President of Semiconductor Solutions, Broadcom
Scaling AI Infrastructure for the Agentic Era | Charlie Kawwas, Broadcom | RAISE Summit 2026
Charlie Kawwas, President of Broadcom's Semiconductor Solutions Group, joins Tony Kim (BlackRock) to discuss scaling AI infrastructure for the agentic era — networking, custom silicon, and the connective tissue of the AI data centre.
Recorded live on the Master Stage at the Carrousel du Louvre, Paris, on July 9, 2026. Livestream powered by Cursor.
RAISE Summit is Europe's leading AI gathering, bringing together 9,000+ founders, executives, investors, researchers and policymakers. Learn more: https://www.raisesummit.com
Subscribe for every session from RAISE Summit 2026.
About Charlie Kawwas
In 2025 and early 2026, Charlie Kawwas, President of Broadcom's Semiconductor Solutions Group, has been a prominent voice on the infrastructure demands of artificial intelligence. He has described AI infrastructure as "civilization's next-generation operating system" and a "critical utility" for the global population. In October 2025, Kawwas joined OpenAI executives Sam Altman and Greg Brockman, along with Broadcom CEO Hock Tan, to announce a partnership to design a custom AI chip and system, with Kawwas stating that Broadcom is "defining civilization's next-generation operating system." He has repeatedly identified energy consumption as the primary challenge to democratizing AI, noting that compute demand is increasing by 10x annually and that power access is the main hurdle for deploying large AI clusters. Kawwas has highlighted Broadcom's engineering focus, stating that over 13,000 of his 15,000 employees are engineers, and has showcased innovations such as the Tomahawk 5 networking chip and silicon photonics co-packaged optics, which he said reduces power consumption by 70%. Kawwas has also shared his personal background and career philosophy in multiple public appearances. He has spoken about arriving in Montreal from Bethlehem with no high school diploma and little English, and credited Concordia University for giving him an opportunity)Skip. In June 2025, he received an honorary doctorate from Concordia, where he told graduates that his formula for success is "WEL": work hard, embrace challenges, and live adaptability. He has advised engineers to choose their own mentors, citing Hock Tan as his own, and urged young professionals to embrace AI or risk being left behind. Kawwas has also discussed his daily rituals of exercising, calling his mother on video, and praying before bed.
Transcript (102 segments)
I
Interviewer0:04
All right, great. I think this is the last one or
C
Charlie Kawwas0:08
Probably? Oh, everyone's tired.
I
Interviewer0:10
Okay, we'll make it try to make it more entertaining. Well, great to see you here. You know, we're in France, Paris. So you know the land of haute couture, and I think that Broadcom in a way is the couture of chips.
C
Charlie Kawwas0:35
Thank you.
I
Interviewer0:36
I think so.
And so, you know, obviously with AI and the rise of the XPU, everyone wants to build XPUs and they all come to the house of couture at Broadcom. So why don't we discuss why the XPU has become such a pivotal thing for basically every hyperscaler and every foundation lab in the world.
C
Charlie Kawwas1:09
So, first of all, thank you for having
I
Interviewer1:10
And we're going to have some show and tell.
C
Charlie Kawwas1:11
Yeah.
I
Interviewer1:12
Yeah.
C
Charlie Kawwas1:12
Yes.
So, thank you for having me here. And this is actually a great introduction about the haute couture. I never thought of it, but in a way,
I
Interviewer1:21
I just thought of it now cuz we're in France.
C
Charlie Kawwas1:24
It's actually a phenomenal idea because look, at the end of the day, this AI wave was really started and driven by what we call general AI compute driven by Nvidia. And Nvidia has done a great job to build this sort of big racks of custom and proprietary platform which is a great platform. But the people who are spending more than 80% of the money are the frontier labs. They're the ones who are consuming all of this hardware and then they build the software that sits on top. What they've realized is this general compute comes with a tax. The tax is first the efficiency of that compute platform. Meaning it's built for everybody. If you have a specific workload in your AI frontier model, guess what? You're paying that extra tax. And then the other tax is the high margins that come with it. Given the big spend that they're doing when it was hundreds of millions of dollars to a few billions, it makes a lot of sense to use a general purpose platform that's pre-built for you in Iraq. But when you start spending 50 billion, 100 billion or maybe 150 billion as you know the top four in the US each spending 150 to 200 billion, that tax is about for each player 100 to $150 billion that they could have gotten in tokens.
And as a result they realized that if there is a way to build their own platform fairly fast at the same pace and rhythm that Nvidia does it which is once a year, it's quite difficult they would love to do that. So the first journey we did was 13 years ago starting with the TPUs from Google. Their initial idea was we do search using text, can we actually do search using voice, and to do that was the creation of the TPU. But as AI developed we realized that these chips are getting bigger and bigger, so the chip we did last year and I actually I'm going to show you these for a reason and I'm going to pass them to you.
I
Interviewer3:38
Okay.
C
Charlie Kawwas3:40
But you have to promise to give them back to me, okay.
I
Interviewer3:41
Because they're very expensive.
C
Charlie Kawwas3:42
So this is for example an XPU that was built and shipped last year. And as you can tell, it's hard to see, but it's made of different chiplets. Some are memory based, some are networking based, and some are compute. Okay. What we're building this year and shipping is this chip. As you can see, it got bigger. So, I'm going to ask you to hold one.
I
Interviewer4:05
Okay. And this is last year's.
C
Charlie Kawwas4:07
This is last year's. And so you can see last year to this year you are now getting three times more the memory. There you have two cubes of memory. Here you have six cubes of memory. And the compute that's in the middle is almost 3x stronger. What ended up happening is we said for next year we actually have to build this.
I
Interviewer4:29
Right.
C
Charlie Kawwas4:30
So now
I
Interviewer4:31
More how many more times memory?
C
Charlie Kawwas4:34
This is 139.
I
Interviewer4:36
Yes.
C
Charlie Kawwas4:36
Okay. So in two years it's almost an order of magnitude.
I
Interviewer4:40
And then when you look at the compute here you have one large compute, here you have four of them.
C
Charlie Kawwas4:45
Yeah.
I
Interviewer4:45
But then you go to the new technology in semiconductors that allows you to shrink the transistors and then you add a lot more.
C
Charlie Kawwas4:52
Yeah.
I
Interviewer4:52
And so I'm going to give you this as well. And then what we're working on for 2028. So this is 2027. Literally this is going to be the size of the platform. That's 20. This is going to be the size. So if you put next to it
C
Charlie Kawwas5:10
Just these. Okay.
I
Interviewer5:12
So this is okay. So 28, 27, 26, 25.
C
Charlie Kawwas5:17
Correct. So the reason I want you to see that is I want you to understand two things. One, the speed at which this technology is moving at. Okay. And hence the innovation that it needs to go through has to be extremely rapid. But then two, and I'm going to open this for you. Literally that's how we build these chips. These are actually modular. This is why people come to the old couture as you said.
I
Interviewer5:44
Yes.
C
Charlie Kawwas5:45
Because we actually take these platforms and pre-build them. So these little chiplets that go in the chip, we pre-build them, make sure they're production ready, work with partners like Samsung and others, create these memory blocks, which is one of the reasons why the prices are going higher because look, we're putting so much more in them. Create these IO Ethernet based networking capabilities and then work with our partners to create now remember we used to have half of this one of these. Now we have 1, 2, 3, 4, 5, 6, 7, 8 times two. This is a double-decker. We're actually stacking these chips back to back or actually the correct term is face to face. So you'll have 16 times what we had two years ago. 16 times in a single chip. And the first chip we deployed a million of.
I
Interviewer6:40
This doesn't even look like a chip, Charlie. This looks like a monstrosity system. It's a beast.
C
Charlie Kawwas6:45
Yes, it is. Actually, that's what I call it at work. This you call it the beast.
I
Interviewer6:48
This is the beast. And but this is we're building four of these for the largest four LLM players.
C
Charlie Kawwas6:55
Okay.
I
Interviewer6:56
And it's modular. So the beauty of this to your point is it's oat couture. You come to us, you tell us what kind of tuxedo or dress you want.
C
Charlie Kawwas7:05
We will customize it and build it for you and remove that general compute tax. So that's gone. And this will be purpose-built for let's say your inference workload, decode workload. And this is what makes this fast and great. And as long as Nvidia continues to develop great technology every year to match that cycle. Yeah.
I
Interviewer7:27
There's only one other place to go to.
C
Charlie Kawwas7:29
Yeah. You mentioned the one-year cadence that Nvidia, it used to be two or three years, right? Correct.
I
Interviewer7:35
And then they accelerated it to one year every year. So the pressure is on. And so for most companies they can't keep up with that cadence. The only one maybe is you then right that with this modular Lego architecture.
C
Charlie Kawwas7:53
You're exactly right. So if you are one of the let's say the big five in the US which today are spending maybe 90% of the spend in the world are OpenAI, Anthropic, Google with Gemini, Meta and xAI. This is it
I
Interviewer8:08
Amongst these five is literally 80, 90 plus percent of the market.
C
Charlie Kawwas8:13
Yes.
I
Interviewer8:14
And four out of the five have realized the Nvidia technology is so good but they can't be buying this general compute platform that's custom at the rack level. And so they come to us each one of them and it's public and they say look what can we do together to get purpose-built platform for me that's tailored using your term actually tailored to my workload and it's not at 80% margin and that's what we do for a living because we focus just on building this. So we make sure each of these Lego blocks are production ready before you come and talk to me and so when you come and talk to me it becomes an integration exercise. So what we announced with Jalapeno for example is it's the fastest chip that was built in 9 months and it's actually something similar to this. They did that in one less than a year.
C
Charlie Kawwas9:08
Yes.
I
Interviewer9:09
And so you have basically weaponized or enabled these foundation labs to build their own silicon in a year.
C
Charlie Kawwas9:16
Correct. And you know and so five years ago your AI chip business was zero and now it's going to be a hundred billion in next year. So zero to 100 billion in five years.
I
Interviewer9:29
More than 100 billion.
C
Charlie Kawwas9:30
More than 100 billion in five years.
I
Interviewer9:31
Remember 100 billion is not good enough.
C
Charlie Kawwas9:33
It's not good enough. More than 100. And
I
Interviewer9:35
But yet how many of these are really scaling, you know, of those four, you know, in the 27 period it's still in early formation, right?
C
Charlie Kawwas9:50
It is, but just to give you sort of an order of magnitude view of this, think of it as this is like about a million a year.
I
Interviewer9:57
Okay.
C
Charlie Kawwas9:57
Think of this is about maybe 2 million a year.
I
Interviewer10:00
Okay.
C
Charlie Kawwas10:01
But remember this is 3x this.
I
Interviewer10:03
3x that. Yeah.
C
Charlie Kawwas10:04
So that's 6x in one year.
I
Interviewer10:06
Yeah.
C
Charlie Kawwas10:07
Then you go to this. This is probably 4 to 5 million.
I
Interviewer10:11
Wow.
C
Charlie Kawwas10:11
And this is 9x this.
I
Interviewer10:13
So now you're at 40, 50x
C
Charlie Kawwas10:16
Of this one.
I
Interviewer10:17
And then the beast.
C
Charlie Kawwas10:17
And then the beast is really 3x of that as well.
I
Interviewer10:21
Right.
C
Charlie Kawwas10:21
Okay. So when you think of the scale at which people are building these things, it's a race toward super intelligence and AGI as you said. And these four to five frontier labs believe the first company in the world that would enable this platform is what I call the agentic AI platform. The first company that would enable a fully integrated GPU, XPU, CPU in a single platform potentially can be the first to reach AGI. So with this in 28, I think we're going to start seeing a lot of capable agents, experts in certain domains that would now finally have the technology that would enable this. So this is the exciting piece about what we're doing and why people keep coming back and saying, I'd love to use the Nvidia chip, but I got a general compute tax and then I got a margin tax. So if I want to drop ultimately my token cost per dollar, okay, I can get two to three times the compute for the same dollar, which means the token price comes down massively. So that's ultimately what they're driving for.
I
Interviewer11:30
So you're effectively a couture design house with low taxes.
C
Charlie Kawwas11:37
Correct. Correct.
I
Interviewer11:39
Okay. And the other thing that's the topic du jour more French.
C
Charlie Kawwas11:45
Yes. See, I told you you speak French.
I
Interviewer11:47
The topic du jour in AI world is code design. And Jalapeno. You brought up Jalapeno and OpenAI and the tight integration of model development helps define the chip architecture and then the chip performance then is optimized to the model architecture. And that this tight integration of code development and the only way to do that I guess is if you have your own XPU.
C
Charlie Kawwas12:16
That's absolutely right. And there's two portions to that haute couture. One is exactly the deep code design for the XPU you described. Okay. The second is at the rack level. So not only we have to do it at the XPU level for their LLMs and their specific use cases, but we actually engage heavily into well how do we make sure that the entire rack is open but actually it's co-designed for them. So part of that is where we bring our networking and Ethernet capabilities and this is where we make sure it's open. So we started off by making the front-end networks Ethernet based. Open. Then we went to scale out. Now everybody's open including Nvidia. The next battle is scale up which a lot of these guys use these blue chips from us, these little chiplets we call them, and they're all Ethernet based. So when we co-design all of this, it's the code design at the rack level, starting from the XPU with each of these blocks, then the entire rack, and then it gives them a choice. And honestly, if Broadcom doesn't give you the best technology, they have a choice to go with somebody else. But this is where it's been multigenerational and quite successful for us because as you said we started from less than a billion dollars two years ago in this space and next year in three years we're going to be over hundred billion dollars just in this space.
I
Interviewer13:48
You mentioned open, you know the openness of the interconnect standards and you guys are full stack. I mean you do networking, you do packaging, you do obviously the XPU design etc. IP blocks. And so they want to be open. They want that open standard. What about open source? And is there the idea of an open-source foundation lab? And could they then also do you see any market for XPU development for open-source foundation labs or open source models? Not the closed foundation labs.
C
Charlie Kawwas14:33
Yeah, actually you're absolutely right. Actually, one good example that I'm going to use here is what we've discussed earlier on today, which is how many of you have heard of a company called Sambanova? Just raise your hand. That's good. A bunch of people here. So maybe 10, that's good. Maybe like 15% of you have heard of Sambanova. So Sambanova system is actually have been co-developed deep engineering with Rodrigo and his team. And we've done three generations of this. And what Rodrigo does is he uses his XPU to run a closed frontier model. So a state-of-the-art, but he brings an open-source model like a Miniax or DeepSeek and runs it on the same chip. And hence he enables a hybrid solution where if you need for specific use cases to use let's say Claude or ChatGPT, you can run that on that XPU, but if you now suddenly need to use something that doesn't require state-of-the-art stuff and you want to use an open-source stack, you use it and you pay zero for these tokens. So that model already exists today and we've proven it with a few of the startups. I think we will start seeing that more and more of the XPUs as we move forward.
I
Interviewer15:53
All right, Charlie, in the last couple minutes here, how do you see the next couple years here and the path forward? You're showing the 28, you're showing something that looks in 2028. What do you anticipate it's going to be the topics for the next couple of years beyond?
C
Charlie Kawwas16:13
So that's I think the $2 trillion question. Okay.
I
Interviewer16:16
I mean Jalapeno, there's more than Jalapeno, right?
C
Charlie Kawwas16:19
Yeah, there's more than Jalapeno. There's definitely more spicy chips.
There's a lot more XPUs that will be coming out both from the large labs and a lot of the startups. But what I see is an insatiable demand in gigawatts. We, for example, we have announced for one of the large five that we've discussed that this year we're putting in about 1 and a half gigawatts. With one of the chips I showed you, next year we're going to probably put in over five. We already contracted
I
Interviewer16:53
With the same customer.
C
Charlie Kawwas16:54
The same. So now we have
I
Interviewer16:55
For XPU.
C
Charlie Kawwas16:56
Yes.
And so we're going to go from one and a half to six and a half because one and a half plus five. The year after that, I'm pretty sure it'll be over 10 incremental. So, if you look at the next three years for each of these lab guys, they're going to go from one to two this year
I
Interviewer17:15
To over 10
C
Charlie Kawwas17:19
But over that three-year journey, they're going to let's say they'll go from two to five to 10. So, that's 17 gigawatts starting from one today.
I
Interviewer17:27
Yeah. Which means the amount of power that's needed, the amount of wafers that are needed.
C
Charlie Kawwas17:36
Yeah.
I
Interviewer17:36
And then the amount of memory that's needed. It's actually going to make a challenge for the rest of the non-AI businesses which at Broadcom we do. So remember we have 16 franchises. Five of them are in AI, 11 are in non-AI. So part of the challenge that I think we all have to live through as we live through the exciting life of this AI as we scale up AI into AGI, we need to make sure that the rest of the businesses survive because AI is not just going to live in data centers. AI has to get to the edge and these other businesses have to actually enable this edge AI. So we've got a great exciting journey on the AI track with power and all of these things and to the point you know we're building our own factories in Singapore to build these big beasts and big chips because existing technologies can't enable that.
C
Charlie Kawwas18:30
Yeah.
But as we do this I think the challenge for all of us is AI has to migrate just like the internet started from data centers all the way to smart devices today. The same thing has to happen with AI. And I think for us at Broadcom, it'll be exciting because we'll play in both pieces.
I
Interviewer18:48
To the couture house of Broadcom. Thanks, Charlie.
C
Charlie Kawwas18:53
Thank you. Thank you.