❯ Cornelis Raises $205M to Challenge Closed AI Cluster Networks With an Open Fabric

Funding plus productCornelis Networks said it raised $205 million with backing from IAG Capital Partners while unveiling Active Compute Fabric. The company sells network adapters, switches, cables and software for AI and high-performance computing clusters. Its new architecture adds programmable compute inside the network so collective operations can be processed as data moves. Funding will support production, deployments and next-generation scale-up and scale-out products.

Intel rootsCornelis emerged from an Intel-related business in 2020 and announced a $29 million Series B in 2022, also backed by IAG. It is now moving from scale-out data-center networks into scale-up links within a rack, as the constraint shifts from individual chip speed to synchronized work across tens of thousands of accelerators. The company is financing CN5000 shipments, CN6000 production and its next open architecture.

Open networkActive Compute Fabric combines lossless transport, in-fabric acceleration and programmable compute using open standards including UALink, ESUN and Ultra Ethernet. Existing technology runs in hundreds of data centers. CN5000 is shipping, while the multi-protocol 800G CN6000 is sampling with customers ahead of broader availability expected in the fourth quarter. Cornelis’ claim that a 100,000-GPU cluster wastes $1.68 billion of compute is a model based on public data, not observed customer savings.

The bottleneck spreadsThe more expensive GPUs become, the less tolerable network waiting time is. Cornelis is targeting the economic value of idle accelerators, giving customers a second source and non-Nvidia chip vendors an open networking route. Its ecosystem, software and delivery scale still trail entrenched solutions. The $205 million supplies the capital to move from samples to production; actual deployments, not simulated savings, will set the next price.

▪ SIGNALAI infrastructure competition is expanding from who owns GPUs to who keeps costly GPUs from waiting for data.