SambaNova to Pair SN50 RDUs with NVIDIA GPUs in Supermicro Systems

In an announcement from San Jose, California, on November 4, 2026, SambaNova said it had entered a strategic collaboration with Supermicro to develop and sell AI inference systems. The initial plan includes 10,000 SN50 RDU chips. The systems will pair SambaNova’s RDUs with NVIDIA GPUs. The companies aim to offer this shared infrastructure to cloud service providers first.

How will the GPU and RDU share tasks in the system?

Under the division of labor described in the announcement, NVIDIA GPUs will handle the “prefill” stage, while SambaNova SN50 chips will handle “decode.” This distributes different parts of the inference process across two types of processors. The approach relies on adding RDUs to systems with existing NVIDIA GPUs so they can work together.

SambaNova describes the SN50 as its fifth-generation RDU chip, designed for high-performance inference. According to the company, RDUs are intended to run large models at high speed and throughput without requiring excessive rack power density. SambaNova also says RDUs can be paired with existing GPUs, allowing data centers to continue using their current infrastructure.

SambaNova and Supermicro say this approach could deliver high-performance inference capacity in 2027 beyond what might otherwise be possible. This expectation is based on the companies’ announcement. No measured performance results, independent comparisons or specific speed improvements were shared.

How will deployment and delivery work?

Supermicro will be responsible for integrating, testing and delivering the systems for customers, as well as providing support that includes on-site service. SambaNova engineers will help run and tune models on the SN50 and assist with Supermicro’s deployments.

The companies plan to focus first on NeoCloud and sovereign cloud providers, then expand to businesses running AI workloads in their own data centers. However, the announcement did not disclose a sales start date, system pricing, supported models or available configurations. The initial plan for 10,000 chips does not, by itself, indicate how many customers the systems will reach or the delivery schedule.

What does this mean for architecture and visualization workflows?

For architecture firms, interior design teams and visualization professionals, this is not a direct feature for design software, but another option for AI inference infrastructure. Organizations running large models in data centers or providing cloud services could consider the approach of using GPUs and RDUs in the same system. However, the source contains no measurements or examples showing how this could change the speed or quality of architectural visualization production.

For a firm or service provider, any potential benefit could come from the backend infrastructure powering its AI service; the impact on actual workflows would depend on model support and compatibility with the existing hardware. The announcement provides no information about integration with architecture and 3D applications, software licenses or purchase costs. More technical and commercial details will therefore be needed before the system can be considered a ready-to-use solution for existing production pipelines.