Cornelis, an AI infrastructure company focused on networking between compute chips, has raised $205 million in a round led by IAG Capital Partners. The company also introduced Active Compute Fabric, a networking product intended to reduce idle time when GPUs and other accelerators wait for data.
The funding reflects a broader shift in AI infrastructure: adding more chips is not enough if systems cannot move data efficiently among them. For operators building large AI clusters, networking performance can determine how much of an expensive accelerator fleet is doing useful work.
What changed
Cornelis said its Active Compute Fabric enables chips to process information and transmit it at the same time. The company’s premise is that conventional data movement can leave GPU capacity underused while systems wait for inputs, outputs, or coordination across a cluster.
Cornelis spun out of Intel in 2020 and has begun shipping its product. It is also developing a new generation that it expects to release later this year.
The $205 million round gives the company capital to pursue a difficult infrastructure market, where buyers demand hardware and software that work reliably at scale and fit into existing data-center operations.
Why networking is becoming a strategic layer
AI workloads increasingly run across many accelerators rather than on a single chip. Training and serving large models require those devices to exchange data frequently, which makes the interconnect between chips a core component of overall system performance.
That creates an opening for vendors that can improve utilization without requiring customers to replace their full compute fleet. In practical terms, a networking improvement can matter if it lets an organization extract more productive work from accelerators it already owns.
Cornelis is positioning its architecture as open: customers can use its networking fabric with different GPUs and accelerator hardware. That approach may appeal to enterprises and cloud operators seeking flexibility as they evaluate a wider mix of AI chips.
The Nvidia challenge is bigger than a chip
Cornelis is competing in a market shaped heavily by Nvidia’s integrated stack. Nvidia’s chips can run on other networking fabrics, but they are optimized for Nvidia’s own software, making the company’s end-to-end platform easier for many customers to deploy.
That distinction is important. Breaking Nvidia’s position does not necessarily require a direct replacement for its GPUs. Infrastructure suppliers can instead compete on individual layers, including networking, software, power, storage, and cluster management.
For Cornelis, an open-networking pitch will need to translate into measurable operational value: higher accelerator utilization, compatible deployments and manageable complexity. Buyers will weigh those benefits against the simplicity and maturity of an integrated Nvidia environment.
What to watch next
The first question is customer adoption beyond initial shipments. Large AI deployments tend to have long qualification cycles, and networking products must prove performance, reliability and interoperability under production workloads.
The planned next-generation product will also be a key milestone. Cornelis will need to show that its roadmap keeps pace with increasingly demanding multi-accelerator systems while supporting the heterogeneous hardware strategy it is selling.
More broadly, the round is another signal that AI infrastructure competition is moving beyond the accelerator itself. As compute clusters grow, the companies that control how chips communicate could have an outsized influence on the economics of AI deployment.
