Cornelis and NextSilicon Team on AI/HPC Reference Architectures

Cornelis and NextSilicon announced a collaboration to develop and evaluate joint reference architectures for artificial intelligence and high-performance computing systems. The effort combines Cornelis’ CN5000 400 Gbps fabric with NextSilicon’s Maverick-2 compute platform, with the companies already conducting joint evaluations aimed at commercialization through OEM partners. The initiative was announced ahead of ISC High Performance 2026 in Hamburg.  

The collaboration addresses two infrastructure constraints commonly cited in large-scale AI and HPC environments: network congestion and compute inefficiency. Cornelis positions its CN5000 fabric as a congestion-free interconnect designed for latency-sensitive AI and HPC traffic, while NextSilicon’s Maverick-2 accelerator employs a runtime-reconfigurable dataflow architecture that adapts execution resources to specific workloads. The first phase of the project focuses on validating fabric and compute combinations across multiple configurations to create tested reference designs for OEM partners. The companies also plan to evaluate Cornelis’ upcoming 800 Gbps CN6000 fabric, scheduled for availability in the second half of 2026.  

Looking ahead, the companies said they intend to focus on emerging AI inference architectures, including Mixture of Experts (MoE) models, agentic AI, and disaggregated inference. These workloads increasingly distribute inference tasks across multiple accelerators and network domains, making interconnect performance a critical component of overall system efficiency. Cornelis and NextSilicon said future evaluations will examine how congestion-free networking and adaptive compute architectures can support these distributed AI pipelines and inform future OEM system designs.  

• Cornelis CN5000 fabric operates at 400 Gbps; CN6000 is expected to support 800 Gbps.  
• Maverick-2 entered volume shipments in late 2025.  
• Initial goal is creation of validated AI and HPC reference architectures for OEM partners.  
• Future testing will target disaggregated inference, MoE models, and agentic AI workloads.  

“Operators keep telling us their most expensive systems sit idle, waiting on the network. We built the CN5000 to end that wait,” said Lisa Spelman, CEO of Cornelis. “NextSilicon challenges the same kind of assumption on the compute side, so this collaboration is a natural fit.”  

🌐 Analysis: As AI inference workloads become increasingly distributed across accelerators, memory pools, and specialized processing elements, network fabrics are moving from a supporting role to a critical component of application performance. The emphasis on disaggregated inference and agentic AI aligns with broader efforts across the industry to address latency and utilization challenges in next-generation AI systems.

🌐 Analysis: Cornelis continues to position its technology as an alternative to Ethernet-based AI fabrics and to NVIDIA’s InfiniBand-centric ecosystem. Meanwhile, NextSilicon is targeting a niche between traditional CPUs and GPUs with its software-defined dataflow architecture. The collaboration suggests OEMs are looking for differentiated system architectures that combine specialized interconnects and workload-adaptive compute platforms rather than relying solely on conventional GPU-centric designs.