theCUBE + NYSE Wired: Physical AI & Robotics Leaders | Vijay Nagarajan, Broadcom & Harminder Sehmi, Blaize & Gopal Hegde, Sima.AI

Clips
More from theCUBE + NYSE Wired: Physical AI & Robotics Leaders

Vijay Nagarajan

VP, Strategy & Marketing

Broadcom

Harminder Sehmi

CFO

Blaize

Gopal Hegde

SVP Engineering

SiMa.ai

play_circle_outline Introduction of key industry leaders and their respective roles in promoting AI technology.

play_circle_outline Enhancing Robotics and AI Infrastructure: The Critical Role of Low Latency, Reliability, and Speed in Edge and Data Layers

play_circle_outline Broadcom's focus on connectivity and edge platforms for AI services and applications.

play_circle_outline Blaze AI's programmable silicon aimed at efficiency and low power for edge computing.

play_circle_outline SiMa.ai's chip designed for multimodal processing across various embedded edge verticals.

play_circle_outline Emphasis on a cooperative ecosystem among chip vendors and device manufacturers.

play_circle_outline Robotics' evolution and the shift towards open standards to foster innovation and interoperability.

play_circle_outline Explaining how AI requires optimizing models for efficiency, not just raw computing power.

play_circle_outline Future predictions on societal changes due to AI integration into work and daily life.

play_circle_outline Focus on privacy concerns and the edge's role in controlling data access and security.

Info
Transcript

Vijay Nagarajan, Broadcom & Harminder Sehmi, Blaize & Gopal Hegde, Sima.AI

Vijay Nagarajan

VP, Strategy & Marketing Broadcom

Harminder Sehmi

CFO Blaize

Gopal Hegde

SVP Engineering SiMa.ai

Vijay Nagarajan, vice president of strategy & marketing at Broadcom Inc.; Harminder Sehmi, chief financial officer of Blaize Inc.; and Gopal Hegde, senior vice president of engineering at SiMa Technologies Inc., join theCUBE’s Dave Vellante and John Furrier at theCUBE + NYSE Wired: Robotics & AI Infrastructure Leaders 2025 event. The trio discusses AI’s expanding footprint across compute layers — from centralized cloud to intelligent edge.

The conversation focuses on scaling AI infrastructure efficiently while addressing performance, cost and power c... Read more

explore Keep Exploring

What event is being discussed in the text and who are the participants involved? add

What are the key themes and discussions surrounding AI infrastructure and its development, particularly regarding data centers, distributed computing, and the future of work? add

What is Broadcom's focus and involvement in edge technology? add

What are the key developments and use cases associated with the company's technology? add

What is the focus and capabilities of the six-year-old Silicon Valley startup mentioned in the text? add

What is the significance of connectivity and architectural considerations in the orchestration of systems, particularly in relation to databases, networking, and sovereignty issues in smart cities? add

What is the significance of open standards and frameworks in the field of robotics? add

What are the considerations and future directions for optimizing data center resources in relation to AI use cases? add

What impact is expected from the advancement of AI technology on productivity and job roles? add

What is the significance of edge computing in relation to data privacy and control? add

bolt Powered by CUBE AI

Vijay Nagarajan, Broadcom & Harminder Sehmi, Blaize & Gopal Hegde, Sima.AI

search

>> Welcome back and we're at theCUBE here in Palo Alto. I'm John Furrier, host of theCUBE with Dave Vellante. We're here for a three-day series on robotics and AI infrastructure leaders. We're chroniclizing all the action going on at the infrastructure and also at the data layer to enable this next wave of generative AI. We're seeing robotics, we're seeing agents. A lot of disruption in a very positive way. A lot of enablement. Got a great panel of experts here. Vijay is the Vice President of marketing and Strategy at Broadcom. Vijay, great to see you again. Thanks for coming on.

Vijay Nagarajan

>> Thank you, John, thank you, Dave for having me here.

>> Harminder is the CFO of Blaze AI. Thanks for coming on. Appreciate it.

Harminder Sehmi

>> Thank you .

>> And we have Gopal, the SVP of engineering, operations, SiMa.ai. Great to have you on. Good to see you.

Gopal Hegde

>> Thank you so much and .

>> Guys, you get a tough panel here, because you got a lot of expertise and one of the biggest themes that everyone's talking about is when are we going to hit the end zone, if you will, for the AI infrastructure? Fully unleash the agents? And so a lot of conversations are around how do I build a data center? If you're an enterprise, who do I talk to? How do I create the future of work? How do I build robots that are going to have the intelligence? So it's basically a distributed computing network we're living. It's hybrid and the top conversation is the speed there. Is it reliable, is it safe and can the packets move fast, low latency? So this is the top conversation, so I'd like you guys to introduce yourself, talk about what you're working on and what you're doing with infrastructure. Vijay, we'll start with you.

Vijay Nagarajan

>> Thank you, John. I'm Vijay Nagarajan and from Broadcom. As you all know, Broadcom does everything I guess from the cloud all the way to the edge of the network, so we take pride in connecting everything specific to this panel on the edge. Our focus on the edge has always been around, again, connectivity and also the edge platform, which is our broadband platform. So we believe that there is an edge play when it comes to your home, where you have broadband gateways that are serviced by the carriers and service providers across the globe. If that has an ML engine and it can actually service LLMs and can support use cases, that's a great way to offer new AI services to .

>> And the networking chips, you guys just announced Tomahawk 6, which is really hitting the performance there, so the chips, thanks for what you do. Appreciate Broadcom.

Vijay Nagarajan

>> Thank you. So that's part of it. So there's the infrastructure chips including our networking on the ethernet side, but also, I think the edge is a very important play.

>> Yeah. Yeah. Robot's a big edge. Harminder, what are you working on? What's cool there? You're the CFO. You look at the numbers. You see the business growth. What's going on in your world?

Harminder Sehmi

>> So I'm signing in for my CEO, as he's traveling. I want to start with the company was founded by four very smart engineers that worked on the GPUs at Intel 20 plus years ago. And what they've developed is, you alluded to it, a purpose-built piece of silicon and in our case, it's programmable. So when you look at what's happening at the edge and what the requirements of the edge are, which is low power, low latency, high efficiency. And if you think about the autonomous driving as a use case, the demands for that are very specific and we feel that bringing that, automotive is towards the end of the decade for us. But in the near term, we've got use cases that are in smart cities, smart retail, industrial automation, defense and we think that the Blaze architecture, which is hardware, software, allows this data that's being generated, a million times more data at the edge, to actually be harnessed and people to take advantage of it.

>> And the multimodal is super hot.

Harminder Sehmi

>> Yes. Exactly.

>> Computer vision.

Harminder Sehmi

>> Yes.

>> Gopal, you guys are doing a lot of work. Share what you're doing and what you're working on, what's cool.

Gopal Hegde

>> Thank you. First of all, thank you for having me on this panel. It's great to be here with the esteemed industry leaders. So we are a six-year-old Silicon Valley startup company and we are focused on what we call a embedded edge market. And as our CEO says, people actually live at the edge. And for us, robotics, as well as gen AI are the key areas that we are actually working on. Our second generation chip, which is in the market, which is this chip here, okay, this is built using TSMC six nanometer process node technology and this is fully software compatible with our first generation product that was mainly CNNs. So this one is multimodal, so we can support voice, text, video processing, natural language processing, all on a single piece of silicon and the applications that we go into cover 12 different embedded edge verticals. Security surveillance, aerospace and defense, agriculture, technology, medical and so on and so on.

>> And that's in the devices, the edge device, that chip?

Gopal Hegde

>> Correct. So what we built here in this chip is what we call a system on a chip, so this is machine learning system on a chip architecture. So if you look at the applications that are at the edge, they're not just about the ML models, so in the cloud, you're mostly looking at models and accelerating models. At the edge, you're actually running end-to-end application, so we actually take the images or voice or text coming from a sensor, whether it is a camera or a speaker or a radar or a lidar and we process it all the way to actually displaying it on the screen or providing a response back to you, so you can actually run your end-to-end applications. And our big differentiation is really it's very purpose-built and at the edge. As you rightly pointed out, power is of paramount importance. Robots, for example, are battery powered and you don't want to keep charging them, so you want to run your as long as possible. Drones, for example, need to be running as long as possible, so power is very important. Performance is very important, so we provide significantly higher performance. Eight cameras. 16 cameras. So for example, in a security surveillance retail application, we are able to support 30, 40, 80 cameras and able to detect different kinds of things with our product.

Dave Vellante

>> So low power in real-time essentially. And you all partner together? You cooperate, you compete in some places?

Harminder Sehmi

>> Well, I know we use one of his parts in our quad PCI and I had a chat with him last week.

>> He was on theCUBE. All right.

Gopal Hegde

>> .

>> Vijay, how's the schedule ?

Dave Vellante

>> It's an interesting market.

Harminder Sehmi

>> But like the SiMa, we have a similar approach, similar kind of use cases, that we're probably solving them in a slightly different way. Our belief is that an ecosystem of ISV partners is crucial, because for us to scale, we need to have people that have applications that may be running on a GPU and maybe perhaps not running as efficiently as it could. That's how we scale.

>> Yeah. We're going to talk about the ecosystem, because the heterogeneous environment we all know is where it's going. There is some homogeneous use cases, self-contained stuff we're seeing out there, so want to put a pin in that. But I want to come back to some of the commentary, you guys, on the intro. I love the edge, because remember the intelligent edge, Dave? It's like, man, that was seven years ago. Intelligent's going be hot. It was basically if you got power and connectivity, you can do something. Okay. So that's it.

Gopal Hegde

>> Absolutely.

>> Now, you're talking about data acquisition, one, using data applications at the edge, but you're also getting data in. So if you've got these devices, you need the connectivity and the power and everything's bounded by power on the compute side and the devices. So now you have data coming in, but you're trying to use data for gen AI and then you've got fresh data coming in that has to be processed on the device. So there's going to be a huge demand, we think, of high performance, small form factor devices, because when you're ingesting computer vision, you don't send it across the network. Maybe you will, but that's just more latency.

Harminder Sehmi

>> So I have a perfect example for that. We're working now with smart city customers. And so if you look at smart city, you've got cameras that are on lampposts, so our chip is industrial grade, so it can already withstand temperatures and vibrations. And you can hook up five, six cameras to one of those edge boxes that's sitting on a lamppost. Now, you get 100 or 1,000 of those coming into a server. The use cases are really good, because you're saving money for a customer by the cameras observing no activity happening at night and you lower the lights and that's a saving. As soon as anything's detected, the lights go up, the frame rate goes up and you do object detection, you send an alert to whoever needs to go. All of that is happening at the edge box.

Dave Vellante

>> You see that all the time , so you go, "Whoa. What just happened?" And that .

Harminder Sehmi

>> Right. But this is now we are working with customers in South America and in Asia-Pac regions and that's been our focus. One of the things that Blaze technology allows you to do is a lot of things for a lot of people at the edge. And so our focus right now is just smart city where we know we're making dumb cameras smarter and then some of the defense applications, we talked about drones, for example. Sadly, the Ukraine War has just told us how you can disable a drone, but also how a drone can be more intelligent, because it's got onboard compute.

Dave Vellante

>> How are people-

Vijay Nagarajan

>> Yeah.

Dave Vellante

>> Go ahead, Vijay.

Vijay Nagarajan

>> And John, you talked about ecosystem and you also asked if we are competing or collaborating. I think the right word I'd like to use is complementary. I think chip set vendors and the types of devices that we work with is very, very complementary. Case in point being what Harminder mentioned. So you have their, say, two edge devices right at the edge and the question comes up on where does the processing even happen? Does it happen right at the edge? Do you have to go back to, say, one level in to another device? So I think having different categories of devices that are actually interacting with each other through connectivity is extremely .

>> Connectivity is important in the orchestration. In a way, an operating system has to be built. So to your point about ecosystem, we used to say it's in the database world and there's no one database to rule the world. Now, you see multiple databases. Every application has a database, but now you have a horizontal plane managing it. When you look at networking, how do you guys see the architecture? Because we haven't even thrown in with smart cities, is things like sovereignty. You got to design the intelligence that has to be at the edge, but also in a data center somewhere, managing policies. I'm oversimplifying it, but like I say, if I'm in France and I want to keep that traffic in the network, that's a sovereignty issue. It's a sovereign cloud challenge or on country.

Harminder Sehmi

>> Yeah. On the cloud.

Gopal Hegde

>> Yeah. So I think typically, all these devices had to have to be provisioned, managed, upgraded, et cetera. So if you think about for the ML at the edge, there is MLOps, which is essentially your life cycle for your ML application, so the edge applications get smarter. For example, you detect certain things in the smart city use case that Herminder talked about. You may be detecting certain types of people or certain types of threats or certain types of events, so how do you keep updating those? So for that, there is a process which actually collects all this information, sends it to the cloud and continuously runs a process in the cloud to find out whether anything new needs to be added. And when something new needs to be added, that model gets trained and then it gets deployed back to the device, so networking becomes very, very important, connectivity becomes very, very important, cloud to edge continuum becomes very, very important. So in that sense, the connectivity that is provided is very, very useful and it's not just about that. So if you think about, for example, gen AI, how do you make these models smarter? If you want a model to be customized for your application, for example, if you're working with a car manufacturer and you want to take a Llama model and you want to make that model specific to that car manufacturer, you can actually feed a car user manual, so the process is called RAG, retrieval augmented generation. So what it does is it reads that document and trains that model with that new information. Now, how do you do that? This thing happens in the cloud and the data then gets sent to the edge device and the edge device becomes smarter. So this is a way to keep your edge device up to date, keep your edge device current and also manage a bunch of edge devices that are actually deployed across your assembly line, across your smart cities, smart factory, smart transportation kind of environment.

Dave Vellante

>> So typically, we would think of edge applications from a persistent standpoint. As a lot of the data is ephemeral, you only send certain data back to the cloud. Take automotive and a deer runs in front of the Tesla. Oh, that's an event. We'll send it back. Video maybe is different, but has machine learning and AI changed that, where that in between, that intermittent data is more valuable, or is it still you're only persisting a small portion of the data?

Gopal Hegde

>> So the challenge with the benefit of the edge is that, so if you want to send all that data back to the cloud, then you have latency problems.

Dave Vellante

>> Yeah. Latency. It's expensive. Okay.

Gopal Hegde

>> So it's very expensive and you may not even have the connectivity. So we are working with customers. For example, so the train track inspection for example. So they use drones now And they inspect the train tracks and they send that information back to the data center. And where people are manually looking over those to find out whether there are any problems or they feed the videos to data center servers. This whole process takes a long time. What if you put a device on the train itself? And while one train is running, it actually detects these things or the drone itself detects some cracks in the track and then informs the data center and then you can actually control. So what you are sending back is not the actual video. You're sending just the event. So the amount of data that goes back is actually very, very simple, so we have an application where our product is being used today. So they have a system where along high speed freight tracks, this is in India, they actually have deployed cameras and they actually look at a freight train as it goes by and they look at is the wheels on the train flat? Is the door on the wagon open? So this information is analyzed and the event is sent to the control center, which is at one of these train junctions and the operator can now make the decision, should I slow down the train? Should I stop the train and fix the problem? So instead of sending the entire video back and having manual processing or cloud processing, which is very expensive and is latency prone, what you're doing is you're just sending that one event.

Dave Vellante

>> Yeah. And we're seeing robots climb bridges and looking for anomalies and so you're just sending back that event.

Gopal Hegde

>> Exactly.

Dave Vellante

>> But you still need good networking and you want it to be high speed, low latency and low cost.

Gopal Hegde

>> Of course. Of course. Of course.

Vijay Nagarajan

>> You hit it. And then there's also, from an edge perspective or a edge to cloud perspective, you're talking about privacy as well. So you have all these vectors that determine where the data gets processed, where it is getting computed. As Gopal said, is it at the edge, is it with the robots, or is it at a localized data center not necessarily all the way back to the cloud? Is it a localized data center, or in the case of a residence, is it at the carrier gateway? So you could have a situation where you could have smart cameras that are detecting events, or you could have cameras that simply transfer the video back to the gateway and the gateway does the processing and transfers the event back to the cloud. So I think it's use case dependent and it's dependent on the variables that you just stated, which is latency, speed, performance, cost and then privacy and security as well. And from a connectivity or a networking perspective, to that end, we have to look at everything starting right from the edge, the devices that Gopal and Harminder talk about, which would require, say, wireless connectivity back to a localized device, local device, or local data center and then back to the cloud through wired networking. So the network architecture has to contemplate every event or everything starting from the edge all the way to the cloud. Includes wireless, wired and all these capabilities.

>> Broadcom, you guys are building the chips. They're smaller, they're faster, they're more energy efficient, reducing the hops on the scale up. we got the scale out. When you look at robotics as well as the edge, you're starting to see use cases in manufacturing, retail, automotive. They're everywhere, where there's a robotic element. Even humanoids, we're seeing that work being done. Covered that here. You're seeing this notion of self-contained unit that needs open standards, but it needs to be highly customized for the performance of the thing, whether it's a robot or a manufacturing line. How do you guys look at that? Because there's been debates right now in the industry. Not debates, but been conjecture, I would say, between homogeneous systems that just work.

Dave Vellante

>> OSP.

Vijay Nagarajan

>> OSP.

Dave Vellante

>> Yeah.

>> And then you got open systems. OSP Broadcom thing.

Vijay Nagarajan

>> .

>> Dave memorized the Charlie Kawwas . Charlie Kawwas, If you're watching, Dave is-

Vijay Nagarajan

>> Dave's never going to forget that.

Dave Vellante

>> Never.

>> He's got it memorized. But also, if you don't have an ecosystem, you can't scale, because there's now interoperability challenges. So there's a trade off, this discussion. When do you hard and top things? When do you keep it open? Or it should be open, obviously, all the time.

Harminder Sehmi

>> Yeah. So a couple of points came up on the previous conversation, which are really important. Privacy and how do you keep your data? So the concept of air gap solutions, you don't actually necessarily have to send. You're collecting a lot of data, you're processing a lot of it, as Gopal said, at the edge, because it's important. That's where you need the result. And then maybe for training purposes and other reasons, the video gets uploaded to the center, but it's still your data. It doesn't go out and to anybody else if you've got facial recognition. You mentioned France or Europe generally. But coming to the robotic type of thing, it's accidental. Samsung was our foundry out of Austin and so we ticked the box of made in America. And therefore, a lot of the defense type of applications for us are open to us. So the box that we had, there's a carrier board which has AI running on it, which can sit on a drone and it makes the drone extremely intelligent versus we've got use cases similar with oil pipeline type of inspection that a lot of the work can be done there and then and what you're just sending is in mile 644 is where there might be a problem as opposed to the video coming back and somebody processing it and then two days later, deciding to go to mile 44. So those are the things that we find that customers are valuing the ability to just take those decisions there and then. .

>> Yeah. But that's a solution though. That's a solution built upon components that are open.

Harminder Sehmi

>> Correct. And it's-

>> So it's not so much that's a heterogeneous chip or whatever.

Harminder Sehmi

>> Yeah. Yeah. Exactly.

Gopal Hegde

>> So to answer your question specifically with regards to robotics, open standards are very important. You cannot have a complete vertical solution for this market and that even the industry leaders are proceeding. So Nvidia has a framework called Isaac. I don't know exactly what it stands for, but it's a framework. A lot of these robots are built on what is called ROS, robotic operating system. And the latest version of ROS is ROS 2. Intrinsic, we work with them. They actually acquired ROS and what they're trying to do is to create a platform which is a software platform for industrial robots. We're working with multiple customers. We recently announced AMR, autonomous mobile robot designing. So what these guys do is they build a common platform that goes into these AMRs and the ROS has multiple components. One component is SLAM. So what SLAM allows the robots to do is to move. And there are different types of SLAM.

>> And it's open source.

Gopal Hegde

>> It's open source. It's completely open source. So it runs on compute, for example, today, so typically, all these run on X86 compute. Now, you can speed it up by actually using ML accelerator, so we are doing that. And there are different types of SLAM. Cartographer SLAM, visual SLAM all are open source. Then on top of that, all these ML models, applications, for example, pose estimation or object detection or tracking, all these different applications which a robot needs can be actually run on SLAM and on ROS for example. So SLAM plus ROS gives you this framework, which is a open source framework. SLAM gives you the control and movement and then the applications which are built using ML give you the ability to run different applications. So that's really what is being developed for robotics. I really like that, because this levels the playing field. Our customers like it, because it doesn't lock them into a particular-

>> It increases innovation too. The entrepreneurial world will jump on this.

Gopal Hegde

>> Absolutely.

>> So you have more technical advances.

Gopal Hegde

>> Absolutely.

>> Interoperability. Vijay, I know this is something close to your heart on open systems. We've had many conversations around open systems, the benefits of it. You guys are making the chips smaller, but the density is still a challenge, because you make the chips smaller, but there's more stuff on stuff, so you got more density but still, chips. Why is open important to you guys?

Vijay Nagarajan

>> For almost the same reasons that Gopal stated, so you want this to be a place where there is a vibrant ecosystem. Open's where you can get a scalable ecosystem, where you need people that are talking to each other, devices that are talking to each other. And for me personally, if I've learned anything from being in Wi-Fi for about 20, 25 years, it's that you need open standards-based technologies that will allow a vibrant ecosystem to thrive. We need interoperability. We need certification around devices that are interoperating with each other with a common protocol and if you have that, then I think everybody wins. A closed walled garden doesn't, but with an open approach, everybody in the ecosystem wins.

Dave Vellante

>> Well, I think we've seen that open drives opportunity.

Vijay Nagarajan

>> And innovation.

Gopal Hegde

>> And innovation.

Vijay Nagarajan

>> Innovation. Absolutely.

Gopal Hegde

>> And innovation.

Dave Vellante

>> And TAM.

Gopal Hegde

>> Yes.

Dave Vellante

>> So you're, today, participating in, let's call it roughly a $4 trillion TAM. And by all estimates, that's going to, let's call it 3X, 4X, maybe even 5X. So I have two questions. One maybe not so controversial. The other one you might not want to answer, but I'm going to ask it anyway. How are we feeling about full self-driving? Harminder, you mentioned it's end of the decade for you guys. What are your thoughts on that?

Harminder Sehmi

>> So by full, you mean level five?

Dave Vellante

>> Yep.

Harminder Sehmi

>> From what I hear the experts talk about, there's your camera technology has to improve. There are all sorts of other things that need to improve. I certainly think L3 plus L4 is possible. We've been working with DENSO. Has been an investor and a customer of ours for seven years and actually, that relationship was super important for us, because they validated the architecture that we built, because the vehicle is the most discerning object which is moving that needs to make decisions. When you get to we did work with another German OEM around L4 type of applications, more a chiplet type approach. So we definitely see that if you are not Tesla and you don't have your own chip and you want to have an alternative to Mobileye or another GPU-based sort of solution, then Blaze is a very, very important partner for you. And we expect to accelerate the work with some of these OEMs and tier ones. And when I say that revenue for us comes at the end of the decade, it takes five, six years for some of these systems to get in place.

Dave Vellante

>> So part of the catalyst for Tesla doing its own design, not only did it want to customize the NPU, but there was issues with Mobileye and who controls the data. I presume that you're open about that.

Harminder Sehmi

>> And here's what all the OEMs and tier ones that we worked with said to us. "Please go sell your chip to every other OEM." And why would they say that? Because the differentiation is not the software layer. I heard a Ford CEO guy talk about a year ago, but there's 100 or so, 150 ECU microprocessors. Okay. That's one problem, but they're from maybe half a dozen manufacturers. Think of all the different software challenges that they have. So talking about open standards is really important, because based on a Blaze architecture, which is programmable, so our SDK is open, we have a low-code, no-code AI studio platform for those people that can't code, that they can deploy AI. So when you put all of this together, so for us, the autonomous opportunity is huge and that's a hockey stick that happens towards the end of the decade for us.

Dave Vellante

>> Second question. You mentioned Samsung, TSM six nanometer. You guys, of course, rely on TSM and you even have your own foundry, I think, for thin film, because of some of the constraints in the marketplace. So how are we feeling about foundry, foundry capacity? There's obviously geopolitical tensions. How important is it to have a foundry based in the United States?

Harminder Sehmi

>> Maybe I'll go. So we've worked with both, TSMC and Samsung. Obviously, for defense, having something in the US is important. We think we don't necessarily need to be at the cutting edge process and node all the time, just because our architecture uses those TOPS in a much more efficient way. So all the foundries that are coming up in Arizona, we would be customers for those.

Dave Vellante

>> Not cutting edge, but near cutting edge. And the-

Harminder Sehmi

>> Near cutting. We don't have to be at two nanometer, for example.

Dave Vellante

>> You guys do in some cases.

Vijay Nagarajan

>> We guys do. We guys do. In some of our three nanometers, two nanometers things that we could look at. See, from our vantage, again, I go back to this thing that a vibrant ecosystem is important and that goes to supply chain as well. From our vantage, yes, we do a lot of our merchant silicon with TSMC, but at the same time, we proudly manufacture a lot of our, say, RF filters here in the US in our fabs. So it's something that we believe in overall, that the vibrant ecosystem needs to exist and supply chain is no exception.

Dave Vellante

>> Mm-hmm.

Gopal Hegde

>> Yeah. I agree. The challenge, I think, for the US in general, speaking about TSMC, and this is public information, that the US cost is higher relative to the cost they have in Taiwan, but they are doing it here as well. So working with someone like TSMC gives us options. If you really want to manufacture in the US, you can in the Arizona fab or you can go to Taiwan. And as Harminder said, we are not at the bleeding edge. By the way, as a small company, we can't even have further, five or a three nanometer, so I'm at six nanometer and six nanometer is actually decent enough for us for the product we deliver. Our architecture is all digital and it is quite efficient from a power and performance standpoint, so that technology can be available here as well. So in that sense, having a ecosystem is important. Having a choice is important. I'm very happy that Samsung has a lot of US presence. Intel is building fabs here and the government is investing, because having options, having a ecosystem for companies like fabless semiconductors like ourselves is very, very critical.

Dave Vellante

>> But could you go to GlobalFoundry or no? Because .

Gopal Hegde

>> I have used Global in the past. So at Cavium, my previous company, we were a big customer of Global. Only reason why we are not using Global is that below 10 nanometer, they're not as good in terms of they have stayed at a higher technology node. And of course, TSMC is world-class.

Dave Vellante

>> But the yields aren't there.

Gopal Hegde

>> Yields and things like that.

Dave Vellante

>> I've been outspoken on this issue. I'm not comfortable with where we are in terms of the foundries in the US. I was hoping that Intel would actually divest for the foundry, but that's not going to happen. And so my feeling is they're going to become maybe a little bit more advanced than GlobalFoundry, but not to the TSM level and TSM, unless they bring expertise here, I see that as problematic, because what we're seeing now with the tensions for things like rare-earth metals and there's some issues that I think are maybe not blind spots, but risks that need to be considered.

Gopal Hegde

>> I agree. I agree I think, but then here is one data point for you. Other than the high end processors or the GPUs, vast majority of semiconductors are at a higher node. So to that extent, having capacity and ability to build those in the US is very, very critical. But do you need a two nanometer process node in the US? Maybe you do for certain products, but not for every-

Dave Vellante

>> Okay. So I agree with that.

Vijay Nagarajan

>> .

Dave Vellante

>> So that level of comfort exists that the US market .

Gopal Hegde

>> Correct.

>> Yeah. Yeah. The market's stable there, because you have use cases that require massive, smaller, bigger throughput.

Gopal Hegde

>> Correct.

>> I guess the question here during this event we're having is robotics has a surge of, I won't say democratization. It's not the right word, but the builders doing robotics and edge is getting very device and specific robotics is booming. Now, there's a huge surge and then infrastructure leaders have to now serve this different edge or more robust edge, whatever you want to call it. So the question is what's the architectural design principles that people should be thinking about? Entrepreneurs, partners, we believe the ecosystem is going to be booming, because once the computing is connected with the power, with the batteries and some of the energy stuff going on, there'll be a surge of innovation that'll be built, enabled by you guys. So what's the playbook?

Gopal Hegde

>> So this is a very interesting question, so I'll tell you this. So the entire world is changing very, very rapidly. So all of us are familiar with DeepSeek, so people actually are very creative, world over. It's not specific to a country or whatever. They're very creative. They get the best out of the technology they have. So I'll give you an example. So for our next generation product, we're looking at what are the requirements for something like a humanoid robot? And my architect came back and told me, "Hey, I need to have this 400 TOPS. I need this requirement, because the large language models are exploding this." I'm talking six months ago. So I said, "Okay. Don't think what is happening today. Think about what is going to happen in five years from now, three years from now, four years from now." Because that's when the chips come out. What is happening is that to make the robots work, to your point, the power is very important, the cost is very important, the area is very important. So essentially, we are going from large language models to small language models. The VLMs and VLAs are actually getting more and more efficient and they're able to do things with less parameters. So what we are finding is that, I'll give you our own example, we are finding that six nanometer technology, we can actually address humanoid robots. It requires software changes. It requires architectural changes that we need to do to our chip, but do I need to go to a two nanometer for that? So a combination of hardware and software becomes the solution. It's not just hardware alone and one of the things we have done at SiMa all along is, to the point Harminder made as well, okay, don't worry about the TOPS, so 400 TOPS doesn't mean anything. What is the metric you care about? Let's say in a computer vision application, if you care about frames per second, can I give you the same frames per second with less TOPS? So a lot of the architectures that are out there are not very efficient. So if you look at a large language model today, they use only a very small percentage of the compute available to them. Now, how do I increase that percentage and how do I make my compute more effective? Through software, gives you the ability to actually not necessarily go and build these high end, high power, high cost chips, but stay at a lower level, but solve the problem with software and I'm very happy to see tremendous innovation happening in this field. So the physical intelligence model, the GROOT model, just so that we are running all those on these chips today.

>> what you're saying is the constraints are critical to identify, so that the software optimization-

Gopal Hegde

>> Exactly....

>> the creative DeepSeek moment, because they basically discovered it by accident, we know what the real story there, is that they didn't really do it for that reason. They were trying to get around the constraint, was they don't have the big GPUs.

Gopal Hegde

>> Correct. Correct.

>> Okay. So what's the position of the constraint? Is it compute, power? What should people be innovating around from a software perspective?

Harminder Sehmi

>> Yeah. So Dannecker comes up. He uses this analogy and it's so simple. When DeepSeek came up, what it demonstrated was that you don't have to wake up the whole chip or the whole model. And his analogy is, hey, somebody says, "What's two plus three?" And there's a math teacher on the model. There's a lawyer. It might be a dental hygienist, whoever. You've got all these people and suddenly, a question comes in and you've woken up everybody. Think about the power that takes. And so what DeepSeek showed was mixture of experts. Hey, this is a math question.

Dave Vellante

>> That's the second-grader.

Harminder Sehmi

>> Right. Now, here's where the Blaze technology is really relevant, because we're able to run multiple things on our chip at the same time, which need not be related to each other. And therefore, if you were to run a DeepSeek type, and the point you're making is very relevant, that all these models are becoming smaller and the accuracy is still pretty high. So running those kind of models, we think on the common architecture that's going to be throughout our roadmap is going to enable these language models to be very efficiently run. And with very low power, because you're not waking up the whole chip.

>> So you're basically saying, and I love that example of waking up only the math teacher, the sub or second-grader as Dave said, but really getting into the fact that in the old school infrastructure days, overprovisioning or over-building was a concept.

Harminder Sehmi

>> Right. Yeah.

>> You're essentially saying that you don't have to over-rotate or-

Harminder Sehmi

>> It becomes modular.

>> So this speaks to some of the things you guys have been doing with the chips. You can think about the data center not as give me the big fat rack and here's all my megawatts for the math question. You don't need-

Vijay Nagarajan

>> Yes. Yeah.

>> Yeah.

Vijay Nagarajan

>> It's not so much the overprovisioning. I think it's the other way around right now with AI, because, okay, you have some constraints and to what Gopal said, look forward to what use cases you're trying to solve three, four, five years hence and the landscape's going to be very different. The models are evolving continuously, becoming more and more effective, so we have to start looking at other parameters to go optimize on and specifically focus on use cases. You talked about 400 TOPS. The types of use cases in our gateways, say, demand less than five TOPS right now, so it's really a function of the use cases. It's not a one-size-fits-all, especially at the edge.

>> Utilization. It's utilization.

Gopal Hegde

>> Exactly. And that I think is, like you rightly said, how do you actually get your product customized for the use cases? And going back to robotics, robotics is a very interesting use case. It's very complicated, super complicated, but at the same time, a robot is doing one thing at a time and sometimes it may be doing a couple of things at a time, so how do you actually build a chip which is cost optimized, power optimized? Because if you basically throw the hammer at it and try to solve that problem with lots of hardware, you're not going to have power efficiency, you're not going to have the cost structure. And if you run very complicated, very large models, to the point Harminder was making, why? Why do you want all those other things that are not required there? So things are becoming much more specialized, things are becoming much more optimized and the software innovations are really key here and we see this every day. And to the other point that Harminder made as well, this open source ecosystem of these models allows companies like us to innovate. So Meta opening up a Llama and making it available was a phenomenal moment. It basically made Google come out and put Gemma out there and-

Harminder Sehmi

>> ....

Gopal Hegde

>> Gemma out there and all these guys. DeepSeek happened and DeepSeek would have never happened if there was no open source model. So this drives innovation and the beauty of this is that people are driven by constraints And they say, "Okay. What do I need to do? I need this robot to be cost-effective." So the humanoid industrial robots, $175,000. How many people can afford that? It needs to be $1,500. To make that $1,500, you cannot use a $500 chip in there. You need a 50, $100 chip. That's your constraint. What can you get in 50, $100?

Dave Vellante

>> Okay. So I-

Vijay Nagarajan

>> Or a different business model.

Gopal Hegde

>> There you go. Yeah.

Dave Vellante

>> Yeah.

Vijay Nagarajan

>> Right. So innovation.

Dave Vellante

>> I was going to ask how far away are we from general purpose humanoid robots and what does the price point have to be? You're saying it has to be under five figures.

Gopal Hegde

>> Yeah.

Dave Vellante

>> How far away are we from that?

Gopal Hegde

>> We are quite far away.

Dave Vellante

>> Like a decade plus?

Gopal Hegde

>> I don't think it is decade plus, but it is not two years either.

Dave Vellante

>> Maybe-

Gopal Hegde

>> Maybe five years.

Dave Vellante

>> Okay. But perhaps this decade, but certainly into next decade, you'll see that.

Gopal Hegde

>> Absolutely. Absolutely. And these robots are actually getting to be quite intelligent. We're building one and we're actually using this device and it's a humanoid robot that actually moves around and it can actually do simple things like pick up a object, all that. We are actually going to build this with four of these chips. And it's a off-the-shelf thing that we plug these things in using these carrier cards, which are and really-

Dave Vellante

>> Yeah. What is that ?

Gopal Hegde

>> Okay. So this is a Nvidia carrier card. So Nvidia has a very broad ecosystem. They've created a really good ecosystem. People build these carrier cards and these carrier cards then get plugged into various types of industrial robots or factory automation equipment, et cetera. So what we've built is we built a module. This is called a SOM module, system on a module, which is a RMNX form factor. So I can plug this in into a carrier card built for Nvidia, so what this allows us to do is to go into all those platforms which are already deployed in the field, in the market with this. So what we are trying to do is to actually take advantage of the ecosystem that is already out there and work with all these partners and that actually allows us a good way to go into market. And back to your robotics problem, how do you build a cost-effective, power-efficient robot? And if I want a robot which is less than five figures, I need to do it in one of these chips, not five chips and-

>> Yeah.

Dave Vellante

>> 1,500, I'm in.

>> So basically, what you're getting at is mind set, because if you throw hardware at the problem, you're going to have a big, fat drone, a big, fat robot versus I've interviewed an entrepreneur on this show where he's like, "Actually, we're going smaller," because his use case is for a small drone. So his now optimization, what he's designing around is form factor, so he's actually working backwards from his goal and now engineering the solution from the constraint.

Gopal Hegde

>> Right. Right. And talking about drones, we're working with one of the drone companies and it's very sad, but the most popular thing they have going for them is the kamikaze drones, so they go one-time drones. They can't afford a very high cost compute system, but then how do you do that? You need a smaller chip, you need a lower power chip, you need a power efficient architecture and basically, the innovation is all in software.

>> Well, we didn't have enough time to talk about the future of, say, work for instance is a topic that comes up a lot on the show, because we love robots and love to get in the weeds on some of the infrastructure, but at the end of the day, the other edge is our life. So I guess the final question is what's the future of society, work? Because now, digital and physical are blending, so now, my relationship with my work, my play, society is face to face. We had a big party last night and everyone goes online. We're doing the digital event here. We're blending. So do I walk into an office space? AI knows what my desk looked like before someone else used it. So what's the future of play, work? Because it's got wireless, it's got AI. It's going to be some sort of edge stuff. Thoughts? We'll start with Vijay. We'll end it on the panel on this question.

Vijay Nagarajan

>> It's a broad question, John.

>> Okay. Go to town. .

Vijay Nagarajan

>> It touches everything from work to life. And as an employee at Broadcom, I see the blending happen gradually and seamlessly. I think we have to increasingly incorporate AI into the scheme of things and at some point in time, the nature of the workforce is going to change to be more AI educated and to be able to compliment what AI can do. So essentially, they got to learn new things and change. That's how I see that happen. But on the flip side, and I've talked to you guys about this in the past, at home, I see tremendous change in terms of how we function on a daily basis. I, for one, believe that the future generation is going to be very, very AI-first, so I'm the first person to go enable my twelve-year-old and eight-year-old .

>> He's a prompt engineer now.

Vijay Nagarajan

>> They are. They absolutely are prompt engineers. So I think they are going to be an AI-first world and so that's the skill that we have to teach them. And so the case in point being whether it's at work or at home, we have to facilitate the preparation. We have to facilitate the education for them, for people to participate in that AI-first future. That's my view.

>> Yeah. And get better chips too on the capacitor.

Vijay Nagarajan

>> Yeah. Yeah.

>> Harminder, what's your take on the future of society, life, work?

Harminder Sehmi

>> Yeah. So at the risk of repeating, I'll just add a couple of points to what Vijay has said. I think privacy is really important, so how we deal with that. Clearly, there are huge benefits still to be reaped from just deploying AI if you think about agtech, if you think about healthcare. But of course, then there is yesterday, one of the guests talked about huge disruption too and is it going to take 10 years to work through that disruption? Is it going to take a generation? We've got to solve those kind of problems. But the point is we know it's here and it's up to the innovation will help us work through it, I think.

>> Gopal, close us out.

Gopal Hegde

>> Yeah. So I think if you look at the waves of technology that has come about over the last 50 years, every wave has made us more efficient, more productive. So what I see AI doing is exactly the same. It is going to make our life more productive. It will take away a lot of the routine things that we are doing. Yes, there may be disruption in terms of really routine jobs now being done by AI, but at the same time, what people have done consistently is they've moved on to higher value jobs and higher value things and use the technology for our benefit. So I see us becoming more productive and basically doing more interesting, more fulfilling things versus doing routine things and the AI taking care of a lot of the routine work. And going back to the point Herminder made about privacy, okay, that's where edge is important. Edge is very important, because if you send all your data to internet and to the cloud, you have no idea what is going to happen there. However, edge actually allows you to keep the data local and what that allows you to do is to control who gets access to that data, because in this new economy, at Marvell, our CEO used to say data is the new oil, so whoever has the most data wins. So that privacy can be addressed at the edge. So I see world changing in a better way and to the point which I made, it's actually very interesting. So we hire a lot of new school graduates now and they are the ones who know AI like back of their hand. So they're the AI experts, so I learn a lot from them and that is what is amazing. So they are actually becoming very good users of how to take AI into places where it hasn't been.

>> Yeah. And this is the shift. It's a generational shift. How data centers are built. How networking is going to work. How utility will be extracted. Gentlemen, thank you so much. I'm really grateful for you coming on.

Dave Vellante

>> Thank you.

Gopal Hegde

>> Thank you so much.

Harminder Sehmi

>> .

Gopal Hegde

>> Thanks for coming on. We're a little bit over. Great mixture of experts here with theCUBE and the NYSE here in our Palo Alto studios with theCUBE. I'm John Furrier with Dave Vellante. Thanks for watching.