Altera on Why Physical AI Needs More Than GPUs

AI models are changing on weekly and monthly cycles, while the connectivity silicon linking compute, memory, and storage takes 18-24 months to design. That mismatch puts pressure on data movement, the layer that has to adapt each time a new model changes how compute and memory are combined.

Patrick Moorhead, Founder, CEO, and Chief Analyst at Moor Insights & Strategy, is joined by Raghib Hussain, President and CEO of Altera, to examine where FPGAs fit in AI infrastructure and physical AI systems. Hussain also covers what customers building intelligent systems asked for ahead of Altera's Innovators Day launch of next-generation FPGAs and agentic AI tools, and how Altera's independence shapes its work with NVIDIA, Arm, and Intel.

For system architects, this implies a split in design decisions: which parts of a system stay specialized for efficiency, and which stay programmable so they can adapt as models and sensors change. In the data center, FPGAs sit alongside CPUs, GPUs, and custom ASICs, bridging the gap while waiting for standard or custom connectivity silicon. At the edge, physical AI systems rely on FPGAs to synchronize sensor data and act with low latency and deterministic behavior.

🔹 Why model churn is outpacing connectivity silicon design cycles. Models evolve weekly or monthly and keep changing how compute and memory are combined, while connectivity products take at least 18-24 months from start to finish. Programmable fabric lets infrastructure OEMs adapt data movement after deployment while they wait for standard or custom silicon.

🔹 How FPGAs work alongside CPUs, GPUs, and custom ASICs. Hussain positions the FPGA as a programmable connectivity companion to those chips, not a competitor. System architects decide which parts of a design stay specialized for efficiency and which stay flexible as data movement needs change.

🔹 Why physical AI needs deterministic, low-latency sensor processing. Physical AI systems must collect, format, and synchronize data from many sensors, then trigger actions with low latency and deterministic behavior. FPGAs deliver hardware-level latency while keeping security and safety logic updatable.

🔹 How independence shapes Altera's partnerships with NVIDIA, Arm, and Intel. As a pure-play FPGA company, Altera works with NVIDIA on reference designs that map a range of GPUs to the right-sized FPGA, with Arm through SoC FPGAs that include multi-core Arm, and with Intel on data center motherboard management. Hussain says Altera sells only FPGAs and has no other agenda.

🔹 What agentic design tools change for FPGA engineering teams. Quartus IQ gives engineers project-level awareness in an agentic flow and compresses time to market, and it is being brought to NVIDIA DGX Station so design data stays on premises. Integration into the NVIDIA CUDA developer environment is intended to let physical AI engineers focus on the application layer instead of rewriting drivers when a sensor changes.

Keeping data movement and sensor interfaces programmable lets infrastructure and robotics teams adapt to new models and sensors without waiting for a new silicon generation.

Watch the full video at sixfivemedia.com, and subscribe to our YouTube channel so you never miss an episode.

Learn more about Altera at altera.com.

Disclaimer: Six Five Media is for information and entertainment purposes only. Over the course of this video, we may discuss companies that are publicly traded, and we may reference their equity share prices. Nothing discussed during this webcast should be considered investment advice or a recommendation to buy or sell any security. We are not investment advisors, and you should not rely on this content as financial advice. Six Five Media collaborates with technology companies and industry leaders to produce research-driven interviews and multimedia programming for enterprise technology audiences.

Raghib Hussain:

This is the moment which I believe FPGAs are built for. We have the tools to enable our customer, solve their problem and create value for them and that's how we all bring AI to the reality as fast as possible.

‍

Patrick Moorhead:
Hey, everybody. Welcome to the Six Five Virtual Webcast. I'm Patrick Moorhead, and today we are talking about something that I love to talk about, and that is how FPGAs fit into AI, deeper into the data center infrastructure, and into the physical world. I am joined by Rageb Hussain, President and CEO of Altera. Rageb, welcome back to The Six Five, and it's great to see you since we met up at the GSA Executive Forum.

‍

Raghib Hussain:
Thank you. Thank you, Patrick, to give me this opportunity. Always nice to meet you.

‍

Patrick Moorhead:
Yeah, it was really great getting to know you as your previous employer, where we worked on a lot of great stuff, and you really made a huge impact there. But it's great to see you in the seat here running Altera. I can't believe that it's been since May 2025. It really seems like yesterday. And isn't it crazy how much AI has changed just since then? I mean, we were literally fixated on models and chatbots. And since then, we are into agents, agent swarms, It's amazing, but also AI reshaped demand across every corner of the semiconductor market as well to support these evolving use cases. Can you talk a little bit about what's behind this increased demand for FPGAs as you're seeing AI scale and shape in different ways?

‍

Raghib Hussain:
Yeah, definitely. So first of all, it has been an exciting, I would say about 18 months. As you know, I joined Altera at the time of announcement itself. So we separated in September, but I joined in May. So it has been exciting. On every front, things are evolving. Markets are moving fast. Demands and needs are moving fast. Specific to the FPGA use case and AI, AI actually is changing the pace at which things need to be adapted. So just to give you an idea, when you look at all these models, they are evolving at the weekly, monthly basis. And if you look at under the hood of that, what is happening really, You are all these models that put that compute and the story which is memory and storage in different combination and permutation to get the most out efficiency out of these model. In other words, the connectivity. between how compute talk to other compute and how compute talk to the memory is changing at a fast pace. And I think that makes the FPGA as the core of this getting adaption to this flexibility. Because by nature, FPGAs are very flexible. You can make the FPGA programmable in the field after deployment. And if you need to evolve your data movement or your connectivity, After deployment to match with the refresh cycles of the of the, you know, compute. This is the best way to do that. So if you look at the data center FPGA has become pretty much, you know, programmable. connectivity companion for the CPU and GPU. I just want to make sure that we are actually companion of the GPU-CPU. We are not kind of competing with them. In fact, we are also companion of these custom Asics, because of the fact that the design cycle of these connectivity products is at least 18 months or 24 months, depending from start to finish. In the meantime, if you come up with a model which you can take advantage of moving data movement in a custom manner or bringing certain compute close to other and so on. You need flexibility in the fabric, and that flexibility is brought by the FPGA. The value driven by the time to market is so high that infrastructure OEMs, they want to take advantage of that flexibility, while they are waiting for the standard product or their custom silicon to come out. So in short, in data center, FPGA has become a companion and goes side by side with the GPU, CPU, and the ASIC. And when you turn around on the on the edge side, it is even more interesting because all these, you know, in a physically I world all these sensors that that needs to be brought together. And you know data needs to be collected data needs to be formatted in a unified form it needs to be synchronized. So all those connectivity between the sensor and step function action, FPGA is the best solution for that. And on top of that, there are known actions that require the low latency and deterministic behavior, and really the security and safety implementation in a very low latency manner. And this is where FPGA brings the hardware type of performance latency along with the flexibility that we can change to address the needs of the security or address the needs of those known action that needs to be done.

‍

Patrick Moorhead:
Now, that's really helpful because quite frankly, the conversation has been dominated by GPUs and accelerators. It's kind of moved to, well, definitely with the Gentic, a lot of talk about CPUs, then obviously the network. It seems like every new conversation is all about some sort of a blockade or a challenge that people are having. And it really sounds like FPGAs work alongside sensor data to networking and data movement inside of the data center as real companion chips to GPUs and accelerators, and not only an AI infrastructure, but also physical AI systems. I'm blown away when I look at a diagram that shows how many FPGAs are inside a biped robot, or even inside of a rack mount systems. And I can just imagine the new fungible unit of agentic token machines is now a pod that has seven racks inside of it. It just must add up and it's a huge business opportunity. for you and also the ability for you to really solve some real problems that your customers are experiencing right now in a lot of unique ways. You're a very different company now, right? You're an independent company. Silver Lake is the majority owner and Intel has a remaining stake, but you're working very closely with companies like NVIDIA, and Arm, who their platforms sit in the center of many of these AI systems. And I'm curious, how is this independence of the company, how does it shape the way you partner across the ecosystem now?

‍

Raghib Hussain:
So yeah, so that's very interesting question. But before I go there, I want to explain that FPGA actually has become the critical function in the entire heterogeneous computer. So it's a critical companion for the GPU and CPU for the reason that I described earlier. It gives the flexibility. And then the designers actually, a system architect needs to decide that, hey, which portion of this overall solution I need to keep it specialized to get the best efficiency and performance, but which portion needs to be flexible so that I can adapt fast with the changing models needs of the data movement and connectivity. And that is why it is important to work very closely with the compute leaders in the industry. Now, being independent, it gives us a very, you know, kind of free, full freedom to be a close partner with the compute leaders, especially with NVIDIA, with ARM and Intel. So with NVIDIA, you know, NVIDIA is AI, you know, compute leader out there. We have partnership, not only the data center side, but more importantly, we have a very close partnership on the physical AI side, where we are working together to bring a complete portfolio of the reference design together from all the way top to the bottom, mapping use cases of the various GPUs from Gesture Nano all the way to Tor, and then mapping the right size of the PGA. That is why having a full portfolio is very important for this partnership and this solution to bring the experience to the end customer where they can really focus on actual physical AI application and they do not need to worry about the overall development environment or even the hardware overall platform where they need to work together. Similarly, with Arm, Arm is present in the overall ecosystem, pretty much in every embedded system out there and now extending into Edge and Intelligent Edge and Physical AI as well. We have a very long-term, very close partnership with Arm and we are putting together, in fact, in many, many designs, we are together already. where we are working very closely with them to bring the overall solution together. Our FPGAs is more of a SoC. Agilex SoC FPGA, they all include multi-core ARM, and that brings a unified solution for the customer, both for the data center as well as embedded world. So especially in data center side, especially in the management of the motherboard and rack has become much more complicated in this AI world. And this is where our integrated solution helps to enable our customer to manage that properly. And with Intel, we have had, of course, a partnership for a long time. We partner with them on the data center side, especially on the motherboard management and so on. But we are also, with this freedom, extending into the AI infrastructure. They are expanding into the physical side of the world. So having this ability to be independent, focus on FPGA, bringing the entire portfolio to serve the needs of the broader market is very important and it is giving us as an advantage sort of a thing to be the close partner for all these.

‍

Patrick Moorhead: 

Yeah, it's interesting. The industry always goes back and forth between integration is the best thing versus specialty, right? And with specialty, you're really focused on a bigger canvas, right? And there's going to be no raised eyebrows. When Altera was part of Intel, like, well, how close can you get to ARM? Right. And at the time, Intel was making, you know, has their own GPUs. Oh, well, how close can you get to NVIDIA? And this just cuts through it as an independent company. And I think your specialty in your market, I think there is strength in that. It's like you're not confused in where you want to take the company. Right. And

‍

Raghib Hussain: 

No, no other agenda. We are here to develop FPGA, drive innovation through FPGA, help our customers all through FPGA and provide them complete top to bottom platform. We are not selling anything else except FPGA.

‍

Patrick Moorhead: 

I think it's great. Hey, I was remiss. I didn't congratulate you on what looked like a very successful innovators day. You unveiled a ton. I mean, collaborations with companies, next generation, FPGAs, new agentic AI tools, which of course everybody needs to do. But question for you, or can you play back, what were customers building intelligent systems asking you for that shape this generation, right? What kind of feedback, what were the core needs that they were asking you for?

‍

Raghib Hussain: 

As you know, Patrick, I believe in working closely with the customer, understanding their problem and solving their problem. That is the mission of and my entire life through our businesses. And that's the mission we have brought in in Altera that, hey, we need to work closely with the customer, understand their problem, and then provide them the solution. So when I came in, we have done a lot of transformation in the last 18 months, but one of the most important one was go out, sit with the customer, get the feedback, and complete the portfolio if there are holes in there, which we did already in the current portfolio in Agilex. But then if we wanted to be the leader in pure play FPGA provider for the entire portfolio, we want to understand what are the needs of the next generation. And obviously, FPGA always lead in the technology adoptions and AI is driving that needs to the next level, right? So having that latest and greatest interfaces, we already have, for example, DDR5 in full production. In the next generation, we need to support the latest memory interfaces, latest PCI interfaces, latest Ethernet interfaces. We have customers using us for 400 gig, 800 gig type of Ethernet solution. We need to support them for 1.6 T or 3.6 gig type of a solution, right? So all those things, if you simplify it because FPGA is such a, I would say, versatile product and it's such a flexible product that you need to at the core, which is the core of the FPGA, you need to make the fabric that can keep up with the bandwidth and performance requirement for all those applications. So one of our focus was how can we improve the fabric to first of all drive even more performance, but then more importantly, drive more power efficiency so that I can really scale it down to really small, you know, power efficient FPGAs. So the Adelix Gen 2 that we have announced covers the entire range of FPGA application from the top to the bottom. In fact, there are different devices to address the needs of that market. That puts us, along with the third generation of our HyperFlex fabric architecture, it puts us in the lead, providing the best performance and best efficiency to the market. In addition, we also understand that many of the application driven by the AI, it requires a lot more custom data movement. And these are really high bandwidth data movement and needs to be done in the fast manner. So for that, we have introduced this new connectivity, hyper-floating, which you already learned in our innovator day. And that gives our customer to implement the best data movement and best performance for their application.

‍

Patrick Moorhead: 

No, I love it. You know, there's a lot of designs out there on FPGAs that have literally been in market for, it looks like forever, right? They're gigantic. And, you know, a lot of manufacturers trying to squeeze out something that's already been designed. By the way, nothing wrong with that, but in the end, brand new designs and new processes I know are a welcome site, particularly for this, uh, this market here, because everybody's trying to squeeze out, um, the, the highest levels of optimization, right. And, and power is just such a, and, and it's not just power, it's density, right. I've seen some FPGAs sitting on, uh, you know, I mean, they're, they're gigantic and they're, you know, the designs are five to eight years old.

‍

Raghib Hussain: 

So this is where if you notice one of the things that we are also doing it, how do we enable our customer with the ease? And this is where our announcement of Quartus IQ comes into the picture, because we are not only enabling AI in the physical and digital world, But we are also using AI to lower the barrier to entry without lowering the bar so you can get you can get the best performance without being the FPGA design expert that at a high level if you ask me that is that is what defines score does IQ, which means it gives the engineer, you know, awareness of the entire project level in an agentic flow to assist them, you know, what component what, you know, IP that needs to be used for to achieve their end system goal, and then assist them through throughout the design process to to achieve the best optimized design. And, and this is this is actually very helpful because it compresses the overall you know, the time to market from the design system intent all the way deploying the design in the field. So not only we are trying with this announcement, we are not only bringing a next level of performance, next level of density, next level of power efficiency, but also next level of ease of use.

‍

Patrick Moorhead: 

Now this is great. And the other big part of your launch was the agentic AI tools that it looks like you're focusing on deployment inside of physical AI, AI systems. And, you know, I'm curious. What change does that really provide to all the engineering teams who are building robots and autonomous machines? Does it also shrink time to market, adding more capabilities and less time? What's the value add here?

‍

Raghib Hussain: 

So I'll double click a little bit here. This is where we are partnering very closely with NVIDIA. So first of all, you know, there's also a concern for folks. Hey, we do not want our design to go out in the cloud or any such thing. So we are working very closely with NVIDIA to bring our entire, you know, agentic AI driven Quartus IQ on DGX station so that you can actually have it in your own facility and then you can still have a very, I should say, all the value of your historical learning from your design. So that's one aspect. On the physical AI side, we are integrating this whole agentic flow and our tool chain into the NVIDIA CUDA SDK CUDA developer environment. And the goal is this, that the physical AI application engineers can focus on just the application layer, without worrying about which sensor am I connecting, which what do you call the interface it is, and what do I need to do to adapt to the, let's say if I change a sensor from one vendor to another vendor, what all do I need to do, what driver do I have to rewrite, and so on and so forth. Our agentic flow actually enables them to adapt to these changing environment and not to worry about. In other words, making every sensor more intelligent AI enabled, think of it that way, and make it very easy to integrate it into the system and really not spend much time in worrying about all these changing dynamics and just focus on the application layer. So that is what we are working very closely together and we are also doing another thing which we are implementing reference design for targeted markets which help them to get it from the reference design to the production in a very fast manner. So those are all the things which are in the work which are going on and we are very thankful for NVIDIA to be a partner to bring all those things. The additional thing is this, of course, once we have all these things in place, we are also, as I said, working closely with other compute leaders as well. So our vision is this, that we should have a solution for all these robotics folks to focus on their application and not worry about the over all the physical world interfaces. I believe we are at a very exciting time in this technological transformation. I think this is the biggest transformation. And this one of the unique part of this transformation is the value of time to market. How do you bring your ideas to the hardware level in a faster manner? This is the moment which I believe FPGAs are built for. So it is a very exciting time for us also that we have the tools to enable our customers to solve their problem and create value for them. And that's how we all solve the world's problem and bring AI to the reality as fast as possible.

‍

Patrick Moorhead: 

Again, congratulations on the event. It definitely looks like it was a hit for you, and it sounds like you got great customer feedback. And again, I really enjoyed our time here with Geb. We're at time, but thank you so much. And I hope we can do this again to chronicle not only what you're doing, your leadership is doing, but the customer stories, the new products that you're bringing out, the new capabilities that you're enabling with all of your new product lines. Thank you. I want to thank all of you for tuning in here. These are my favorite conversations, CEOs and presidents of leading tech companies out there. And you know, I'm a little biased on Silicon. I love it. And this isn't the first time we've been talking about FPGAs on the show, but What's really unique here is the specialty focus, a full line from top to bottom, new leadership, new partnerships under Rick Ebb's leadership. So hit that subscribe button, follow us on the social and check out all the coverage at Six Five Media. Take care.

‍