Cornelis Raises $205 Million to Challenge Nvidia's Networking

Most of the money and attention in the AI hardware race goes to the chips. The wiring between them rarely makes headlines. Cornelis Networks is trying to change that. On September 14 the company, an Intel spinoff, said it had raised $205 million led by IAG Capital Partners, and it used the moment to argue that the network sitting between thousands of accelerators has become the bottleneck that matters.

Cornelis calls its approach Active Compute Fabric. The pitch, laid out in its coverage by Network World, is that the fabric should do some of the work rather than simply shuttle data. The design pairs lossless transport with programmable acceleration inside the network, so certain operations happen in transit instead of waiting for a GPU to pick them up. In a large training or inference cluster, a lot of time is lost to chips sitting idle while they wait for numbers to arrive from somewhere else. Shaving that idle time is the whole game.

The case against lock-in

The competitive angle is unmistakable. Nvidia sells not just the accelerators but the interconnect that binds them, and that vertical grip is a large part of why it is hard to build a serious AI cluster without buying deep into one vendor's stack. Cornelis is pushing an open alternative, one that TechCrunch describes as GPU-agnostic, meaning an operator could mix accelerators from different suppliers behind the same fabric.

Whether that pitch lands is a separate question. Buyers tend to value a single throat to choke when a cluster misbehaves, and Nvidia's software ecosystem is a formidable moat on its own. But the appetite for options is real, and it is the same appetite driving the wider fight over who profits from AI infrastructure, from custom silicon deals to the strain the data-center boom is putting on local power grids.

What the money buys

Cornelis says the round will fund production of its CN5000 switch, which is shipping now at 400 gigabits per second, and its next generation CN6000 at 800 gigabits. The company spun out of Intel in 2020 and has been working the high-performance networking market since, so this is less a debut than a bid to matter at the scale AI clusters have reached.

The broader point is that the AI buildout is no longer only about who has the fastest chip. It is about the plumbing, and the plumbing is where a challenger can still find room. Nvidia's own dominance runs through its interconnect as much as its GPUs, the same reason its recent acquisitions have drawn scrutiny. A well-funded rival arguing that the network should compute, not just connect, is a reminder that the map of this industry is not finished being drawn.