Transcription
Well, what does it say about the, um, expectations that investors have placed on companies like yours that you can release results reaffirming forecasts, $400 billion in certain types of revenue, surpass estimates in a lot of different ways, and still there to be this sense of disappointment? What does, what does it say about the where we are in the hype cycle around AI? I guess we are kind of like in a very surreal environment. Frankly, I don't think about it. I know it's hard not to, but know, just focus on fundamentals, create value, and stop thinking about your stock price. Trouble is, very hard to do that, but we tried to do it. Well, I mean, it is, we were just talking with Ali, you know, every month or so, every couple of months, there's a new lamb, and it sort of changes the conversation around who's ahead, who's behind.
Um, I'm, I'm, I'm curious about, you know, you have this massive partnership with Google. Very important working with them on their TPUs. Um, I, I and I'd love to get, I'd love to drill down a little bit with you on that. And in particular, you know, how concerned you might be about Google sort of taking more of the chip development and design process in-house, the term customer-owned tooling. Is that something that, you know, keeps you up at night? And are there any provisions that you built into your partnership with them that sort of might preclude that?
Well, you know, that's something we should expect in technology. And we've been in semiconductors for over 20 years. I think that's how long I've been doing this with Broadcom. That's what we do all the, every year, all the time. We have multiple product divisions, each of which we are the best in what we do. And we just keep investing. And essentially, you try to out-engineer any competition. Customer-owned tooling that Google is trying to create, which is trying to design it themselves with the help of perhaps some smaller partners like Marvell, maybe Media, things like that. I call them kind of like angled guidance. But they basically, they are working for Google to try to achieve customer-owned tooling. So we just compete against my own customer in a different part of it. And the whole idea is to be able to create differentiated product and technology that beats what they have. And all these very nice. It's been helped by the fact that the competitor, the real competitor facing all this, is the GPU out of Nvidia. And as long as Nvidia keeps coming generation after generation, that is superb technology, Google has to create their equivalent technology to match that. And that's where we come up, right? In the TPU process, it's quite, it's quite involved. And I'm wondering if there's anything that Google and that anything you can do to help Google democratize that process a little bit.
How do you mean democratizing? Just in the sense of, you know, making, making TPUs more accessible to mainstream developers and not just the elite AI labs?
Well, that's, I think Google is trying to do that by making available TPU, that TPU environment through cloud, Google Cloud, GCP. But a big part of that is the software, the operating system, the software, the compilers that go with it. And that's very much in Google's purview. Of course, would like to help, but I think they got that pretty much out of control in terms of trying to push that program.
Yeah. Um, I want to talk a little bit about Anthropic and the ways that, um, these foundation model builders are obviously in great need of capital in order to finance their needs for, for chips and the compute that are needed to build and develop. Um, can you talk a little bit about the way that that relationship is being structured and financed and the provisions that you are having to take to kind of protect yourself against downside? How concerned are you about the end-user demand materializing and enabling them to meet their commitments?
I am. Frankly, I think in this particular case, uh, we, we were partners with Google to provide compute capacity through TPUs, which we called design, Google to Anthropic. And frankly, there's no shortcut about it. When we first started almost a year ago now, which is, we basically is a leap of faith that Anthropic and generative AI and Anthropic through its business model of doing it, addressing enterprises with code assistance, coding, coding tools would make a difference. We went in on that leap of faith, and I guess to date, that was a great bet. They, they do seem to be in the ascendant, very ascendant at the moment, given their, you know, the fact that they have leapfrogged ahead of OpenAI on their valuation, filing confidentially for an IPO.
Um, so have you not had to worry about, you know, have you not had conversations with investors who might be worried about you getting left holding the bag if things don't materialize the way Anthropic anticipates them?
Well, as I said, we were making a bet on Anthropic. We've been us and Google to jointly doing that. And I guess we were not just making a bet on Anthropic and its business model. We also making a bet that generative AI would be something that would, uh, take off, take off in a fairly substantial way, particularly among enterprises, the way Anthropic goes. And sure did. And that's, and that's pretty good reason for doing this. Started a year ago. I know I'm probably a fraction of what I know today, and what I know today, I probably would be a fraction of what I'll know in six months' time. These things are changing very fast, but the tools we see coming out of come here. Interesting. And the same applies OpenAI and Gemini and out of Google. They are just phenomenal in the sense what they can achieve to help productivity. In Broadcom, we use it. We use those tools for engineering design, code assistance. And it's pretty remarkable what can be achieved in terms of enhancing productivity and enhancing productivity. I think even making to product designs then are probably better than you do it with AI engineers. And these are very good engineers.
That we talk a little bit more about how you're advising your staff to use these different tools, and have you found yourself having to put limits on that? Where we hear about the, you know, the era of token maxing, which I hate that term, but the idea of letting people just use without limits tokens. Have you had to put limits on that? And what does that look?
No, we don't. And frankly, on a bigger picture, we on that stage yet. So maybe it's because we're behind other guys who got to that point and start to put that throttle on what they do. Now, a big part of it is when you use an, uh, a tool like Opus 4.7 and the engineers start using it, they go through a learning process, even how to use it. You can get a tool to be very productive and start, but as you keep using it fairly and you use it long enough, you get very good at it, and then the tools become much more productive. And as you get much more productive, just thinking about it, you can get one senior engineer to, uh, to produce an application design in one week. What it would the other way. What? I take ten engineers, which we paid $300,000 a year and take them three months to produce the same thing. The return on investment is still pretty compelling. So to say we're throttling probably is a question of return on investment and what application you're using. But it is, it's also been able to be trained to be able to use it better and better.
Last year, you announced, uh, a landmark, uh, agreement to work on AI chips with OpenAI. Could you give us an update on where things stand in kind of bringing that to fruition, in particular addressing, uh, what another publication wrote about, maybe some snags there. And you're wanting Microsoft to agree to buy a certain percentage of the chips. Just how are things going with finalizing that agreement?
Oh, no. This, by the way, creating chips with pundits. Pesky example. You said earlier my great relationship with Google. While they tried to do custom on tooling, it's very, very friendly in a way. I and very, and we all have the same goals, competition. We have a great and deep partnership with this guy, as we do with any other members of a pot customers cycle. And, but it's a journey for them. So the first generation of chips would tend to be replaced by a second generation, which gets better and better and better. So for them, it's a journey towards creating custom. So I accelerate the silicon to basically optimize on the workload to replace what is the de facto now, which is tends to be invaded. When we do that, and it, but we only have a few customers, to be exact, six. Each of them on a journey, a different stage of the journey towards, uh, creating their own silicon, which they will use. And obviously, Google is way ahead in being much earlier. OpenAI is one of them. We're engaged with, and we've engaged now with three in over two years. It's pretty remarkable how much we have achieved because we have the I celebrate the silicon working very well in the, in their, in their labs, in the data centers. And we are on track to go into production late this year, and that's it. Off we go to the races.
So there's nothing holding back that partnership at this stage, whether it's needing Microsoft to make any kinds of agreements or anything like that?
No, not at all.
Great. I want to talk about M&A because historically Broadcom has used M&A in, um, relatively substantial way. Um, does the AI revolution sort of temper the need for you to supplement growth using mergers and acquisitions, or are you kind of entering a post-M&A phase, or is that the wrong way to think about it?
No, you, you hit it right on time. And it's an interesting question. It's a very valid question. Does it cross my mind and then on my board, uh, on an ongoing basis? Not regularly, as do we do M&A. So I'll be very flip and give you a straight answer. In the last two years, I between 24 and 26 this year, I would be doubling my revenues. I would create over $50 billion per year annualized on revenue. I'm looking around. What can I buy that even come close to that? And that's the tricky part. I mean, it's a distraction, only I mean, there's a distraction to acquire, go to regulators, further distraction to integrate. Another year. Meanwhile, organically, this phenomenon we call generative AI and our ability to ship, uh, ship them picks and shovels the compute capacity into this demand, which is almost insatiable, makes it very hard to choose M&A over focusing and succeeding in, uh, generative AI compute.
Is there any area where you think you might want to sort of make an exception to this, whether it's maybe photonics, just to name one area, photonics or you mean optics? I'm just throwing anything.
Yeah. We ran an abyss. I ran a business model. 20 years. I tried to very hard. I know, try to avoid bright, shiny objects. All right.
Um, I, uh, I want to ask you, uh, some of the same questions that I've asked some of the other, um, uh, keynote speakers about, um, you know, when you look at the price sensitivity around token usage and the price sensitivity around, you know, when businesses are making prioritizing spending on this, we've seen a lot of companies start to kind of cut back on throttling, use of tokens, etc. as you, as we talked about earlier. Um, We've also seen indications that certain Chinese large language models are starting to gain a lot of traction, particularly given how efficient they can run. Whether it's DeepSea, the list of Alibaba, Kwan. Um. Are you seeing that at all? From your vantage point?
Mean from my one I heard. I can't say we're doing it here in Broadcom, but one I've heard a lot of. And I assume you hear from the same source, a lot of company leaders, CEOs of companies talking about, hey, you know, we're trying to find ways to reduce token costs and all that. And maybe they're ahead of the game where we are. But my sense of enterprise console usage of AI was still in the early innings of this game. And I think we're still trying to learn what we can get out of this phenomenal tool that we have created and continue to enhance and improve with each succeeding generation. So from my personal viewpoint, as I said earlier, and from the experience running at Broadcom, yeah, I don't want my guys to go nuts over spending on tokens. But on the other side, if you can create value out of it, it's and then you have a return on investment that you otherwise could not. Yeah. Why stop. And and this is not going to a video on Kate when we are kids and play a game. And at the point when the monster is about to eat you, your gun run runaway ammunition so you feel more tokens in some of these mines? Not quite.
I'd. I'd be remiss if we didn't talk about the other businesses that Broadcom is known for, networking in particular. Um, you know, your gear has become, you know, increasingly important to the modern AI era. We're starting to see other companies like Cisco wanting to kind of play in the areas that you've dominated. And I'm just curious what you think of the prospects for them to kind of, um, you know, create competition for you there.
Well, we do business in a. I mean, we're very fortunate. I used to really not tongue in cheek. We had a ribbon semiconductor for 20 years, and we've accumulated, uh, portfolio of about 17 semiconductor product divisions in specific call areas where we are literally number one, or we're not even number one, and we continue to invest. And about five, six of those divisions happen to be very important in creating AI chips and AI clusters. Right. So all became that way. And one of them is networking, I should say for switching, switching gears. And that's taking off like a rocket. But we also try to be, as we always have been, when we sell our products to be very, I like to call it strategic. When we look at look at the customer, are they using for particular purpose? Are they sustainable today? One not just one generation, the next and the next and advanced sustainable. And so we push, we prioritize our products to those kind of strategic customers and demand, as I say, for AI and AI networking today is just massive. So it's okay for us that we have such strength and the way we look at businesses, the overflow. I'm happy for you to go to Cisco.