The round
Cornelis Networks has raised $205 million in a round led by IAG Capital Partners. The company, which spun out of Intel in 2020, used the announcement at the AI Infra Summit in Santa Clara to unveil an architecture it calls Active Compute Fabric and a strategic collaboration with Qualcomm, SiliconANGLE reported.
Cornelis sells an open networking architecture that works with a range of GPUs and accelerators rather than one vendor’s chips, TechCrunch reported, which puts it in competition with Nvidia. It did not disclose a valuation or name other investors. The company is already shipping product, and TechCrunch said a new generation is expected later this year.
What the fabric is meant to do
In most AI clusters the network moves data between accelerators and does nothing else. Cornelis wants the fabric to do work while the data is in transit. “The payload does not arrive the way it left,” chief marketing officer Brandon Draeger told SiliconANGLE.
According to the company, the fabric can assemble KV cache data, coordinate expert dispatch for mixture-of-experts models and accelerate collective operations before data reaches its destination. The aim is to cut the time GPUs spend waiting.
The numbers behind the pitch are Cornelis’s own. Draeger said accelerator utilisation in large AI deployments typically sits near 50% of installed capacity. The company said its pre-production simulations show network traffic falling by up to 50%. Those are simulations, not measurements from a customer cluster, and no independent results have been published.
“AI infrastructure is reaching a point where faster endpoints alone are not enough. The fabric has to become an active part of the compute system,” said chief executive Lisa Spelman.
The Qualcomm piece
Cornelis and Qualcomm are “in advanced stages of joint technology evaluation focused on rack-scale inference across both scale-up and scale-out environments”, SiliconANGLE reported. Scale-up links accelerators inside a rack so they behave as one system; scale-out connects racks to each other. Neither company named a product or a date.
Why networking is contested
Nvidia sells the accelerators and much of the networking that ties them together, and it is extending that reach. Last week d-Matrix said it would use Nvidia’s NVLink Fusion to connect its own Raptor accelerators at rack scale. Cornelis is betting that customers mixing chips from several vendors will want a fabric that is not tied to any of them.
What to watch
The next milestones are the new product generation TechCrunch said is due later this year, and whether the Qualcomm evaluation turns into a named product. The claim that matters most, that in-network processing raises GPU utilisation, will need figures from a production cluster to be tested.
