Supermicro to Build and Sell SambaNova Inference Systems Paired With Nvidia GPUs

Supermicro will integrate, validate, deliver and support the systems, including on-site service, SambaNova said. SambaNova engineers will tune AI models to run on the SN50 and work with Supermicro on customer deployments. The Next Web reported that the racks will join Supermicro's Data Center Building Block Solutions, its range of pre-configured data centre systems.

The collaboration starts with an initial batch of 10,000 SN50 chips. SambaNova said the rollout will be phased. Neoclouds and sovereign clouds, which keep data and compute inside one country, come first. Enterprises that run AI in their own data centres come later. SambaNova said it expects the partnership to bring significantly more high-performance inference capacity online in 2027. Neither company disclosed pricing or financial terms, and the 10,000-chip figure is an initial plan, not a count of delivered systems.

Splitting the Workload

The systems use an approach known as disaggregated inference. A large language model answers a request in two phases. In prefill it processes the prompt, and in decode it generates the answer one token at a time. SambaNova says prefill is limited mainly by compute, where Nvidia GPUs are strong. Decode is limited by memory, which its Reconfigurable Dataflow Units are designed to handle. The SN50 is the company's fifth-generation RDU and is built for inference.

SambaNova says agentic AI workloads spend most of their time in decode, which makes them a good fit for its hardware. "Agentic AI is the next wave of inference, and it has to run everywhere," said Rodrigo Liang, SambaNova's co-founder and chief executive, according to The Next Web. "Now customers can order SN50 RDUs alongside GPUs, from a world-leading partner they trust."

Supermicro president and chief executive Charles Liang said in SambaNova's announcement that the collaboration "expands our broad DCBBS portfolio by enabling heterogeneous architectures that integrate GPU computing with specialized inference acceleration."

Power is a large part of the pitch. SambaNova said the previous week that a SambaRack SN50 draws an average of 20 kilowatts, The Next Web reported. The company says its chips can serve the largest models without extreme power density per rack. That would let operators add inference capacity in existing data centres instead of waiting for new facilities held up by power and cooling limits, and keep their current GPUs in service longer.

Why It Matters

This is the latest of several commercial deals for SambaNova in recent weeks. At its AI World conference on October 27, Oracle said it plans to deploy SN50 chips in its cloud data centres and offer the capacity to AI model builders. On October 29, SambaNova signed services firm Emergys to deploy and run its platform for enterprise customers, according to The Next Web. The San Jose company was founded in 2017 and raised funding at an $11 billion valuation in July.

The arrangement removes a common obstacle for companies trying to compete with Nvidia. Buyers can now order a specialised inference chip from a server vendor they already use, with that vendor's support behind it, instead of buying direct from a start-up. And because the system is built around Nvidia GPUs for the prefill stage, it adds to existing Nvidia deployments rather than replacing them.

The early focus on sovereign clouds also matters for governments building national AI capacity. For many of these projects, power, cooling and the cost of running models now count for as much as raw training performance.