U.S.-based Cornelis, which develops networking technologies for AI accelerators, has raised $205 million in a new funding round. The company wants to help makers and operators of AI infrastructure reduce their dependence on Nvidia’s solutions and use GPU computing power more efficiently.

The round was led by investment firm IAG Capital Partners. At the same time, Cornelis unveiled a new technology called Active Compute Fabric, which is designed to reduce GPU idle time when they are forced to wait for incoming data.

GPUs can sit idle while waiting for data

Modern AI systems combine large numbers of graphics processors and other accelerators. However, high computing power alone does not guarantee efficient operation: some of the time, chips may spend waiting for data that must arrive from other components of the system.

Cornelis proposes solving this problem with a network architecture that allows chips to process data and transmit it to one another at the same time. The company hopes to improve the efficiency of computing resource usage this way, without simply adding more GPUs.

The new technology has been named Active Compute Fabric. Cornelis has already begun shipping its product and is simultaneously developing the next generation of the system, which is expected by the end of 2026.

Cornelis wants to offer an alternative to Nvidia

The company emerged in 2020 after spinning out of Intel and is now trying to carve out a place in a market largely shaped by Nvidia.

One of Cornelis’s main differences is its open architecture. Its networking infrastructure can be used with different GPUs and AI accelerators, whereas Nvidia offers a tightly integrated set of hardware and software solutions.

Technically, Nvidia accelerators are also capable of working with networking technologies from other vendors. However, they are optimized for Nvidia’s own software, so for customers it is often simpler and more cost-effective to use the company’s entire infrastructure stack.

It is precisely this dependence that new players in the AI infrastructure market are trying to gradually reduce. Cornelis is one of the companies proposing to break AI systems into separate components and give customers the ability to choose hardware and networking solutions from different vendors.

For the market, this means an attempt to challenge Nvidia’s dominance not by directly replacing its GPUs, but through individual infrastructure elements that connect numerous accelerators into a single computing system.