Eridu
Last reviewed
Jun 3, 2026
Sources
7 citations
Review status
Source-backed
Revision
v1 · 1,494 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 3, 2026
Sources
7 citations
Review status
Source-backed
Revision
v1 · 1,494 words
Add missing citations, update stale details, or suggest a clearer explanation.
Eridu is an American startup developing networking chips and switches built specifically for large-scale artificial intelligence computing. Based in Saratoga, California, the company emerged from stealth on March 10, 2026 with an oversubscribed Series A round of more than $200 million, bringing its total funding to roughly $230 million. [1][2] Eridu argues that the data center networking layer, rather than compute, has become the binding constraint on building ever-larger AI clusters, a problem it calls the "network wall." Its founders are designing networking silicon from scratch to push far more bandwidth between GPUs than existing equipment can deliver. [2][3]
Eridu's central claim is that spending on AI accelerators is being wasted because the network connecting them cannot keep up. "Billions of dollars of investment in AI data centers are being wasted because of the network wall," chief executive Drew Perkins said when the company launched. [4] The argument is that as operators move from clusters of a few thousand chips toward systems with hundreds of thousands or millions of GPUs, the bandwidth each chip needs from the network has grown faster than switching hardware has improved. Perkins framed it simply: "Networking was not keeping up with compute, and this gap was growing." [3]
The company breaks the problem into three dimensions. Scale-up networking connects GPUs tightly within a single training domain, the role filled today by interconnects such as NVIDIA's NVLink. Scale-out networking links those domains into a larger cluster fabric, typically over Ethernet or InfiniBand. A third dimension, which Eridu calls scale-across, covers the emerging need to connect multiple clusters or even multiple data centers. [3] Chief technology officer Promode Nedungadi has tied the pressure to changes in how models run, noting that techniques such as mixture-of-experts routing and the splitting of inference into separate prefill and decode stages move more data per token through the fabric. [3]
Eridu is building a network switch and the silicon underneath it as a clean-sheet design rather than an incremental update to existing parts. The thesis is that off-the-shelf merchant switching chips cannot reach the scale AI needs. "You don't get to an order-of-magnitude higher scale using off-the-shelf silicon," Perkins said. [3] The company is designing its own high-radix switch chips, using a chiplet-based approach with advanced packaging, and pairing them with a custom software stack that spans the silicon, the package, and the full system. [3][5]
The headline performance claims describe a switch that can place far more GPUs under a single switching tier than current designs. For scale-up, Eridu says its system can connect thousands of GPUs in a single hop, with throughput on the order of tens of terabits per second per GPU. [4][5] For scale-out, it says a single tier can support more than 10,000 GPUs, and that two tiers can reach into the millions, removing the need for the third networking tier that large clusters typically require. [4] The company contrasts this with a hyperscaler that today maxes out at roughly 500 GPUs in a single scale unit using conventional switches. [4] Eridu also markets efficiency gains, citing up to 40 percent lower capital expenditure, up to 70 percent lower networking power consumption, and roughly a tenfold performance improvement over existing solutions. [4][5] These figures come from the company and have not been independently benchmarked, since Eridu has not yet disclosed detailed specifications or a general-availability date. It has said it plans to release more technical detail and partnership news later in 2026. [3]
Manufacturing relies on TSMC for advanced process technology and system integration, a relationship reflected in the cap table through TSMC's venture arm. [1][5]
Eridu was founded in 2023 by Drew Perkins, Omar Hassen, and Mike Capuano. [3][6] Perkins, the chief executive, has a long career in networking that predates the commercial internet. He began working in the field in the 1980s and helped develop the Point-to-Point Protocol, a component of the TCP/IP stack still used to establish network connections. [2] In 1999 the optical switching company he co-founded, Lightera Networks, was acquired by Ciena for more than $500 million. He went on to co-found Infinera, an optical networking firm that went public and was sold to Nokia for about $2.3 billion in 2025, as well as Gainspeed and the augmented-reality contact-lens company Mojo Vision. [2][3]
Perkins built the team around people who had spent their careers in networking silicon and systems. Omar Hassen, the chief product officer, came from chip design roles at Broadcom and Marvell. [2] Mike Capuano, the chief business development and marketing officer, joined as the third founder. [6] Promode Nedungadi serves as chief technology officer but is not listed among the founders. [3] The company began by questioning the assumptions of conventional switch architecture. "You have to start from a different place, otherwise you can hit the ceiling," Hassen said. [3] By the time it left stealth, Eridu employed roughly 100 people. [2]
The Series A was oversubscribed at more than $200 million and was led by Socratic Partners, with John Doerr, Hudson River Trading, Capricorn Investment Group, and Matter Venture Partners among the lead participants. [1][4] Counting earlier financing, Eridu has raised about $230 million in total. [2] Doerr, the chairman of the venture firm Kleiner Perkins, invested personally; he had backed one of Perkins' previous companies and was introduced to Eridu by Wen Hsieh, a founding managing partner at Matter Venture Partners. [2][7] "AI requires a new generation of infrastructure that Eridu is pioneering," Doerr said. [7]
A long list of strategic and financial investors joined the round. Reported backers include Bosch Ventures, Eclipse Capital, Fusion Fund, the chipmaker MediaTek, Osage University Partners, SBVA, TDK Ventures, VentureTech Alliance (a TSMC-affiliated fund), and Zelda Ventures, among others. [1][7] The presence of component makers such as MediaTek and TSMC's investment vehicle, alongside trading firm Hudson River Trading, points to interest from both the supply side and the heavy compute users who would deploy the technology. Perkins declined to disclose a valuation, describing it only as in line with other Series A rounds of comparable size. [2]
| Funding detail | Reported figure |
|---|---|
| Round type | Series A (oversubscribed) [1][4] |
| Series A amount | More than $200 million [1] |
| Total raised | About $230 million [2] |
| Lead investors | Socratic Partners, John Doerr, Hudson River Trading, Capricorn Investment Group, Matter Venture Partners [1][4] |
| Founded | 2023, Saratoga, California [3][6] |
| Headcount at launch | About 100 [2] |
Eridu enters a market it sizes at around $200 billion, and it does so against entrenched incumbents. [2] In switching silicon and systems, its most direct rivals are Broadcom, whose Tomahawk and Jericho chips power much of the industry's Ethernet switching, along with Cisco, Arista Networks, and Marvell. [2][3] NVIDIA is both a partner-by-necessity and a competitor: its GPUs are what these networks connect, but its NVLink scale-up fabric, Spectrum-X Ethernet, and InfiniBand products directly overlap with what Eridu wants to sell. [3]
The startup's pitch is that the incumbents are delivering incremental improvements on an architecture that was not designed for AI scale, which is why Eridu chose to start from the silicon rather than from a finished switch. [3] That positions it among a cluster of companies trying to rethink AI networking rather than iterate on it, a group that includes DriveNets, which disaggregates the network into a distributed routing fabric, and industry efforts such as the Ultra Ethernet Consortium and UALink that aim to standardize open alternatives to proprietary interconnects. Whether a new entrant can displace established switch vendors in production AI data centers is unsettled. Eridu has strong financing, a founding team with multiple prior exits, and a foundry partner in TSMC, but it has not yet shipped a product or named a customer, and its performance claims remain unverified outside the company. The next test will be the technical detail and deployment partners it has promised for later in 2026. [3][4]