Cornelis Raised $205m to Put RISC-V Cores in the Network Fabric

16 SEP 2026 — Cornelis Networks has raised $205 million and used the announcement to argue that AI infrastructure loses money in the time accelerators spend waiting, not in the accelerators themselves.

The company, which spun out of Intel in 2020, is putting programmable compute inside the network itself. It calls the result Active Compute Fabric. The pitch is that a switch processing data in transit gives expensive accelerators fewer reasons to sit idle.

What "active" means here

The fabric embeds RISC-V cores directly into the network silicon, in both the interface cards and the switches. Two workloads are named as targets: acceleration of the key-value cache that large models use during inference, and the routing of tokens between experts in mixture-of-experts models.

In both cases, a conventional network is a courier. Data leaves an accelerator, crosses the fabric untouched, and arrives to be processed somewhere else. Putting a small amount of compute in the path means some of that work happens in transit.

Putting compute in the data path is harder than it sounds, which is why this is a funding round rather than a finished market. A switch that transforms data has to do it without reordering packets, without becoming a failure domain of its own, and fast enough that the detour beats sending the work to an accelerator. Those constraints have kept in-network computation mostly in research clusters and high-performance computing rather than in commercial AI serving.

The design keeps the parts of the company's Omni-Path heritage that were about moving packets reliably under load: congestion management, credit-based flow control and packet spraying. The CN6000 platform runs at 800 gigabits and carries both Ethernet with RoCEv2 and Omni-Path modes on a single piece of silicon, and works with Broadcom Tomahawk switches when it is in Ethernet mode.

The claim, and how to test it

Lisa Spelman, the chief executive, put the value proposition in the terms a data-centre operator cares about: "We're giving you your GPUs back...you can drive that GPU utilization up five points, 10 points."

Five to ten points of utilisation is a large claim and a checkable one. On a cluster whose accelerators cost more than the building they sit in, a ten-point improvement is the difference between buying the next tranche of hardware this year or next.

It is also the sort of number that depends entirely on the workload. A training run that is bound by computation will not improve much; an inference service that is bound by memory traffic and collective operations may improve a great deal. Anyone evaluating this should measure it on their own serving mix rather than on the vendor's.

Standards rather than silicon

The strategic position is open standards. For rack-scale interconnect, Cornelis is backing UALink and ESUN rather than Nvidia's proprietary NVLink, and its fabric is meant to work with a variety of accelerators rather than one vendor's.

This is the standard argument from every challenger in this market. The alternative is that the company selling the accelerators also sells the only fast way to connect them, which gives it two margins and the customer one supplier.