What does this mean: cornelis Networks raised 205 million dollars led by IAG Capital Partners and announced Active Compute Fabric alongside a Qualcomm collaboration.

Cornelis Networks raised 205 million dollars led by IAG Capital Partners and announced Active Compute Fabric alongside a Qualcomm collaboration.

What does this mean: the CN7000 adds RISC-V cores and SRAM to every NIC and network switch controller, spreading compute across the fabric itself.

The CN7000 adds RISC-V cores and SRAM to every NIC and network switch controller, spreading compute across the fabric itself.

What does this mean: target workloads are KV cache acceleration and mixture of experts routing, both of which currently spend GPU cycles on moving data.

Target workloads are KV cache acceleration and mixture of experts routing, both of which currently spend GPU cycles on moving data.

What does this mean: the AI hardware conversation is mostly about chips, while the bottleneck has moved to what sits between them.

The AI hardware conversation is mostly about chips, while the bottleneck has moved to what sits between them.

Computing

Cornelis put processors inside the network to stop GPUs sitting idle

16 September 2026 · 2 min read · By Future Technology

Key takeaways

Cornelis Networks raised 205 million dollars led by IAG Capital Partners and announced Active Compute Fabric alongside a Qualcomm collaboration.
The CN7000 adds RISC-V cores and SRAM to every NIC and network switch controller, spreading compute across the fabric itself.
Target workloads are KV cache acceleration and mixture of experts routing, both of which currently spend GPU cycles on moving data.
The AI hardware conversation is mostly about chips, while the bottleneck has moved to what sits between them.

A great deal of expensive GPU time is spent waiting for data to turn up. Cornelis Networks, an Intel spinoff, has raised 205 million dollars on the argument that the fix is to stop treating the network as plumbing and start treating it as compute.

The round was led by IAG Capital Partners and arrived with two announcements: a product called Active Compute Fabric, and a collaboration with Qualcomm.

What Active Compute Fabric does differently

This is not a smart NIC, and it is not a DPU. Those offload networking work from the host processor. Cornelis is doing something else with the CN7000, which adds RISC-V cores and SRAM to every network interface card and every network switch controller. The result distributes compute engines and memory across the fabric rather than parking them at the endpoints.

In plain terms, the network can process data while it is forwarding it, instead of acting purely as a courier between chips that do all the thinking.

The workloads it is aimed at

Two named targets, and both are specific rather than aspirational. KV cache acceleration deals with the attention state a model stores as a conversation grows, which has to move between devices and gets larger with context length. Mixture of experts routing sends each token to a subset of the model's experts, which means constant shuffling across the cluster.

Neither of those is arithmetic in any interesting sense. Both currently burn GPU cycles that could be doing inference. Move that handling into the fabric and the same racks produce more useful work from the same silicon. Inference is the clearest near term case, because that is where the cache traffic lives.

Why the money is going here now

The AI hardware conversation is almost entirely about chips, and the constraint has quietly shifted to the gaps between them. That is the same reading behind Lumilens raising 700 million for optical networking: when accelerators outrun the links joining them, the links become the product.

It also sits against enormous capital going into the silicon itself, from Samsung's 648 billion dollar programme to new fabs such as the Tesla and SpaceX Terafab in Texas. Those investments only pay off if the machines can talk to each other fast enough to stay busy.

What to watch

Cornelis has a funding round, a partner, and a described architecture. It does not yet have volume deployments anyone outside the announcement has measured. The number that matters is utilisation: how much idle GPU time this actually removes in a live cluster, reported by someone who did not build it. Until that lands, this is a credible read on where the bottleneck moved rather than proof that anyone has fixed it.