In the race to build bigger AI systems, the fastest chips in the world often sit idle, waiting for data to reach them. Cornelis, a networking company focused on fixing that bottleneck, has closed a $205 million funding round led by IAG Capital Partners, and it wants to use the money to challenge one of the most entrenched positions in tech: Nvidia's end-to-end control of the AI hardware stack.

Training and running large AI models requires thousands of processors to work together. Each chip is only as productive as the network feeding it, and when information moves slowly between processors, expensive compute capacity simply waits. For companies spending heavily on GPU clusters, that idle time translates directly into wasted money and slower results.

Cornelis is building networking technology designed to help AI chips talk to each other more efficiently. Its argument is simple: the next big gains in AI performance will not only come from faster processors, but from smarter ways of connecting them.

Alongside the funding, the company revealed a new offering called Active Compute Fabric. The technology is aimed squarely at the waiting problem. Rather than forcing chips to finish processing before passing data along, the fabric allows them to compute and transmit information simultaneously, keeping hardware busy and data flowing.

Cornelis says it is already shipping the product to customers and is developing a follow-up generation, which it expects to release before the end of the year.

Cornelis separated from Intel in 2020 and has positioned itself as the open alternative in AI networking. Its fabric is built to work with processors from multiple makers, giving buyers more freedom in how they design their infrastructure.

That openness matters because of how Nvidia has built its business. Nvidia GPUs can technically run on third-party networking fabrics, but they are tuned to perform best with Nvidia's own software and networking. That tight integration makes the full Nvidia stack the path of least resistance for many customers, which in turn reinforces the company's dominance.

Cornelis belongs to a growing group of AI infrastructure startups trying to break that dominance apart one layer at a time. Instead of attacking the GPU itself, these companies target the pieces around it, such as networking, interconnects, and software, where a better or more flexible option can win customers without requiring them to abandon their existing chips.

For startups and enterprises building AI products, a credible open networking layer could reduce vendor lock-in and improve the return on every GPU they buy. For investors, the round signals continued appetite for infrastructure plays that attack specific inefficiencies in the AI stack rather than competing head-on with the largest chipmakers.

The real test will be adoption. Nvidia's integrated ecosystem is deeply embedded in data centers worldwide, and Cornelis will need to prove that its performance gains and flexibility are compelling enough to justify moving away from the default choice.