Semiconductor Insiders - Podcast EP366: How Cognichip is Changing Chip Design with Faraj Aalaei

Episode Date: September 16, 2026

Daniel is joined by the CEO of Cognichip, Faraj Aalaei, a successful visionary entrepreneur with over 40 years of distinguished experience in communications and networking technologies. As a leading entrepreneur in Silicon Valley, he was responsible for building and leading two semiconductor companies through IPOs as a founder… Read More

Hello, my name is Daniel Nenny, founder of Semaywiki, the Open Forum for Semiconductor Professionals. Welcome to the Semiconductor Insiders podcast series. My guest today is the CEO of Cognithip, Farage, Alahi, a successful visionary entrepreneur with over 40 years of distinguished experience in communications and networking technologies. As a leading entrepreneur in Silicon Valley, he was responsible for building and leading two semiconductor companies through IPOs as a founder and CEO. Welcome to the podcast, Farage.

Thank you. Glad to be here. It's great to speak with you. And I know we're going to be a meeting at a couple of conferences coming up. And it's a great opportunity to talk and get to know you a bit. Can we start with telling us how you ended up in the semiconductor industry in the first place? Very interesting question. So I didn't start in the semiconductor industry. I actually started as a designer of systems.

And I spent the first three, four years of my life designing circuits for satellite communication. Then I went to Bell Labs for 10 years and built systems that Bell Labs was putting into data centers of those times. They used to be called central offices and so on. In any case, when DSL came around to replace dial-up modems, that was my first entry. into thinking about building a chip company. So Centillium Communication was that company, where I came to it from a systems perspective and my co-founders were actually chip designers.

So that was my first entry back in 97. And the reason I decided to flip over to the chip side of this is because anything that was on the table in those days to go replace dial-up modems, which just actually not possible to build systems, because of the power consumption, the number of chips you needed, just the complexity of just serving one user was not possible. So the thought occurred to me that, hey, you know,

if we can do X, Y, and Z, we can reduce 32 chips that was being offered by large companies into two chips, and now you can actually deliver service at scale. So that was sort of like the motivation to come into the chip business back in 1997. and I never left. I'm still here in the chip business and in its industry that I love because a lot of magic happens when you're designing chips, right?

That's where the core IP is. That's where the possibilities come together. And so it's very exciting to be at always the leading edge of what comes next into the consumer's hand or into enterprises. So it's a great place to be and now it's even better because everybody's starting to pay attention to the chip company.

Yeah, I agree. Semawiki is 15 years old and the traffic of the last couple of years has just exploded. And, you know, the semiconductor industry is now front page news and it's quite an exciting time. But, you know, you've kind of done a pivot now. You know, you've taken two chip companies public and ran a billion dollar segment at Marvell. So tell us a little bit, what made you decide to start Cognit chip? Yeah, so after I left, I sold my company to Marvell and I left, I started to start a little bit. I started to take two or three years and actually make investments. And a lot of these investments were in software, ironically.

I would get calls from chip designers saying, hey, I have this chip in mind, you know, would you look at it? And the answer was no, I'm not going to invest in chip companies because it's the devil I know and I know it's damn too expensive. And as a seed investor, I was looking for software companies. And of course it became the age of AI. And I made investments in AI companies

and spending time with these AI scientists and entrepreneurs, I was learning enough about AI to be dangerous. And so, but that's when the light went on, is the light went on and saying, well, all the problems in these last two or three decades have been building these companies. What if AI could solve the problems? So fundamentally the problems,

as I perceive them, where it takes too long to build these chips, it's way too expensive. And just my own experience, when we built the first company, Centilium, we raised $50 million. When I took it public, we still had 17 of that 50 in the bank. The second company, Aquanchia, we raised $200 million as a private company,

and we took it public probably sooner than we should have because we didn't wanna raise any more money, from private, private markets. And today, as you can see, it's like hundreds of millions of dollars to go do chips. And so my motivation was, what if with AI we could collapse that time

and we could collapse the amount of capital you need, just like going back to essentially the days of when we started Centillion. Now you get entrepreneurs excited, you get the venture capital firms excited to fund these companies to some point of, fruition of their vision and then go from there. That is not possible today. Today, chip design is the venue for large companies

or entrepreneur, chip entrepreneurs who are actually building vertically integrated companies. They build a chip, they build a system, and sometimes they go off and actually offer the service as well to be able to afford to tape out those chips. So I came to this industry, I switched over to AI side to be able to help the industry I love, be able to reinvent itself and keep up pace with the software industry.

Yeah, I agree. You know, I've been in the semiconductor industry for 40 years and it, just personal observation I share with you is that, you know, chip development hasn't sped up the way software development has. And I mean, it's just cost too much to build chips and that limits the market of people that can build chips, right? So just a bottom line, I mean, is AI going to fix this? I believe so. I believe that if you fundamentally approach the problem, AI is the vehicle to

help us solve this problem. Now, AI is sort of like a broad term everybody throws around AI. Not all AI is made for all kinds of problems, right? So, you know, and you take the clues from, you know, looking at, for example, the pharmaceutical industry, the medical industry, a lot of times, when you're thinking about AI, you have to kind of think, all right, what problem is it trying to solve? And is it, is every AI system that's out there made to solve the problem you're trying to solve? And our answer to that from the beginning was no. We knew enough with my co-founders to understand that in order to solve a problem for semiconductor, you have to train these models to understand what chip design is because number one, it's very complex. And number two, at the end of the day, it's not like vibe coding. You know,

you can't be wrong in chip design. And these coding that you do for chip design eventually ends up being a transistor that sits next to 100 billion other transistors and how they interact and how the collection of the transistors you arrange have to solve a problem. Well, you know, large language models by and large do not understand that because there's been no data available for them to train on. So we took it on ourselves to go back to fundamentals and say, yes, AI can solve this problem. But in order to create that AI to be able to solve the problem, we have a lot of work to do. Got it.

So let's talk a little bit about cognitive chip and the product you call. ACI. So ACI is a physics-informed foundation model rather than an LLM applied to design. So what's the actual distinction and what is the model trained on? So it's a chip-specific frontier model, right? And as I mentioned before, it's built from ground up to help chip designers bring about their creativity, but move faster through the process without making mistakes. Okay. So a general purpose LLM is it's trained on pros. It reads text, it creates codes, some hardware examples. It's very good at sounding fluent in everything, but it does not understand chip design because it was never trained on the physics. And one of the reasons that

it wasn't, it's because there's not a lot of open source chip design examples out there for you, like it is in software. Like as you know, software has been around for a very long time in terms of a lot of open source data available. And so the models could train on that and be pretty good actually at creating code. But for chip design, those models aren't there. Those data sets are on there for models to train.

They don't understand timing constraint. They don't understand power budget. What are the trade-offs in area? And what is the judgment that when an engineer look at those, pray those, how does it make that judgment? What information does it need to make that judgment? Those things are just not built into LLN. So what you have to do is reduce this guessing,

which is expensive liability in our industry, if you're guessing about some kind of design. But then train a model on large curated and govern data sets using techniques at the frontier of synthetic data generation, reinforcement learning, noitry reward extraction. There's a lot of techniques in the AI world that has been used in some of the other physical industries,

I would call them physical industries because there's a physics at the end of the day that have been developed. And so take some of those learnings from those industries and bring them into semiconductor and apply them to our field and then create models that actually understand chip design. Oh, interesting.

So you've said, in-house silicon teams at Apple, Amazon, Google, or proof that the industry needs more specialized chips. If ACI works the way you intend, who gets to build the custom silicon that can't today? Yeah. So, you know, as I've said before in many forums, is that, you know, because of the expense associated with chip design, it's really now the land of very large chip companies, or very, very highly funded chip startups. But the need for this purpose chips is not limited to these large companies. And it actually stops innovation when chips can only be designed in a few large companies.

You know, so the workloads are getting more specialized. You got edge devices. You got energy constraint systems. You got a physical AI system like humanoid robots, autonomous vehicles, industrial automation. Who builds ships for them? Right. And so the challenge here is to figure out a design system where lower budget organizations who need specialized ships can also do this, can get into the act of creating their own barriers to entry, creating their own.

their own IP that is unique to them and it differentiates their product. Chips are the fundamental way you differentiate your product in the hardware world. So giving tools to these people so that let that imagination fly and not be constrained by the capital or the time needed to get to the market with those chips. So you're working with some of the top 20 chip makers. What are they actually using it for today? And what results have surprised you the most? Yeah, so these chip makers that were engaged with, their usage varies seriously by customers.

At least the initial usage is varying. And so the variation from my perspective is the best evidence that this is an infrastructure rather than a point solution. I tell my colleagues in the semi-industry that what you have to think about is what AI is going to do to your business three years from now. don't think just about the efficiency you get in the next six months. You got to be thinking about what does this world look like three years from now. And what this world looks like three years from now is to make sure that your organization is organized in a way where creativity and thinking and architectural definition is where your focus is, not the daily grind of getting into the process of working out the design. right so so that the work shifts to the front end the human intelligence gets focused more on the front end of the design process this lengthy design process because that's where market decisions

are made that's where your differentiation is built and let the AI systems get you through the process the rest of the way to a tape out yeah so you've raised 93 million you added lip-butan to board, who's one of my favorite semiconductor industry leaders. What's next for cognitivehip? Well, thank you so much for that question. Lipu is an asset to any organization. You know, this is my third company with Lipu. He's supported my adventures in this entrepreneurship world

for the third time in a row. And we've been successful the last couple of times. And hopefully this will be, there'll be even a bigger success. But, you know, ACI operates today, like a well-trained designer. But the aim for us is to reach beyond that capability. So that ACI can be better than any living chip designer. That's our goal. That's our ultimate, we're our North Star, where we're reaching.

The frontier ahead, though, is the physical AI. You know, every one of these robots and vehicles and automation system ultimately a physical AI model that runs on a physical chip. And most of these systems will need custom silicon rather than general purpose compute to perform in the real world. So now the challenge is, how do we get these physical AI devices to get to the market at the speed of software? So from the time you decide you want to do something

to the time that new model is running inside a humanoid, or the next update to your automation system, that time needs to be shrunk significantly. You cannot wait every five years for your human aid to be upgraded in this kind of world, right? So for us, how do we bring that speed to physical AI? How do we bring it to the edge of the network? And how do we make sure that you can actually

get through that transition as fast as possible? We're not built to be a feature. We're built to be an infrastructure for the next generation of hardware and how they get built. That's the way we look at ourselves, and that's what we're striving to become. Great. It's a pleasure to meet you. Great conversation. I'll see you at the AI InfraSummit next month, so we can talk a little bit more there. I'm very much looking forward to it. Thank you for having me. That concludes our podcast. Thank you all for listening, and have a great day.