology and stick it in here quickly, and bingo, we got it for Arista Networks 7800R. That's great. Go. That was 8, 4,000, 8,000 nodes.
They go to 1632, so we have to come out with a new generation of Jericho AI switch, and we launched a Jericho3AI, just to kick up the PR on it. It's really meant for AI switching, the successor to the 7800R, for use to some of the hyperscalers who wants to use Ethernet to drive their AI networks. That's almost no question. When you go up to scale up so big, 16,000, 32,000, 64,000, it's so much easier to use Ethernet if you can make Ethernet meet certain criteria of lossless and high latency, something Ethernet has not been good at doing, but InfiniBand has been very good at doing. InfiniBand's problem is scaling out that much is tricky. It's not meant to scale out to the size it scale out.
It's possible to be done, you just have to stage the architecture that makes it harder. Ethernet, on the other side, is just known to be just all out there, but it loses packets and latency, not so good. What we do is we put in the technology that makes it lossless and improve the latency to the level is acceptable. What it's meant by acceptable is that you can run 16,000 GPUs simultaneously, exchange the data, and keep running it. That's it, and that's a whole new scheme. To answer your question, beginning to get clearer and clearer, what is AI network switch and what is AI fabric versus non-AI fabric? Problem that now it is. Before that, not so clear.
We just had to engage with our customers who work with us. You get a sense what they use for AI versus what they don't. Fortunate thing is there are not that many customers out there.
Got it. You think customers are still investing in non-AI switches of Broadcom this year?
Of course. You gotta run workloads, you know, it's a cycle. I mean, many of the cloud guys, unless they want to sweat their equipment, even then there's a limit in which they can sweat their existing equipment, you need to upgrade, you need to refresh, you need to expand to run a bigger thing. Don't forget, your AI networks are meant for very high performance workloads. You don't necessarily want to use that network to run your traditional enterprise workloads. Makes no sense.
What proportion, Hock, do you think, roughly, of workloads are AI-related at hyperscale today?
I have no clue.
Do you think it's, like, less than Like, if I were to take the proportion of your switch sales, right, versus overall switch sales, is that a reasonable way to look at that metric?
No, because I have an install base by the billions, that's still running-
Right
... traditional workloads. This new little stuff is barely up and running. This is not that huge compared to what's running out there. No. This, the traditional workloads is still real solid. If you look at my numbers for networking, strip out AI, I'm still growing in networking. Strata , I offload Strata the $800 million this year, I still am growing. It still is a need. Enterprise is still buying the thing. Even hyperscale is still buying. They need it to run workloads.
Got it. Traditionally, you know, we have seen decisions on computing and networking kind of done, you know, a little bit separately, you know, best of breed in each. In AI, do you think that, you know, just given the tight connection, right, and the low latency requirements, that does it give an edge to your GPU competitor because they have InfiniBand switches, because they have a new Ethernet switch? Do you think just by the virtue of the fact that they have a larger share in accelerators, does it give them a chance to gain switching share versus Broadcom?
You know, best way to answer it, because we're talking about new phenomenon, but if you go back in history, I mean, how many companies, semiconductor companies in particular, have come out historically, and many of you invest in semiconductors over the years have seen it, where just because they do a core chip, before you know it, they want to sell the whole platform. I would almost say it's not natural. If you have seen over the years, platforms disappear because customers want flexibility and choices. Just like why I say merchant gets to win at the end of the day, they want choices. They want a disaggregated and distributed model in data centers. Now, in AI, and generative AI in particular today, guess who are the customers? They're the bunch of the largest, smartest guys around, the hyperscale. Guess what?
Small enterprise, maybe they buy a black box and be done with it. Hyperscale, no, they don't.
Right.
They look at it, they assess it, and, no, they want flexibility.
Right.
Which is why we engage with them on AI networks.
One last kind of technology topic that's also gaining some industry attention is the consideration of co-packaged optics, right? Or whether the industry will stay with the current format, which is to use these pluggables with their separate DSPs, right? Where does Broadcom fall on that spectrum, right? Do you think it's gonna be co-packaged optics in a big way, or you think those are niches where they make sense versus it's pluggables for the most part?
Well, it's back to my customers at the end of the day. I mean, my whole business model is about addressing in technology and engineering what the customer wants. By the way, I may not ship what they want, but I give them the technology they want. It's basically this, it could exist both ways. Just like, because some customers will prefer, even as you offer co-packaged silicon, like as fabric or switch, and double the speed and yet a 10% reduction in power, double the speed. You think for AI networks that will run at high performance, lots of power coming out, that power is becoming the constraint to building AI data centers in scale. You think they want co-packaging.
The drawback with silicon photonics, which is co-packaging, is effectively what it is the active components of fiber optic modules that connects a switch to the servers or to other switches, is integrated into my switch. One big thing. Cool, right? A lot of content, lower power, much better proximity, performance better, all that nice stuff. Yeah, but guess what? Pluggable optics gives you more flexibility if you are building a data centers, right? You, you may not want to populate the entire switch in day 1. You just want to populate maybe half. Well, you can't if you buy my co-package switch, I give you the whole thing. I can see that there are many situations where they don't want a co-package, and other situations, particular power constraints, that they may use a co-package.
I've seen the cloud guys, they pick and choose to optimize their very heterogeneous data centers.
Now, Hock, the other way Broadcom has kind of been unique is, through your investments in infrastructure software, right? Whether it was CA, whether it was Symantec or Brocade. How would you grade those investments if you take a look back? Like, so if you know, go back to what was your original strategy about investing, right, and diversifying outside of semiconductors, right, versus the results that you have seen, how would you kind of grade those investments?
Software is a great financial business for us. It has been. I mean, we make great margin, great cash flow, and bloody pay your dividends, guys. It is. I don't have to touch the semiconductor, guys, it almost pay your dividends every year. That's pretty good. It's boring, which is great. You look at me like I've gone nuts. It's. See, it's the way we buy the assets, then we manage it. We buy, then we focus on large footprints of customers.
Right.
Today, we still invest, but we give these customers the best, literally, support and feature request requirements they need. That's it. We have a limited number of customers, 600, 700, among our software group. You can imagine we can spend to give them that experience, and we can focus to send an experience. That's great.
Got it.
That's been a great model.
Got it. Has it driven synergies? Like, have you seen synergies between the semiconductor side and the infrastructure software side? If there aren't, you know, is that a good or bad thing? Like, how do you look at driving synergies between the two groups?
We never look to drive synergies. I mean, my 17 product division, semiconductors, they don't try to help each other. They don't try to. Each is run by general manager, each has a management team, each are told to be totally selfish. You know, don't take one for the other guy. Come on. If the other guy cannot stand on his own, he doesn't deserve to exist. That's why they need to be, to try, to be in an environment where they learn how to survive and prosper on a standalone basis. That's the key job requirement of the general manager and the team. You survive on your own. If you have to depend on somebody else to prop you up, that's bad. We don't do that, very limited synergies.
The only synergies is, you know, when we go to leading-edge silicon technology or packaging technology, yeah, we develop it centrally, and we share the know-how and IP. Other than that, and some core intellectual property, in like SERDES, A to D converters, stuff like that, high speed. Other than that, no, they are each on their own. You think software and hardware, even less.
Got it. Is there a certain mix, like if you look at Broadcom, say, five or 10 years from now, is there a certain mix between semiconductor and software that you're looking for? Or you think whatever develops naturally based on the strength of each segment?
From my last answer, you already heard it. No, we don't. We don't have any vision in my mind that I should be 50% hardware, 50% software. We are really a technology company, and we look at tech, areas in technology to invest in, to buy first, invest in, develop, even grow, where they, we can be very successful. We look at VMware the same way. Except in VMware case, unlike the other software companies I have today, this business, we actually see to grow. That's why we're investing in it. We actually believe it can grow, and we will invest big time.
All right. When you make R&D decisions, Hock, right, Broadcom's R&D intensity is among the lowest, right, in this industry. Obviously, you're still growing it, right? Because the overall business is growing. How do you make that trade-off? Like, if you were to invest an extra $1 and you were able to drive, you know, 100 basis points faster growth, how do you make those kinds of decisions? Like, is it to maintain the profitability at odd levels, or is it to grow faster? How do you make that trade-off?
We never manage any of those division in semiconductor as an example, by numbers. We don't. You notice each of them, 17, have different levels of profitability. I'll be honest, I won't tell you what each one is, but it ranges. Never mind, it ranges very much-
It's okay if you tell us.
It's not okay for my customers.
Yes.
More than that, it's more natural what each business is like, how, what you need to invest. Don't forget, the management team and general manager, believe it or not, which makes my job and Charlie's job easier, they are empowered to invest as much as they need. Seriously, you think we're kidding? No. They can invest as much as they need in their business to sustain and grow, in some cases, their business, to sustain their business as much as they need, provided they just focus on that market, that basic vertical they are in. They're not allowed to go beyond guardrails. That's a requirement. If you don't go beyond guardrails and just focus on investing as much as they need, we'll leave you alone. They invest as much as they need. I'm telling you, we don't tell them how much to invest, how little to invest.
What they invest is what they show, and we pull it together, we show you the numbers. People say, "Are you underinvesting?" No, we're not. They are each allowed to invest. It's kind of telling what's happening out there, that if you focus on being the best and just focus in this narrow area, you don't have to spend all that money that's out there. Just investing more, the R&D dollars doesn't get you anywhere either, because it's not just about spending dollars, it's about getting the best people.
Right.
We believe in semiconductors especially, we probably have the best teams in the specific areas we're in the world, bar none.
All right. Just the last thing, Hock, when you look at the, just the broader semiconductor industry, there is a contention that the industry can grow faster than it has in the past because pricing is a lot more resilient, right, versus it has been in the past. At the same time, some end markets, such as, you know, consumer, are not growing as quickly. Let's say if in the past we thought semis would grow 5%, 6%, do you think they grow, if you look at the next 5-7 years, they grow roughly at that pace? Or you think that it is a new industry, and we can actually grow at much higher base than we have managed to grow?
I don't know the answer to that, Vivek. That's a very good, thoughtful question, and, you know, not sure. You can see that at some point when supply exceeds demand, you know, there'll be a lot of price pressure because customers are doing that. Have customers changed their behavior, is what you're saying?
Right.
To some extent, we see that, but it's still early. After two years of customers, literal customers, biggest customers, almost having existential issues over supply, they're kind of like, okay, they're more well-behaved. Let's see how long that lasts. I like to tell customers, "Hey, make a long-term agreement with me. We are partnership. I give you engineering, I show you supply, you work with me closely on your roadmap, not current, but future." Yeah, we have some of this that exists, and I like customers to believe that is a route to grow in an industry that has matured and is consolidating. I like to believe that. We see some of that, but I think it's still too early to come to any conclusion that that will translate into a growth kicker for the entire industry.
Got it. Then finally, Hock, you know, Broadcom has, right, some of the most premium margins in the industry. How do you look at the progression of those margins over time? You still think there is more leverage, still the ability to expand profit margins further? Or you think that those premium margins can attract more competition over time?
No, nothing to do with that. There will be, it's nothing to do with competition or anything else if you are the best of the best in the narrow, in the specific verticals you are in to doing it. What we've always articulated out there is, you know, the nice thing about in technology, especially semiconductor, semiconductors, is simply this: it doesn't say stay still. Your customers, their end users, always want a better product, better features, lower power, higher speeds, more ports, whatever. You have a chance to always come out with the next generation. In industrial, that next generation or automotive may take five years. In consumer, like smartphones, it comes out almost every 12, 18 months, networking, two years, two and a half years, faster and faster. Storage, four, five years.
It does, and each time we come up with new generation, you literally create more value to the, to your customer. Literally, more value. Think about Tomahawk 5 versus Tomahawk 4. The bandwidth doubles. You think I'd go to a customer and say, "Can I have 2x the price," since I'm doing that? They save, I think, topology 4x. No, we raise the price something like 40%. It happens, every generation gives you a chance to extract more value. If you average across a broad portfolio, like we have, across different end markets, and know that every year we have a new feather in the new generation of products, depending different areas, we have what we've seen empirically over the last 10 years in Broadcom.
Our gross margin, product margin, indication of value of our products to our customer, on averages grow 100 basis points of gross margin. You take that on an OpEx that will grow roughly even the level of sales, if not lower, you have an operating margin that translates into an expand, constant expansion. And we have seen that happen over 10 years, and I expect to keep continuing to see that happening over the next 10 years, given that industry, semiconductor technology, continues to evolve.
Terrific. Thank you so much, Hock. Really appreciate your time.
Thank you.
Thank you.
Glad to be here.
Thanks, everyone.