After nearly three decades at Intel, Wilfred Gomes set out to rethink AI infrastructure from the system up. 

AI’s next infrastructure challenge is physical. As models grow larger and compute demand accelerates, data centers are running into constraints that cannot be solved by improving a single chip or component. Power, cooling, memory, connectivity, cost, and reliability are increasingly intertwined, and addressing them requires rethinking the system itself.

That is the premise behind Mueon. Led by co-founder and CEO Wilfred Gomes, a former Intel Fellow, Mueon is developing a new system-scale architecture for AI infrastructure through Cubelets™, modular building blocks that bring compute, memory, power delivery, and thermal management in a compact new form. The ambitious goal is make high-performance AI infrastructure denser, more efficient, and easier to deploy beyond the traditional rack.

For Wilfred, Mueon is the culmination of nearly three decades at Intel, where he worked across CPU and GPU design, memory, packaging, power delivery, thermals and advanced system integration. A named inventor on more than 250 patents and patent applications, he came away from that experience with a simple conviction: the next leap in computing would not come from optimizing one part of the stack, but from redesigning the stack as a whole.

Wilfred also knew that pursuing a system-level architecture would require expertise across nearly every layer of semiconductor design and manufacturing. He and co-founder and CTO Jack Hwang had worked together on Foveros at Intel, where Jack also helped advance backside power delivery and high-volume fabrication processes. Together, they assembled a group with deep expertise in lithography, wafer thinning, dielectrics, memory, power delivery, reliability and process integration. Several had worked together at Intel and that shared history has created a high-trust team of hands-on technologists accustomed to taking difficult semiconductor ideas from first principles into high-volume production.

Mueon was also the first investment from the Geodesic Alliance Fund. It reflects what drew us to the company from the beginning: a deeply technical team taking on a foundational infrastructure problem with the potential to reshape how AI systems are built and deployed.

We spoke with Wilfred about the physical limits of today’s data centers, how Cubelets could change where and how AI compute is deployed, and how his experience at Intel shaped the vision for Mueon. 

After nearly three decades at Intel, including helping develop technologies such as Foveros, how did that experience shape the vision for Mueon?

WG: At Intel, we were working with the latest process technology and thinking about how to build entirely new kinds of products. Over time, it became clear that products would increasingly be constrained by power, performance and yield—and that improving a single chip would not be enough. 

That realization helped lead to 2D and 3D integration, including Foveros. Instead of treating each component in isolation, we began thinking about how different elements could be brought together as a system. Mueon grew from taking that idea further: applying system-level integration not just within a package, but across the computing infrastructure itself. 

Was there an “aha” moment that convinced you that the next major breakthrough in AI infrastructure would need to happen at the system level—not through another incremental improvement to a chip, server, or rack? 

WG: Absolutely. At Intel, I was leading the work on what would be integrated into CPU products. We were looking at system-level innovations across everything from graphics and chipsets to memory. 

It became very clear, especially when looking at memory, that the future would require system-scale integration. Even when we were writing about this in 2015 and 2016, it was becoming obvious that a single chip was not enough. You had to build full systems—and large systems—to really realize the potential of compute.

As AI workloads scale, what limitations in today’s data-center architecture are becoming most urgent?

WG: The first one is power. If you look at the data center buildouts today, they’re limited by the amount of power that can actually go into a building.

Second, cost is becoming very important because, in the end, the economics have to work.

And third is reliability. Even if you have the power and are willing to spend the money, you still need to make these systems reliable so that you can get good answers with good fidelity. 

What is the core idea behind Mueon’s system-level approach to AI infrastructure?

WG: The idea came from looking at what the industry is doing today. We’re creating these massive, incredibly complicated data centers, and there are reasons for that. But it reminded me of what happened in the 1980s.

At that time, you had workstations and big mainframes, and people were enthralled by them. Then the PC came along, and eventually those systems were supplanted because the PC became much lower cost. Then came the phone. Every decade or so, a new form factor emerges and replaces the old one, because Moore’s Law has always been about scaling things.

So one of the fundamental questions we asked was: the industry is spending billions  of dollars on fab technology. Can we use that existing fab ecosystem to essentially bring FAB technology and the existing supply chain  to  a full data center?

That’s really what Mueon is about: system-scale integration and architecture using existing fab equipment to build systems. The goal is to make data-center-scale computing dramatically smaller, denser and more power-efficient. 

Can you walk us through a Cubelet in practical terms? What does it replace, how does it change the physical design of a computing system, and what would a customer notice compared with a conventional rack-based deployment?

WG: A Cubelet is a compact, stackable building block that brings compute, memory, power delivery and thermal management together as an integrated system. Today, those elements are distributed across chips, boards, servers, racks, cables and cooling infrastructure. A Cubelet is designed to compress that architecture into a much smaller form factor. Let me walk through three potential users of a Cubelet.

Today, if you go into a data center, the data center itself is massive. Industry  has talked about Manhattan-scale, gigawatt-plus data centers that could cost $50 billion to $70 billion to build. These are very large systems. If Moore’s Law were really at work, what if you could shrink all of that to one-hundredth of its size? You could put that compute into a normal office building or another location where power is readily available. Today, you may need something approaching nuclear-power-plant-scale generation for a gigawatt data center, while many large industrial sites have access to something closer to one to 10 megawatts.

Second, if you go inside the data center, you’ll see racks that today can consume 200 to 600 kilowatts and cost roughly $10 million. What if I could give you that capability on something closer to a large chip that could operate as an appliance in your garage or another local environment, consuming perhaps 10 to 100 kilowatts? There are many organizations that want to curate their own data, keep it on-premise, or potentially even launch compute into space. If you dramatically reduce the weight and form factor, those kinds of applications become much more practical.

Third, think about today’s GPUs, which can consume 10 to 20 kilowatts in these systems. What if that became 100 or 500 watts? Suddenly, everybody could have an AI workstation at their disposal and run applications without having to send their data always to the cloud.

Today, it is very difficult to create additional data centers because they are so constrained. But if you can change the form factor and power this dramatically, I think the use of computing could increase by orders of magnitude. You could have what I call “conscious computing” or perpetual AI, always at your disposal. Compute would go from being a very exotic activity to something much closer to ubiquitous AI.  Our exponential computing platform enables this new paradigm and will be the new Moores paradigm law for System Scale architectures.

Why did you choose Geodesic as an investor and partner, and how has the team contributed as Mueon moves from an ambitious technical vision toward real-world deployment?

WG: We were the first investment out of the Geodesic Alliance Fund, so we were very fortunate to meet Rayfe Gaspar-Asaoka, Neb Mela, and Tom Gillespie. We’re really proud to have Geodesic as an investor in Mueon.

I think there were three things that made the fit strong. First, the Geodesic team likes companies that are solving very hard, fundamental problems. There’s a belief that solving hard, fundamental problems can also lead to strong investments.

Second, Mueon is very tied to the semiconductor ecosystem, national labs, and defense, where we see a lot of potential applications. That was clearly of interest to them.

Third, we’re deeply connected to equipment suppliers and to partnerships between Japan and the U.S., which aligns very well with Geodesic.

Since the investment, the team has opened doors for us across those ecosystems. Rayfe, Neb and Tom have also become trusted partners as we build the company.

Is there anything else you want people to understand about Mueon?

WG: I think we are asking a very simple question: What is the appropriate architecture for AI?

Nobody has really had the opportunity to revisit that from the ground up. We have, and one of the things that has become clear is that this only works if you co-optimize at the system level.

Mueon is really a system-scale company. We are also very fortunate to be partnering with the broader ecosystem, because that’s how the semiconductor industry moves forward and how everybody benefits. We have the opportunity to work with hyperscaler partners, equipment partners, fab and foundry partners, and others across the industry.

Bringing this kind of innovation to market is difficult. But the industry has given us an opportunity to carve out a space for this work, and we are grateful for that.

Learn more about Mueon at mueon.com.