Cornelis and NextSilicon to Build Joint Reference Architectures for AI and HPC

Key Takeaways

• Cornelis CN5000 400 Gbps fabric pairs with NextSilicon Maverick‑2 accelerator
• Joint reference designs aim to eliminate network‑induced idle compute
• Upcoming CN6000 800 Gbps fabric slated for H2 2026 testing
• Maverick‑2 claims up to 10× GPU performance at half power
• Collaboration targets disaggregated AI inference and Mixture‑of‑Experts models

Pulse Analysis

The AI and high‑performance computing markets have converged on workloads that generate a torrent of tiny, latency‑sensitive messages. Traditional Ethernet fabrics, designed for bulk traffic, become a choke point, forcing expensive GPUs and CPUs to sit idle while waiting for data. Cornelis’s CN5000 fabric, launched in 2025, tackles this problem with a congestion‑free, lossless architecture that delivers 400 Gbps of deterministic bandwidth. Paired with NextSilicon’s Maverick‑2 accelerator—built on an Intelligent Compute Architecture that reconfigures at runtime—the duo promises a holistic solution that keeps data flowing and compute busy.

From a systems‑integrator perspective, the value lies in a validated reference design that eliminates the guesswork of mixing untested components. Early joint evaluations already show that the combined stack can sustain higher utilization rates, translating into up to tenfold performance gains over conventional GPUs while consuming roughly half the power, according to NextSilicon. The roadmap includes scaling to the 800 Gbps CN6000 fabric in the second half of 2026, positioning the partnership to support emerging disaggregated inference workloads such as Mixture‑of‑Experts and agentic AI, where the network is part of the compute path.

Industry analysts see this collaboration as a strategic counter‑move to the dominant silicon vendors that rely on monolithic, CPU‑centric designs. By offering OEMs a ready‑to‑market blueprint, Cornelis and NextSilicon can accelerate time‑to‑revenue and lower total cost of ownership for data‑center operators. If the joint reference architectures deliver on their promises, they could reshape procurement decisions, encouraging a shift toward modular, software‑defined compute and networking stacks that adapt to evolving AI models. The partnership therefore not only addresses current bottlenecks but also sets a foundation for the next generation of scalable, energy‑efficient AI infrastructure.