Cornelis Networks, a networking startup spun off from Intel in 2020, has raised $205 million in a funding round led by IAG Capital Partners, the company said this week, alongside the launch of a new architecture it is positioning as an open alternative to Nvidia's dominant AI-networking stack.
The new product, called Active Compute Fabric, is designed to address a persistent inefficiency in large AI training and inference clusters: expensive GPUs sitting idle while they wait for data to arrive from other chips in the cluster. Rather than treating the network purely as a transport layer, Cornelis's architecture combines programmable compute, lossless data transport and in-fabric acceleration so that some processing happens on data as it moves between accelerators, rather than only after it arrives. The company says this reduces the amount of time GPUs spend waiting rather than computing.
An Open Bet Against a Closed Leader
The pitch is explicitly aimed at Nvidia, whose NVLink and InfiniBand-based networking technology is deeply integrated with its own GPUs and has become a significant part of the company's competitive moat in AI data centers. Cornelis is positioning Active Compute Fabric as vendor-agnostic, meaning customers could pair it with GPUs and accelerators from multiple manufacturers rather than being locked into a single chip supplier, according to TechCrunch's reporting on the raise. That openness has become a common selling point among the wave of infrastructure startups trying to chip away at Nvidia's roughly 90 percent share of the AI accelerator market.
The new funding will go toward scaling up production of the company's CN6000 networking silicon, supporting customer deployments already underway, and continued development of next-generation Scale-Up and Scale-Out products, according to SiliconANGLE. Scale-Up networking connects chips within a single server rack, while Scale-Out networking links racks together across a data center; Cornelis is arguing that both need to be rethought as AI clusters grow into the hundreds of thousands of chips.
Part of a Broader Infrastructure Land Grab
Cornelis is one of several venture-backed challengers, alongside companies working on optical interconnects and custom silicon, trying to capture a share of the AI infrastructure buildout as hyperscalers look to diversify away from a single networking vendor. Whether an open architecture can match the performance hyperscalers get from Nvidia's tightly integrated hardware-software stack remains an open question that will likely be settled in production deployments rather than press releases.
The company has not disclosed its total funding to date or a valuation tied to this round. Cornelis said it has already begun shipping product to customers and expects to detail its next-generation platform later this year, an indication that the AI infrastructure market's appetite for alternatives to Nvidia's networking stack, however capital-intensive to build, has not slowed even as broader AI spending faces new scrutiny.
