Cornelis Raised $205 Million for AI Networking
The startup launched an architecture designed to reduce accelerator downtime by offloading compute to the network.
Updated on Sept. 25, 2026 in Data Centers
Live Poll
Do you believe new AI hardware infrastructure will effectively reduce data processing costs?
Cornelis has raised $205 million to advance its Active Compute Fabric, a new networking architecture designed to optimize large-scale AI clusters. The system, which is currently shipping in its CN5000 networking product, aims to mitigate the bottleneck of communication and synchronization overhead in massive deployments.
Why it matters
As AI clusters scale, networking demands often lead to significant accelerator downtime, causing an estimated $1.68 billion in lost capacity annually. Cornelis targets this inefficiency by shifting compute functions directly into the network fabric to keep processors active.
The Active Compute Fabric integrates scale-up standards UALink and ESUN with scale-out Ultra Ethernet specifications to move operations into the fabric. By offloading collective operations, the architecture aims to reclaim portions of the 500 gigawatt-hours of energy annually consumed by idle GPU capacity in massive systems.
The players
Cornelis
A networking technology company based in Wayne, Pennsylvania, focused on developing programmable fabric architectures for high-scale AI infrastructure.
The details
The Active Compute Fabric functions by performing data operations as information moves through the network. It combines lossless transport, in-fabric acceleration, and programmable compute to handle synchronization tasks that typically saturate accelerator capacity. This approach specifically targets the inefficiencies found in 100,000-GPU systems, where communication overhead often prevents maximum hardware utilization.
Timeline
The CN6000 product is expected to have expanded availability in the fourth quarter of 2026.
The market opportunity for open-standard AI networking is projected to reach $55 billion by 2030.
The Tech Race
Cornelis is positioning its Active Compute Fabric against existing proprietary cluster interconnects by embracing the open-standard Ultra Ethernet movement. This strategy aims to capture market share from established networking leaders by facilitating interoperability in large-scale AI deployments.
Enterprise data center operators and AI researchers can currently deploy the CN5000 product to begin integrating these networking capabilities. While the higher-spec CN6000 is currently sampling, broader availability is not expected until the fourth quarter of 2026.
The takeaway
Cornelis is betting that standardizing AI networking will solve the massive efficiency losses currently plaguing large GPU farms. Observers should track the Q4 2026 rollout of the CN6000 for benchmarks confirming whether in-fabric acceleration can successfully reduce idle accelerator time.
Further reading
For more on the changing requirements of physical infrastructure, visit Data Centers.
Source note: This article includes information reported by MyChesCo.
Live Poll
Do you believe new AI hardware infrastructure will effectively reduce data processing costs?
