There's a peculiar kind of confidence that comes from having already sold two companies for billions of dollars. Drew Perkins has that—twice over, in fact—and on a Tuesday morning in March, he brought it to bear on what he believes is the most pressing infrastructure problem in artificial intelligence: the network can't keep up.
Perkins, who co-founded Infinera in 2000 and watched Nokia acquire it in 2024 for roughly $2.3 billion, emerged from stealth mode this week with Eridu. The Saratoga-based startup is building what it calls purpose-built AI network switches. And it arrived with something rarely seen even in frothy times: more than $200 million in Series A funding.
The money—$230 million in total capital raised, according to TechCrunch—suggests investors think Perkins might be onto something. The oversubscribed round was led by Socratic Partners, with backing from an unusual coalition: John Doerr (the Kleiner Perkins chairman who wrote checks to Google and Amazon back when), Hudson River Trading, Capricorn Investment Group, and Matter Venture Partners. Doerr invested personally. "AI requires a new generation of infrastructure that Eridu is pioneering," he said in a statement that, for once, didn't sound like pure venture hyperbole.
But the real tell is in the semiconductor names on the cap table. TSMC's affiliated VC arm VentureTech Alliance joined the round—alongside MediaTek, Bosch Ventures, TDK Ventures, Eclipse Capital, Fusion Fund, and SBVA—effectively turning the investor roster into a strategic map of the semiconductor and AI infrastructure stack. TSMC isn't just writing a check; it's a manufacturing partner, which signals early alignment on advanced packaging and process technology. "We are excited to work with Eridu... using TSMC's advanced process and advanced system integration technologies," Dr. Lucas Tsai, TSMC North America VP, said in the company's press release. That kind of language from a foundry partner this early? It matters.
The Widening Gap
Perkins' argument is disarmingly straightforward: GPU performance doubles every couple of years. Network switches from incumbent vendors? They deliver incremental 2x gains every two to three years, he told SiliconANGLE. The math isn't working. What AI models demand—massive, tightly coupled compute fabrics shuffling weights and gradients at blinding speed—is outpacing what datacenter networks can realistically deliver.
The numbers, at least as analysts see them, back this up. A January analysis by 650 Group projects the data center networking market will approach or possibly exceed $200 billion annually by decade's end. AI scale-out Ethernet alone could top $100 billion by 2030. Those figures assume someone figures out how to actually build the switches to handle it. Eridu is betting it can capture a meaningful slice by rethinking how switches are architected in the first place.
The pitch is not subtle. The company is developing a high-radix network switch built from custom silicon, advanced packaging, and proprietary systems integration. Eridu claims its architecture can replace up to 30 conventional switches with a single unit—a 10x radix increase that collapses sprawling multi-tier networks into simpler topologies. The result, if it works: single-hop scale-up domains supporting thousands of GPUs, and two-tier scale-out fabrics that could span more than a million GPUs.
Performance targets are aggressive, perhaps more than some hyperscalers expected. The company promises up to 40% capital expenditure savings and up to 70% reductions in networking power consumption, according to a March press release. Patent applications filed in mid-2025 and published in early 2026 hint at 3D die-stacking and scalable packaging architectures designed to push both radix and efficiency beyond what's currently possible with merchant silicon.
The Infinera Playbook, With New Variables
If the strategy sounds familiar, it should. Perkins spent three decades building optical networking companies that reshaped telecom infrastructure. Before Infinera, he co-founded Lightera Networks, which Ciena bought for more than $500 million in 1999. That's the playbook: vertical integration, custom silicon, systems-level thinking. Eridu is applying it to a different layer of the stack.
His co-founders bring similar pedigrees, though the domain has shifted from optical to compute fabrics. Omar Hassen, co-founder and chief product officer, held senior roles at Ventana (RISC-V server chips), AppliedMicro, Marvell, and Broadcom. Mike Capuano, co-founder and chief business development and marketing officer, spent years at Cisco, Juniper, and Pluribus—the latter acquired by Arista in 2022. It's a team that's built ASICs before, scaled products before, and navigated the politics of hyperscaler procurement before.
The company has grown to around 100 employees, per TechCrunch, with an R&D center in Bengaluru led by Utkarsh Rai. Job postings for ASIC design verification and network switch SDK engineers in both Saratoga and India suggest the company is deep into hardware development. Perkins told SiliconANGLE that Eridu is working closely with "leading hyperscalers" under non-disclosure agreements, and the product definition is "completely validated." No customers have been publicly named—standard operating procedure at this stage.
A Crowded, High-Stakes Battlefield
Eridu is entering a fiercely competitive landscape, which is both opportunity and risk. Nvidia's NVLink and Spectrum-X dominate much of the conversation around AI networking. Broadcom's Tomahawk family, paired with custom SUE/ESUN stacks for hyperscalers, owns significant market share. Marvell's merchant ASICs, Cisco's AI datacenter fabric—everyone with a semiconductor roadmap and datacenter relationships is staking a claim.
The difference, Perkins argues, is that Eridu isn't iterating on existing architectures. It's replacing them. That's a bold position to take against incumbents with established relationships and proven supply chains. But it's also the only bet that makes sense at this valuation. Incremental improvements don't justify $230 million in Series A capital.
Network World reported the company plans additional product and partnership announcements later in the year. The initial customer targets are hyperscalers—the handful of companies building AI infrastructure at planetary scale—with eventual expansion to neoclouds, sovereign clouds, and large enterprises. That sequencing makes sense: prove the technology where the pain is most acute, then broaden.
Whether Eridu can execute is an open question. Building custom silicon is expensive and unforgiving. Datacenter switches require not just performance but reliability, interoperability, and supply chain resilience. And hyperscalers, for all their complaints about incumbent vendors, are deeply conservative when it comes to network infrastructure. They'll test new architectures, but they won't bet their AI training runs on unproven hardware.
Still, the capital is real, the team is credible, and the technical thesis—that AI networking needs to be rebuilt from the ground up—is hard to dismiss. Perkins has done this before, in a different era and a different part of the stack. Now he's betting he can do it again, at a moment when the industry might actually need him to.
