In this exclusive interview with Swapnil Bhartiya at TFiR, Drew Perkins, CEO and Co-Founder, and Omar Hassen, CPO and Co-Founder at Eridu, deep dive into the critical constraints of legacy networking in AI data centers and how their clean-sheet approach is unlocking unprecedented GPU performance.

The AI Compute vs. Network Mismatch

While GPUs have continued to scale aggressively to meet the demands of massive AI workloads, network infrastructure has struggled to keep pace. This creates a severe bottleneck where processors remain idle, waiting for data.

Q: What are the exact constraints that the current breed of network is posing on AI data centers?

Drew Perkins: “Networking technology, in particular Ethernet switch chips, has been doubling in capacity every silicon node. Every time we’ve gone from seven nanometers to six nanometers, five nanometers, and four nanometers, these switch chips have doubled in capacity. However, GPUs have been growing far faster than that. GPU and computation requirements have been growing far faster than networking has been able to support. This causes a problem because GPUs need to ingest and process data very quickly, but they can’t do that if they don’t have the data. The network is simply not keeping up with the need to feed them data as fast as possible. Our observation is that we need to find a way to let the network keep up with the computational side of the system, and that takes new technology.”

To address this mismatch, Eridu’s goal is to fundamentally rethink the entire network switch architecture—including the silicon, packaging, system design, and optics—to ensure data flow matches the computational processing rate.

Complacency in the Networking Industry

Many established networking giants failed to anticipate the explosive growth requirements brought on by AI models, relying on historical upgrade cycles instead.

Q: Why couldn’t well-established networking companies solve this problem?

Drew Perkins: “I think the networking industry has grown complacent over the last several decades, assuming that doubling capacity every two years—keeping up with Moore’s Law—has been good enough. The entire industry is based on that assumption. In the CPU space, performance was also doubling, which matched the networking pace. But GPUs came about and required 10X growth every couple of years instead. The networking vendors didn’t get that message. Two-times growth has worked well economically and financially, but we observed that the computation industry is leaving the network in the dust. We were the first to do something about it and try to rebuild a company from the ground up designed to keep pace with the computation side.”

Eridu’s Clean-Sheet Network Architecture

To overcome the limitations of legacy designs, Eridu is rethinking everything from chip architecture and packaging to system-level optics, aiming to build a more scalable, flat network.

Q: What is your approach to solving this problem? Are you tweaking existing technologies or using a clean-sheet design?

Drew Perkins: “It literally requires looking at everything from the ground up—from how semiconductors are made to how they are packaged. With the massive growth in the numbers of GPUs in the AI data center and the way they ingest data, we realized you need networking switches that have higher rates and a higher number of ports. This allows you to build a network that is flat, higher performance, and lower latency. We are attacking all those problems at every level we can in order to achieve a faster network that keeps up with the GPUs.”

Drew Perkins: “It just required us to rethink how the chips are built, how they are architected, how the packages of those chips are done, and how you could possibly build something with a higher radix and a higher number of ports that is a flatter network. What we are doing here is building a more scalable network that is designed for scalability far more than today’s networking switches.”

The Hyperscaler Ecosystem and Target Audience

The immediate beneficiaries of this infrastructural leap are the entities building the world’s largest supercomputers.

Q: Who is the ideal target audience for this next-generation networking technology?

Drew Perkins: “Our primary customers are the hyperscalers that are deploying the GPUs and the networks that interconnect them to build these AI data centers. Of course, those are then being used by the people building the models—organizations like OpenAI and Anthropic. Google is combined in that case, and Amazon and Microsoft to some degree. But fundamentally, our target is the organization that is designing the GPU and data center system.”

CapEx, Power Consumption, and Economic Value

Beyond performance, inefficient networking actively cannibalizes power and capital that should be dedicated to revenue-generating compute operations.

Q: Will this technology augment or replace existing switches, and how does it impact data center economics?

Omar Hassen: “In addition to the performance benefits, you also get additional economic benefits from the higher radix switches, which include lower network power. As it stands today, the amount of power being used by the network is growing because you need multiple layers of these smaller switches. That is essentially taking away power from the compute, which is where the tokens are generated and where data centers are actually creating economic value. By lowering the network power, you can create more tokens by virtue of dedicating more power to compute. Additionally, the cost of the network is growing as a percentage of IT spend in a data center, crossing over 25% and sometimes nearing 30%. With our solution, we are able to reduce network power by 70% and the cost of the network by about 40%. That is a significant amount of CapEx and power that can be redirected to the economic benefit of the data center operators.”

Q: Based on studies, how much GPU capacity is currently lost due to network constraints?

Omar Hassen: “There are a number of published studies, and Meta in particular has done great work measuring this and publishing real results. If you look at what companies like Nvidia are saying—specifically Jensen Huang in his keynotes—he mentions how the network can help improve performance by up to 25%. He noted that even a 10% improvement in network performance results in billions of dollars in economic value for these data centers. There is an industry-wide recognition that network performance directly impacts the performance of compute.”

Revolutionizing the Pace of Innovation

As AI models evolve, the underlying physical infrastructure must innovate at an unprecedented velocity to keep operations viable.

Q: Is this a natural evolution of networking, or a revolution forced by AI workloads?

Drew Perkins: “There is a fine line between evolution and revolution. When evolution moves at a much faster speed, it looks more like a revolution. We are trying to move things at a faster pace than has historically happened and bring new technologies to bear more quickly. You might look at some of the things we are doing as revolutionary, or you might say we are just evolving faster. In any case, our goal is to bring new technologies to market faster than would otherwise happen, improve network performance, and grow this into a fantastic leading company.”

Watch the full TFiR interview with Drew Perkins and Omar Hassen here.