Eridu exits stealth with $200M to rebuild AI networking

Another startup is emerging to take on the challenge of AI networking. This time, it has a name taken from the mists of history.

AI networking startup Eridu came out of stealth on Tuesday with more than $200 million in funding. The company is targeting what it says is a fundamental architectural mismatch between the scale demands of AI data centers and what current switching technology can deliver.

The Saratoga, Calif.-based company was founded by Drew Perkins, who previously co-founded Lightera, acquired by Ciena, as well as Infinera which was acquired for $2.3 billion by Nokia. The oversubscribed Series A was led by Socratic Partners, with participation from John Doerr, Hudson River Trading, Capricorn Investment Group and Matter Venture Partners, among others. Perkins named the company after Eridu, the ancient Mesopotamian city widely considered one of the earliest in recorded history.

Eridu is developing a clean-sheet network switch built around custom silicon. Stated targets include hyperscalers, neoclouds, and sovereign cloud operators. The company has not disclosed product specifications or a general availability date. More technical details and partnership announcements are planned for later in 2026.

The competitive field Eridu is entering is not small, with the company taking aim at large incumbents including Broadcom, Cisco, and Nvidia. The company argues that the existing vendors have been delivering incremental improvements on the same underlying architecture rather than addressing the actual scale requirements of AI infrastructure. “AI data center networking needs to be rethought from the ground up with a clean sheet of paper, because it’s different than cloud data center networking,” Perkins told Network World. “Every time there’s a new application, a new need, it has different requirements. From the chips up through the packages, through the system, through the software, the optics, everything needs to be thought of again for a different kind of application.”

AI networking is a different problem from cloud

The demands of AI networking are different from the networking technologies that have been developed and deployed for the cloud.

Perkins said the core difference comes down to scale. A typical cloud data center runs roughly 100,000 servers, each using tens of gigabits per second of bandwidth. AI data centers are being built around millions of GPUs, each requiring even larger amounts of bandwidth. “Networking has been stuck on the same 2x every two years kind of improvement for 30 years, certainly 20 years,” he said. “Networking was not keeping up with compute, and this gap was growing.”

That gap is not static. Promode Nedungadi, Chief Technology Officer, said the architectural and algorithmic trends driving AI are making the network problem harder, not easier. Techniques like mixture-of-experts models and the disaggregation of inference into separate prefill and decode stages all require more data movement.

“Every one of those requires more data to be moved around,” he said. “The amount of data being moved per token is growing.” The scale challenge also has more than one dimension. Perkins described three: scale-up, which refers to interconnecting GPUs within a single training domain; scale-out, which covers the broader cluster fabric; and what he called scale-across, an emerging requirement that standards bodies are beginning to address.

“We think that scale-across is quite interesting as well,” Perkins said.

Architecture: silicon, packaging, and software

A key differentiator for Eridu will come from silicon.

“There’s no doubt that we are developing our own silicon. We’re developing the most advanced silicon in the networking sector, bar none, period, and that’s absolutely necessary,” Perkins said. “You don’t get to an order-of-magnitude higher scale using off-the-shelf silicon.”

Eridu has a partnership with TSMC for process technology and advanced system integration. Perkins said TSMC sees the networking bottleneck as tied directly to its own business. The silicon approach is likely to benefit from chiplet-based architecture and advanced packaging.

“We believe you need to be on a different technology arc than what the mainstream technology is,” Omar Hassen, Chief Product Officer, told Network World. “In terms of things like advanced packaging, you’ve got to take advantage of everything you can from chiplet-based architecture, clean-sheet design, and advanced packaging. We believe we’re on the right technology arc that can take us beyond what the existing incumbents are doing.”

Fundamentally, Eridu’s approach is an attempt to break through the architectural ceiling facing incumbents to get an order-of-magnitude improvement.

“You have to start from a different place, otherwise you can hit the ceiling if you just use the same architecture as everybody else,” Hassen said.

Eridu at a glance:

• Founded : 2023
• Funding: $200 million.
• Investors: Socratic Partners, John Doerr, Hudson River Trading, Capricorn Investment Group, and Matter Venture Partners.
• Headquarters: Saratoga, California
• CEO : Drew Perkins
• What they do: AI networking