Qualcomm and Cornelis Want AI Networks to Do More Than Move Data

AI clusters do not only need faster accelerators; they also need to exchange data quickly and predictably. Qualcomm’s newly announced collaboration with Cornelis Networks is aimed at that constraint: improving the role of networking in efficient AI infrastructure, particularly as inference and heterogeneous clusters scale.

The important qualification is that the announcement is a strategic collaboration and a joint appearance at AI Infra Summit, not the launch of a jointly shipping Qualcomm–Cornelis product. Qualcomm says Cornelis’ vision of an open, programmable fabric aligns with its interest in improving AI utilization and economics. 57

The central idea: turn the network into an active compute layer

Cornelis calls its approach Active Compute Fabric. Rather than treating switches and network interfaces solely as devices that forward packets, the architecture puts programmable compute and acceleration into the fabric. Cornelis says this allows some communication and collective-work processing to occur as data moves through the network, instead of placing all of that work on hosts or AI accelerators. 515

The architecture spans two related networking jobs:

• Scale-out: connecting servers or nodes across a cluster.
• Scale-up: connecting accelerators more tightly within a rack-scale system.

Its stated goal is to make data movement less likely to leave expensive GPUs or other accelerators waiting for inputs. That is especially relevant to distributed AI workloads, where collective communication and many-to-one traffic patterns can become a bottleneck. 121

How lossless transport and in-fabric compute are supposed to help

Cornelis combines three elements in Active Compute Fabric: lossless transport, in-fabric acceleration and programmable compute. 5

For its Omni-Path-based CN5000 platform, Cornelis describes a credit-based lossless transport design, fine-grained adaptive routing, incast-aware flow control and link-level replay. In practical terms, the company says these mechanisms are designed to prevent buffer overruns and congestion from disrupting data delivery. 18

The potential payoff is straightforward: if communication completes more reliably and with less accelerator-side overhead, a larger share of accelerator time can go to model execution. That could improve the effective economics of training or serving models, since the networking layer helps protect utilization of much costlier compute hardware.

That is a design thesis, not a settled benchmark result. Cornelis publishes performance comparisons and utilization claims, but operators should evaluate them against their own models, collective patterns, software stack and cluster topology. 15

What the product roadmap says today

Cornelis’ current and planned products matter because the architecture is not all at the same maturity level.

• CN5000 is a shipping 400 Gb/s, end-to-end Omni-Path scale-out fabric comprising SuperNICs, switches, software and cabling. Cornelis positions it as a lossless, congestion-controlled system for AI and HPC. 17
• CN6000 is the next-generation intelligent networking platform for AI training, inference and HPC. Reporting says it adds Ethernet support and is being prepared for broader availability, while Cornelis describes it as part of its next generation of scale-out products. 68
• CN7000 is the planned native scale-up platform. It is the roadmap component most relevant to a direct alternative for tightly coupled, NVLink-style rack-scale interconnects. It is not yet a broadly deployed replacement for Nvidia’s scale-up technology. 219

This distinction is crucial. Cornelis already has a commercial scale-out product, but its fullest scale-up proposition remains future-facing.

Why openness and heterogeneity are part of the pitch

Cornelis positions Active Compute Fabric as an open architecture across scale-up and scale-out environments. Its roadmap references Ethernet, Ultra Ethernet, RoCE and UALink-related technologies rather than a network tied to one accelerator supplier. 1119

That is attractive to infrastructure operators seeking flexibility: in principle, they could pair the fabric with different GPUs, CPUs, custom accelerators and other components instead of choosing one vertically integrated compute-and-networking stack. Cornelis also says CN5000 interoperates with accelerators from AMD, Intel, Nvidia and others. 23

The trade-off is execution. Open, heterogeneous infrastructure has to demonstrate robust interoperability, available hardware volume and mature software across the full system—not merely compatibility at the component level.

How it fits Qualcomm’s data-center strategy

The Cornelis relationship complements Qualcomm’s larger effort to become a supplier of AI data-center compute and connectivity.

Qualcomm and Amazon have announced a multigeneration collaboration around customized silicon for large-scale AI inference and optical connectivity reaching up to 1.6 Tb/s. Reuters reported that Amazon could purchase up to $60 billion in Qualcomm AI data-center chips and related products under the long-term arrangement, subject to its terms. 3324

Qualcomm also completed its acquisition of Alphawave Semi, adding high-speed wired-connectivity technology that it says complements its Oryon CPU and Hexagon NPU processors. 44

Together, those moves suggest a broader strategy: participate in both the compute and connectivity pieces of AI infrastructure. Cornelis adds a possible open-networking dimension to that strategy, but the companies have not announced a complete integrated platform.

The competitive reality: a credible direction, not a displacement yet

Cornelis is entering territory where Nvidia has established products and a deeply integrated accelerator, networking and software ecosystem. Active Compute Fabric’s appeal is modularity: a network designed to contribute computation and operate across heterogeneous hardware rather than lock the buyer into a single proprietary fabric.

Cornelis raised $205 million to advance its scale-up and scale-out products, manufacturing and deployments, giving it additional resources to pursue that ambition. 57

Still, funding and architecture announcements do not establish market equivalence. The case for Cornelis and Qualcomm will depend on delivered CN6000 and CN7000 products, software support, interoperability, supply and reproducible production results. For now, the collaboration is best understood as a bet that better networking can raise AI infrastructure utilization—and that customers will value an open alternative as much as peak performance.