Eridu emerges from stealth with $200M Series A to solve AI networking bottlenecks

Veteran entrepreneur tackles AI networking with $200 million war chest

Eridu, an AI networking startup led by serial entrepreneur Drew Perkins, has emerged from stealth mode with an oversubscribed $200 million Series A funding round 1

1

. The round was led by Socratic Partners, renowned venture capitalist John Doerr, Hudson River Trading, Capricorn Investment Group, and Matter Venture Partners, bringing the Saratoga, California-based company's total funding to $230 million 1

1

. The investor lineup also includes chip industry heavyweights MediaTek and Taiwan Semiconductor Manufacturing Co. through its VentureTech Alliance investment vehicle, signaling strong industry confidence in Eridu's approach 2

2

.

Perkins brings decades of networking expertise to the venture. He helped create the Point-to-Point Protocol that became integral to TCP/IP in the 1980s, co-founded optical switch company Lightera Networks (sold to Ciena for over $500 million in 1999), and led Infinera through its IPO before Nokia acquired it for $2.3 billion in 2025 1

1

. His epiphany came in February 2023 during a conversation with OpenAI CEO Sam Altman, who explained that ChatGPT's success required enormous amounts of compute—initially around 4,000 GPUs, now scaling to millions 1

1

.

Breaking through the network wall in AI data centers

Perkins argues that billions of dollars invested in AI data centers are going to waste because of what he calls the "network wall" 2

2

. The core problem is that GPU compute and memory bandwidth are improving by roughly 10x per year, while data center switches from Broadcom, Marvell, and Cisco are only improving 2-3x every 2-3 years 1

1

. This growing gap creates networking bottlenecks that prevent AI workloads from scaling effectively, no matter how many GPUs are deployed.

Existing network architectures were designed for traditional cloud data centers, not the massive AI factories being built today. When more networking capacity is needed, operators add more boxes, increasing the number of hops each data bit must travel and driving up latency—the delay between typing a prompt and receiving a response 1

1

. Perkins told SiliconANGLE that existing networking companies only promise incremental enhancements because they're focused on improving legacy architectures that are being pushed to their limits 2

2

.

Clean-sheet silicon architecture promises order-of-magnitude improvements

Eridu's solution centers on a completely redesigned AI network switch built from a clean-sheet silicon architecture that takes full advantage of the most advanced silicon and packaging technologies 2

2

. By late 2023, Perkins had partnered with co-founder Omar Hassen, whose background includes networking chip design for industry giants like Broadcom and Marvell, and together they founded Eridu in 2024 1

1

.

The company's approach integrates more networking functionality directly onto the chip itself, reducing reliance on optical connections that Perkins identifies as the least reliable part of networks 1

1

. A single high-radix Eridu switch can replace up to 30 lower-level radix switches, creating a much flatter network architecture 2

2

. Radix refers to the number of input/output ports on a switch, and increasing this metric is critical for improving GPU communication at scale.

Promising metrics for hyperscalers and enterprise customers

Eridu promises fewer network tiers to reduce both latency and network jitter, enabling data center operators to dramatically increase the scale of their GPU clusters 2

2

. The company claims its design will support single-hop scale-up domains with thousands of GPUs and scale-out domains of millions of GPUs, while delivering up to 40 percent savings in CapEx and upwards of 70 percent reduction in networking power consumption 3

3

. Improved tokens-per-watt and tokens-per-dollar metrics would directly impact the economics of running large language models and other AI workloads 2

2

.

Dylan Patel, an analyst with SemiAnalytics, endorsed Eridu's approach, stating that the networking technology is "purpose-built for massive AI scale, supporting both scale-out and scale-up" and will "enable more bandwidth, larger AI clusters and lower costs" 2

2

. While Eridu doesn't have publicly announced customers yet, Perkins said the company is working closely with leading hyperscalers who have validated the product definition and been deeply involved in shaping its design, though NDAs prevent naming them 2

2

.

Market positioning and go-to-market strategy

Eridu will eventually sell complete systems that occupy the same position in AI data centers that classic network gear providers like Arista Networks hold in traditional data centers 1

1

. The company currently employs around 100 people and plans to use the Series A funding to complete development of its offering and accelerate recruiting efforts 1

1

2

2

.

Initial customers will be hyperscalers through direct business relationships, but the company plans to work through channel partners once it expands to neocloud, service provider, and large enterprise customers 3

3

. This positions Eridu to address what it describes as a $200 billion AI networking market 3

3

.

Perkins declined to comment on valuation but indicated it's comparable to others raising similar amounts in Series A rounds, and he wants his employees to benefit from their stock options 1

1

. The fundraising process itself became competitive, with Perkins noting his phone was "ringing off the hook" after Wen Hsieh, founding managing partner of Matter Venture Partners, introduced the opportunity to Doerr, who had backed one of Perkins' previous startups 1

1

. If Eridu delivers on its promise, it will position itself at the center of the largest data center build-out in history, armed with the deep technical experience increasingly rare in Silicon Valley's current startup landscape.