Supermicro will build and sell AI inference systems that combine NVIDIA GPUs with SambaNova SN50 chips, the companies announced on November 3. According to Data Center Dynamics, Supermicro is initially planning for 10,000 SN50 reconfigurable dataflow units, or RDUs, to build and sell disaggregated inference systems.
The systems will pair NVIDIA GPUs with SambaNova RDUs, splitting the work of inference between two kinds of silicon. NVIDIA GPUs process the prompt, while the SN50 chips, built to generate responses quickly, handle the output. Supermicro will integrate, test, deliver and support the systems for customers, including on-site service, while SambaNova engineers ensure AI models are tuned to run on the SN50 and support Supermicro with deployments.
Customers will be reached in phases. Deployments start with neoclouds and sovereign cloud providers before expanding to enterprises that want to run AI in their own data centers. That sequence targets two fast-growing groups that are building AI capacity for customers who want options beyond the largest hyperscalers.
SambaNova CEO and co-founder Rodrigo Liang said agentic AI is the next wave of inference and has to run everywhere. He said Supermicro builds AI infrastructure for data centers globally, and that customers can now order SN50 RDUs alongside GPUs from a world-leading partner they trust, which is how fast, efficient inference gets to scale.
Charles Liang, president and CEO of Supermicro, said the collaboration expands the company’s broad Data Center Building Block Solutions portfolio by enabling heterogeneous architectures that integrate GPU computing with specialized inference acceleration, helping customers deploy efficient, scalable AI solutions.
Founded in 2017, SambaNova originally focused on training workloads and pivoted in 2025 to become an AI cloud services provider. Its Suite product lets companies train their own generative AI models based on popular foundation models, either on-premises or in the cloud. The company introduced SambaNova Cloud in September 2024, offering cloud-based inference on its SN40L chip, which launched in September 2023 and can run models with up to five trillion parameters.
For data center operators and AI cloud builders, the combination offers another option for serving inference workloads, a fast-growing share of AI infrastructure demand. Pairing with NVIDIA GPUs, rather than replacing them, also lets customers who have standardized on NVIDIA tooling add a specialized inference tier.
StorageReview explains the logic behind the split. Inference has two phases: prefill, which processes the prompt and is limited by compute, and decode, which generates the response one token at a time and is limited by memory. SambaNova assigns prefill to the NVIDIA GPUs and decode to the SN50. The company argues that agentic workloads spend most of their time in decode, because an agent makes thousands of model calls and each answer gates its next action, so a faster decode stage shortens the whole loop.
The scale is easier to picture with rack math. StorageReview notes that SambaRack SN50 carries 16 RDUs per rack, so a 10,000-chip plan works out to about 625 racks if Supermicro keeps that density. The SN50 rack is air-cooled and scales to 256 accelerators, and SambaNova says it offers five times the compute per accelerator and four times the network bandwidth of the previous generation. The company says that profile lets operators add inference capacity inside existing facilities and keep the GPUs they already own on prefill duty.
Both companies describe 10,000 chips as a planning figure for a phased rollout. Neither has disclosed a purchase commitment, pricing, a ship date or customer names, and SambaNova has not published throughput, latency or power figures for the combined systems. The racks join Supermicro’s Data Center Building Block Solutions catalog, which already includes NVIDIA Vera Rubin NVL72 racks, so buyers can order SN50 alongside GPUs from one vendor responsible for the whole system.
Leave a Reply