AI infrastructure startup Cornelis has raised $205 million to expand a networking architecture designed to reduce one of the most expensive inefficiencies in AI computing: GPUs waiting for data.
The supplied source says the funding round was led by IAG Capital Partners.
Cornelis also announced Active Compute Fabric, a networking technology intended to let chips process and transmit information at the same time.
The company’s pitch is built around utilization.
AI accelerators are costly, and the economic value of a cluster depends on how much time those processors spend doing useful work.
When GPUs sit idle waiting for data to arrive, customers are paying for capacity that is not producing output.
Networking therefore becomes part of compute efficiency rather than a secondary component.
Cornelis is also attacking Nvidia’s ecosystem from a specific angle.
The company offers an open architecture that can work across multiple GPU and accelerator platforms.
Nvidia’s own networking products are tightly integrated with its GPU and software ecosystem, which gives customers a simple full-stack option.
That integration is a major competitive advantage.
It can also create an opening for customers that want more hardware flexibility or do not want every layer of the infrastructure stack tied to one supplier.
Cornelis spun out of Intel in 2020 and has already begun shipping its current product.
The company is developing a new generation expected later this year.
The investment significance extends beyond one private startup.
If open networking fabrics gain traction, Nvidia could face more competition in the part of the AI stack that connects accelerators together.
That does not threaten Nvidia’s GPU position directly.
It targets the economic value Nvidia captures around the GPU through networking and full-stack integration.
The key challenge for Cornelis is ecosystem adoption.
Customers need confidence that an alternative fabric is reliable, scalable and supported across hardware and software combinations.
Nvidia benefits from years of integration and customer familiarity.
What investors should watch: customer adoption of Active Compute Fabric, new accelerator partnerships, shipment growth, next-generation product launches and whether hyperscalers begin separating networking decisions from GPU purchases.
BTI’s bottom line: Cornelis is not trying to replace Nvidia’s GPUs. It is trying to make the networking layer more open and more efficient, which could chip away at Nvidia’s full-stack advantage if customers value flexibility enough to adopt a second ecosystem.