Cornelis Networks raises $205 million for AI networking technology
The Intel spin-off is partnering with Qualcomm and bringing an open alternative to Nvidia's networking technology to market with Active Compute Fabric.
Illustration · AI-generated (AI IN LIFE)
At a glance
- Funding: $205 million, led by IAG Capital Partners
- New partnership: with chipmaker Qualcomm
- Company origin: spun off from Intel in 2020
- New technology: "Active Compute Fabric" — programmable compute built into the network
- Products: CN5000 and CN6000 switch lines, production scale-up planned
Data-centre networking specialist Cornelis Networks has closed a funding round of 205 million dollars, led by IAG Capital Partners. At the same time, the company announced a strategic partnership with chipmaker Qualcomm.
Cornelis spun off from Intel in 2020 and specializes in congestion-free networking systems for AI and high-performance computing. The fresh capital will mainly go toward scaling up production of its CN5000 and CN6000 switch lines.
At the heart of the announcement is a new, open architecture called Active Compute Fabric. It builds programmable compute directly into network cards, combining lossless data transport with processing power that works both within a server (scale-up) and across multiple servers (scale-out).
The practical benefit: because network cards can compute while simultaneously moving data, expensive graphics processors spend less time idle waiting for data. Cornelis also promises that network functions can be adapted to new AI workloads via software, without new hardware.
With the move, Cornelis is explicitly positioning itself as a challenger to market leader Nvidia, whose proprietary NVLink technology currently dominates networking in large AI data centres — more on Nvidia's own upcoming platform in our report on Rubin NVL72. Whether Cornelis's open architecture can keep pace in practice remains to be proven in real large-scale deployments.
FAQ
What makes Active Compute Fabric different from existing networking technology?
Traditional network cards only move data. Active Compute Fabric additionally builds programmable compute directly into the networking hardware, so part of the data processing can happen while it is still in transit.
Why does this matter for AI data centres?
Graphics processors are expensive and are meant to compute as continuously as possible. If they have to wait for incoming data, that wastes compute time. Networking technology that computes on its own can reduce that waiting.
Is Cornelis going head-to-head with Nvidia?
Yes, at least in networking technology for large AI clusters. Nvidia currently dominates this space with proprietary, closed technologies like NVLink; Cornelis explicitly positions its solution as an open alternative.
