AI chip interconnect technology company Cornelis announced that it had raised $205 million in a funding round led by IAG Capital Partners. The announcement coincided with the launch of a new technology called Active Compute Fabric, which aims to reduce the time graphics processing units (GPUs) spend waiting for data to arrive instead of performing processing operations.
A Network for Processing and Transmitting Data in Parallel
Cornelis says its technology relies on a network fabric that allows chips to process and transmit information at the same time. The design targets an operational problem in AI infrastructure: some GPU time may remain unused when data does not reach the processing unit quickly enough.
The source provides no figures on the improvement achieved by Active Compute Fabric, nor does it mention additional technical details about the product’s architecture or the tests conducted on it. Therefore, in its current form, the announcement is more an indication of the product’s direction than a quantitative presentation of its performance.
A Bet on an Open Architecture
Cornelis is competing with Nvidia by adopting an open architecture that allows customers to use a variety of GPUs and accelerators with its network fabric. This approach differs from relying on an integrated ecosystem tied to a single vendor, an important distinction in a market where AI workloads require coordination among processing, memory, and networking.
The source notes that Nvidia chips can technically work with other networks, but they are optimized to integrate with the company’s networking software, making the use of Nvidia’s full stack easier and more attractive to customers. Cornelis is therefore betting on separating infrastructure components and offering a networking alternative that does not require a single type of hardware.
What Actually Changed?
The announcement is not limited to a funding round: Cornelis said it has already begun shipping its product and is simultaneously working on a new generation expected to be released later in the year. This gives the company a product in use, but the source does not identify the customers or markets, or specify the scale of shipments.
Cornelis was founded after being spun out of Intel in 2020 and is part of a wave of AI infrastructure companies attempting to reduce dependence on Nvidia’s ecosystem by targeting specific parts of the computing stack. The importance of the news lies not only in the funding, but also in the company’s attempt to address a practical bottleneck in accelerator operations: delivering data to processing units at the right time.
However, Cornelis’s ability to change the market balance will depend on details not disclosed in the announcement, including its actual performance compared with Nvidia’s networks, the extent to which the technology is compatible with different types of accelerators, and customers’ willingness to replace or separate infrastructure components they prefer to obtain from a single provider.