Supermicro to Build Inference Racks Pairing SambaNova's SN50 Chips With Nvidia GPUs

The arrangement gives SambaNova, one of the start-ups trying to win inference work from Nvidia, access to the sales and service network of one of the largest builders of AI servers. For Supermicro, it adds a non-Nvidia accelerator to its catalogue without asking customers to give up GPUs they already own. The pitch is aimed at operators who want more inference capacity now but are short of the power and cooling that the newest GPU racks demand.

The systems split inference into two stages. Nvidia GPUs handle prefill, the compute-heavy step of reading the prompt, while SambaNova's Reconfigurable Dataflow Units handle decode, the memory-bound step of generating a response token by token. SambaNova argues that AI agents spend most of their time in decode, so a faster decode stage shortens each loop an agent runs.

"Agentic AI is the next wave of inference, and it has to run everywhere," Rodrigo Liang, SambaNova's co-founder and chief executive, said in the companies' press release, dated November 3. Charles Liang, Supermicro's president and chief executive, said the collaboration would add heterogeneous architectures to the company's portfolio and deliver "flexible, high-performance AI infrastructure".

The SN50 racks will sit in Supermicro's Data Center Building Block Solutions range. Supermicro will integrate, test, deliver and support the systems, including on-site service, while SambaNova engineers will tune models to run on SN50 and help with deployments. The companies will sell first to so-called neoclouds and sovereign cloud operators, and later to enterprises running AI in their own data centres.

SN50 is SambaNova's fifth-generation chip. The company says it can serve the largest models at high speed without the extreme rack power density of the latest GPU systems, which would let operators add capacity in existing buildings. SambaNova claims its SambaRack SN50, which is air-cooled, delivers five times the compute per accelerator and four times the network bandwidth of the previous generation; those figures have not been independently tested.

The announcement leaves several questions open. Both companies describe the 10,000-chip figure as a planning number for a phased rollout rather than a purchase commitment, and they disclosed no pricing, ship dates, customers or system-level performance and power figures, according to an analysis by StorageReview. The companies said only that the partnership should put substantially more high-performance inference capacity into service in 2027 than would otherwise be possible.

SambaNova, founded in 2017 and based in San Jose, made a first close of $1 billion in financing at an $11 billion valuation in July. JPMorganChase is deploying its SN40 and SN50 systems for on-premises inference, and the company also has a heterogeneous inference collaboration with Intel.