Alan Fagan: Welcome back to the Critical Lowdown, everybody. I'm Alan Fagan, VP of Sales for the Americas at EPS Global. Today, we're going somewhere a little bit different. For the last few years on the podcast, we've talked about open networking for service providers, ISPs, and data centers. Today, we're talking about what is arguably the most demanding network environment ever built: the AI factories being stood up by Neo Clouds and sovereign data center operators around the world. We've got the perfect person to dig into it with. Matt Free from Aria Networks is someone who has spent his career at the intersection of enterprise networking and AI. Matt, welcome to the show.

Matt Free: Thank you for having me, Alan.

Alan Fagan: Could you tell us a little bit about yourself and the path that brought you to Aria Networks?? 

Matt Free: Sure. I have 25-plus years in the IT data center realm. I started off in the late '90s slinging storage, then turned to Fibre Channel networking and SANs. I spent most of my career at Cisco—14 years leading data center teams there. I spent some time at Nutanix and then Juniper. If you look at where those companies were focused, it was around cloud networking. Now, it's a totally different game around AI networking. I met Mansour Karam when I was at Juniper; he was doing some very innovative things. When he left and realized there was a point in time where AI networking was going to be a major shift, he addressed that challenge, and I couldn't not go. So, I'm happy to be here.

Alan Fagan: Aria came out of stealth in April and made a big splash with a $125 million raise, so congratulations on that for a start! You came out under the banner of "networks that think." You mentioned the founding team—we actually knew Mansour from his Apstra days—with backgrounds from Apstra, Arista, Juniper, and Meta. For people hearing the name for the first time, what is Aria, and what is the core problem you are going after for AI cluster operators?

Matt Free: Thanks for the question. It's really the whole system. It's hardware, telemetry, and intelligence all rolled into one. You can't do piece-parts without doing the whole thing. If you look at AI networking, it's fundamentally different in terms of speed and performance. Most vendors look at their switch and only see their portion of the problem. If you look at legacy or current vendors, the operators are looking at second-level telemetry, but we look at microsecond telemetry. We have hardware—Tomahawk 5, Tomahawk 6—with SONiC burned in. We also have something called an FDE (Forward Deployment Engineer) that really goes out and makes sure the integration points with the systems are all working together. It's a professional service that is bolted on.

Alan Fagan: Taking a step back and looking at the market—we're seeing it in our own business, selling a lot of optical transceivers from Coherent; the demand is like nothing we've ever seen before. Neo Cloud revenue is projected to be over $25 billion in 2025, but the really exciting thing is that Synergy and others are forecasting it to be close to $400 billion by 2031. You're in front of the operators every week. From your perspective, what's changing about how they're building infrastructure now versus 12 to 18 months ago?

Matt Free: That's another great question. I'd say the networking piece and the hardware were good enough. Legacy vendors, through acquisition and over time, let software work around some of the issues, which can't be done anymore. There are gray failures, speed issues, and congestion that can't be seen. What Aria is doing is pulling data right off the chip, which allows us to get microsecond telemetry. We see what other vendors cannot see; they're blind to the gray failures that the network suffers from. That's number one.

Number two, if you look at AI infrastructure as a whole, about 10 to 15 percent of it is the network. But when it's running properly, it equates to a tremendous amount of revenue. The network has moved away from being a cost center to more of a revenue multiplier.

Alan Fagan: What do you think the difference is now between how Neo Cloud operators think about infrastructure compared to a traditional enterprise or hyperscaler? 

Matt Free: I liken it to the farming industry. You had the land grab—everyone was going after power, cooling, and space. Now, we've moved from that to needing to produce a harvest. When you're standing up infrastructure, that's fine, but once you start losing or wasting compute, that's where the big issues are. If you've got a farm and your irrigation system has a leak, in a traditional enterprise it might not matter much. But if you're trying to really reap the harvest, you need that system to be in lockstep, providing the best water you can to that field so you can yield fruit from it.

Alan Fagan: Aria's view is that the network sitting underneath an AI cluster has historically been a bit of a black box. You know it's there, you know it's busy, but you can't really see what it's doing. You guys have designed yours the opposite way—from the chip up, as you mentioned earlier—so it's constantly reporting on itself in real-time. In plain terms, what does that actually let an operator see and do that they couldn't do before?

Matt Free: It's what we call "deep networking." Deep means a deep inspection of not just the switch, but the entire fabric:

• The NIC
• The transceivers
• The network

Typically, vendors can't see what's going on regarding degradation. That telemetry we get is really the secret sauce that helps the operator look at where problems are. On top of that, we use AI. We don't just slap on an LLM; we are built with AI for AI. We use AI to:

• Auto-detect
• Auto-discover
• Auto-fix

We're talking to our networks, not yelling at them. We're able to address issues before they happen.

Alan Fagan: An interesting part of that is that the system is not just designed for human network engineers; you're designing it for use by AI agents too.

Matt Free: Exactly.

Alan Fagan: What does that mean in practice, and how is operating an Aria network different compared to a traditional setup?

Matt Free: As I mentioned before, typically, folks have to rewrite software to find out what hardware issues are going on. Now, you see a much broader view of performance degradation. It's also at scale. We're not talking about a hundred servers; we're talking about tens of thousands of GPUs. There's a term called MFU—Model FLOPs Utilization—which wasn't a term many years ago. A 1% improvement in MFU can turn into a 5% to 6% increase in revenue for a customer. It can also pay for the network for five years. You can't ignore those types of numbers. Quite frankly, this is why CEOs and co-founders of these Neo Clouds are so interested in talking to Aria. Even CROs are involved, because it's a revenue generator, not a cost center.

Alan Fagan: The relationship with EPS is a relatively new one—well, you're a startup, so every relationship you have is pretty new! But talk us through how the relationship with EPS came together and what makes a value-added distributor like us the right route to market for you, rather than going direct.

Matt Free: There are other distributors out there, but EPS Global obviously has a global presence. We can't be everything to everyone out there. EPS has the ability to execute; they're talking to the same customers, the same partners, and solving the same problems. When you look at transceivers, you guys have that experience. We're small, we have an amazing product, and we're getting all the attention, but we can't scale without partners like you.

Alan Fagan: Yeah, we have a relationship with Coherent going back 25 years, believe it or not, and we're their largest global distributor. We've never seen demand quite like it is now; it's unprecedented. We're building a reference architecture that's specifically aimed at the Neo Cloud technology stack. Aria sits as one layer in that stack, which spans everything from the optics we just mentioned, to switches, down to power, cabling, and install. From a customer's day-one perspective, how do you see that helping them?

Matt Free: Most Neo Clouds think of the network as an afterthought. They're so concerned about getting the GPUs: where's the power, where's the cabling, how do we get funding? What this does is reduce the risk. Your experience, combined with our engine and end-to-end visibility of the network, allows them to stand things up very quickly. It takes the risk out for the Neo Cloud. It's the life of Ricky Bobby: "If you ain't first, you're last." These Neo Clouds want to get up, stay up, and increase their efficiencies.

Alan Fagan: Great analogy to use given that we're sitting here in Indianapolis! Looking forward, we talked about how much money there is out there. There's a huge amount of capital landing right now in this space—tens of billions across CoreWeave, Microsoft, and sovereign builds in Europe and the Middle East. Looking ahead to the next 12 to 18 months, where are the genuine bottlenecks? Power, optics, software, or something else?

Matt Free: I think it's going to be power and optics for a while. Some of the vendors have a hard time with the supply chain. They've been in traditional networking, and now the supply chain has caught up with them. But for the most part, Neo Cloud CEOs need to look at the network not as a cost center anymore; it must be a revenue-generation machine. If they start looking at it that way, skate to where the puck is going, and do it quickly, the network is something they won't have to worry about in the future.

I'd also say, with our FDE deployments, these Neo Clouds don't have the opportunity to spin up and hire 100 people to work on the network. They need something that's proven—proven by EPS Global and Aria—that's readily available in terms of the supply chain to get up and running as fast as possible.

Alan Fagan: We're seeing that as demand shifts from training to inference, the data centers look a bit different. They're not hyperscalers. We feel like they need more help; they need help from Aria and EPS to put these solutions together because they haven't got the army of engineers to do this.

Matt Free: No, and it takes a long time to hire them, if you even can.

Alan Fagan: Exactly. And fairly enough, they don't have to. They can find a partner that can help them put this solution together.

Matt Free: 100%. They can use that money to buy more GPUs.

Alan Fagan: And more networking!

So, hopefully, we'll have some Neo Cloud operators watching this video or listening to this podcast. Whether they're a year into a build or just about to start one, what is the one piece of advice you'd want them to be thinking about when specifying the network layer?

Matt Free: I would say: don't neglect the power of the network, and don't assume that it's just going to work. Look at other accelerators out there:

• AMD
• Positron
• D-Matrix
• Cerebras

They've all made big announcements. Operators want to make sure that the network is going to work in whatever GPU or accelerator environment they choose, especially moving from training to inference. So, don't underestimate the network or treat it as an afterthought.

I'd also say, with our FDE deployments, these Neo Clouds don't have the opportunity to spin up and hire 100 people to work on the network. They need something that's proven—proven by EPS Global and Aria—that's readily available in terms of the supply chain to get up and running as fast as possible.

Alan Fagan: Well, that's it for the questions, Matt. Thanks very much, this was very informative. I'm grateful you took the time. We're very excited about this partnership and really looking forward to seeing where it goes.