A $205 Million Bet to Help AI Chips Talk Faster

Cornelis Networks announced on Monday, September 14, 2026, that it raised $205 million in a funding round led by IAG Capital Partners. TechCrunch

The company is based in Wayne and makes networking equipment that helps AI chips share information quickly. Think of its products as highways between chips. It spun off from Intel in 2020 and uses an open design, so customers can use many different types of GPUs and accelerators with its networking fabric, the shared system that moves data.

The funding comes alongside an expansion into scale-up networking with a product called Active Compute Fabric. Scale-up means the very fast connections inside a tight group of chips working together. The technology is designed to let chips do calculations and send data at the same time. Cornelis says Active Compute Fabric puts small amounts of computing at each step in the network, so traffic control and group data sharing run in the network itself rather than on the main AI chips. The product is shipping.

Cornelis plans to use the $205 million to make more of its CN5000 and CN6000 network switches. SiliconANGLE The company said the money will support its expansion and bring its next generation of products to market. It is working on a new generation expected later in 2026.

The CN5000 SuperNICs, plug-in network cards for servers, offer one or two ports at 400 Gb/s and fit in a standard low-profile x16 PCIe Gen 5 slot. Cornelis earlier introduced the CN6000 800 Gbps Ethernet SuperNIC for large AI and supercomputing systems. Its switches provide 48 ports at 400 Gb/s for 38.4 Tb/s in both directions and come with air or liquid cooling. Director Class Switches offer 576 ports at 400 Gb/s for up to 230.4 Tb/s in a two-tier fat tree layout inside one large box. Cable choices include passive copper, active copper and active optical cables at 400 Gb/s per link. Cornelis

Cornelis OPX Software is free, open-source software covering host drivers to network controls and built on libfabric, a common way for programs to talk to the network. Cornelis says software written with MPI, OpenSHMEM and RCCL, common tools for using many processors together, runs without code changes on its system. The CN5000 Omni-Path family won the HPCwire Readers' Choice and Editors' Choice award for Best HPC Interconnect Product or Technology.

Cornelis shared the Active Compute Fabric, $205M funding and Qualcomm collaboration news in a Business Wire press release titled Cornelis Expands into Scale-Up Networking with Active Compute Fabric $205M in Funding and Qualcomm Collaboration at AI Infra Summit. Business Wire It announced work with Qualcomm and presented the expansion, funding and collaboration at AI Infra Summit. In April 2026, Cornelis separately announced it was adding new federal partners and sellers to help get CN5000 into more supercomputing and data-heavy uses.

The broader context here is that AI data centers have two kinds of connections. Scale-out connects many servers across a building and has received most of the Ethernet attention. Scale-up connects chips that must act as one machine, with very high speed and very short delays, and has been harder to open to new suppliers.

In my view, that is why Cornelis is worth watching. If existing programs can run without changes, it costs customers less time and effort to try hardware beyond the usual supplier.

Worth flagging for buyers is whether the company can deliver at larger volume. Making current cards and switches is one test. Increasing supply of the 48-port and Director-class models while releasing the next generation later in 2026 is another. Work with many chip types and with Qualcomm will need to be proven in everyday use, including handling failures and delays, not only top speed.

Looking further ahead, if those tests go well, the long-term result should help customers. More choice in networking can lead to better prices, steadier supply and more flexible designs for supercomputing and AI. That result is not certain. It is the practical change this funding aims to make possible.