The archive · Hardware & Devices · Strategic decision · 2024–2026
Eridu bets AI's bottleneck is the network; exits stealth with $200M
Clean-sheet AI data center switch maker Eridu exited stealth in March 2026 with a $200M Series A (total $230M), betting networking, not GPUs, is AI's wall.
Eridu
What the business is
Designs networking chips, switches and complete systems for AI data centers, replacing many tiered optical hops with on-chip communications.
Starting capital：$230M total: $200M oversubscribed Series A led by Socratic Partners, John Doerr, Matter Venture Partners, Hudson River Trading and Capricorn, with SBVA, MediaTek, Bosch Ventures, TDK, Eclipse and TSMC's VentureTech Alliance
How it started
After ChatGPT launched, Perkins concluded from a 2023 conversation with Sam Altman that the next bottleneck would be how millions of GPUs talk to each other, not access to compute. In 2024 he and Hassen founded Eridu to build a brand-new kind of network chip and system.
What happened
The team designed a clean-sheet switch that puts more functions on-chip, claiming one unit can replace up to 30 legacy switches and cut data center capex by up to 40%. The founders' decades of relationships — Wen Hsieh at Matter Venture Partners, then John Doerr — turned a few calls into an oversubscribed Series A. Eridu came out of stealth on March 10, 2026 with roughly 100 employees; SiliconANGLE reported it had no official customers yet, with hyperscalers engaged under NDA.
No ending yet — it is still running.
Background
Eridu builds networking chips, switches and systems for AI data centers, betting that the wall stopping AI scale is not GPUs but the network between them. Drew Perkins — who helped create PPP, sold Lightera Networks to Ciena, and co-founded Infinera (later sold to Nokia for $2.3B) — got the idea from a 2023 chat with Sam Altman, who said what enabled AI was enormous amounts of compute, then 4,000 GPUs, now millions.
In 2024 Perkins and co-founder Omar Hassen, whose background is networking chip design at Broadcom and Marvell, founded Eridu to rethink networking from the silicon up: a high-radix switch that replaces many tiered optical hops with on-chip communications, cutting latency, power and cost. Perkins argues GPU compute improves roughly 10x a year while data center switches improve only 2–3x every 2–3 years, so incremental upgrades cannot keep up.
On March 10, 2026 Eridu came out of stealth with an oversubscribed $200M Series A — about $230M raised in total — led by Socratic Partners, John Doerr, Matter Venture Partners, Hudson River Trading and Capricorn, with chip-industry names including MediaTek and TSMC's VentureTech Alliance. The company has roughly 100 employees and, per SiliconANGLE, no official customers yet, though Perkins says hyperscalers are deeply engaged under NDA.
The bet is unproven: the switch has not shipped. But Eridu's founders bring decades of networking credibility, and its investors are pricing the largest data center build-out in history, where whoever owns the network between millions of GPUs owns a critical layer of AI infrastructure.
What has to be true
- AI progress shifted from chips to the network between them: as clusters grew to millions of GPUs, switch bandwidth improving only 2–3x per generation became the bottleneck.
- The founders had the credibility to raise: Perkins' two prior networking companies were acquired for over $500M and $2.3B, and Hassen designed chips at Broadcom and Marvell.
- They bet on a clean-sheet silicon architecture rather than incremental upgrades, claiming one high-radix switch can replace up to 30 legacy switches.
- The round was oversubscribed with strategic chip investors (MediaTek, TSMC) who could both validate and manufacture the design.
What can be applied
When a category improves 2–3x per generation while demand grows ~10x a year, a clean-sheet redesign can be the right bet — but $200M for unshipped silicon still has everything to prove.
Aftermath
As of March 2026, Eridu is still building: the clean-sheet switch has not launched and it reports no official customers, though Perkins says hyperscalers are co-designing the product under NDAs. The $200M Series A funds completion of the silicon, systems and software, plus recruiting beyond roughly 100 employees. The company targets the AI networking market with systems that could support domains scaling to millions of GPUs, replacing legacy tiered-optical networks. Its path to revenue and proof of the order-of-magnitude claims — lower latency, up to 40% capex savings — remains ahead.
Sources
- AI network startup Eridu emerges from stealth with hefty $200M Series A
- With $200M in funding, Eridu wants to break through the network wall holding back AI
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card