Supermicro, SambaNova plan 10,000-chip inference deployment

Supermicro’s SambaNova collaboration begins with a planned order of 10,000 fifth-generation SN50 Reconfigurable Dataflow Unit (RDU) chips for disaggregated AI inference systems sold by the partners.

The number marks an initial plan, not confirmed shipments. Supermicro will integrate the SN50 chips alongside NVIDIA GPUs. The announcement gives no delivery schedule, unit price, named customer or performance benchmark.

What does the partnership cover?

Supermicro will build and sell the systems, which will join its Data Center Building Block Solutions (DCBBS) portfolio for server and storage deployments across customer sites.

The framework creates optimised server and storage configurations. Disaggregated operation means dividing an inference request across processor types. Supermicro says the design should improve inference performance within the electricity and cooling limits of existing data centres.

RDUs are purpose-built for high-throughput inference and large-model serving without extreme rack power density. Rack power density means the concentration of electrical demand within a rack. The announcement provides no comparative power, latency or throughput figures for that claim.

How does the processor split work?

The design follows a two-stage pipeline, meaning the request passes between two specialised processors. The architecture is heterogeneous, meaning different processors handle tasks suited to their design.

The architecture assigns prompt preparation to NVIDIA GPUs and response generation to SambaNova RDUs, creating a heterogeneous path through each model request for inference workloads.

The partners expect this division of labour to raise overall system performance and inference efficiency. That expectation remains a design objective, not an independently validated result.

• Prefill: NVIDIA GPUs read and prepare the incoming AI prompt.
• Decode: SambaNova RDUs generate the model’s response.
• Objective: Combine GPU computing with specialised inference acceleration.

Charles Liang, Supermicro president and CEO, said:

“Supermicro is committed to delivering flexible, high-performance AI infrastructure that supports the evolving requirements of our customers. Our collaboration with SambaNova expands our broad DCBBS portfolio by enabling heterogeneous architectures that integrate GPU computing with specialised inference acceleration, helping customers deploy efficient, scalable AI solutions.”

The quotation sets out the intended architecture and commercial benefit. It does not supply the measurements needed to test either claim.

Can existing data centres absorb the racks?

The design addresses facilities that cannot support next-generation, high-density racks by adding RDU capacity while extending the useful life of GPUs already owned by operators.

Supermicro says operators can add inference capacity without waiting for next-generation data-centre builds constrained by power and cooling. It also claims that offloading inference extends the useful life of installed GPUs.

SambaNova and Supermicro expect substantially more high-performance inference capacity by 2027 than would otherwise be possible. The announcement defines neither the baseline nor the projected increase, so the 2027 claim cannot be sized. It also describes inference demand as soaring, but supplies no demand series or market forecast.

Rodrigo Liang, SambaNova co-founder and CEO, said:

“Agentic AI is the next wave of inference, and it has to run everywhere. Supermicro builds AI infrastructure for data centres globally. Now customers can order SN50 RDUs alongside GPUs, from a world-leading partner they trust.”

He presents the combined offer as a route to scaling fast, efficient inference. The reference to customer orders does not specify availability, volume or delivery timing. He says SN50 runs in existing data centres and complements GPUs already installed. That is a deployment proposition, not proof of comparative efficiency.

Who will sell, deploy and support the systems?

Supermicro controls integration, testing, delivery and service, while SambaNova supplies engineers to deploy and tune models on the SN50 hardware for customers during initial deployments.

Supermicro will manage the full customer lifecycle, covering integration, testing, delivery, ongoing support and on-site service. SambaNova engineers will load and tune models on SN50 chips and assist with customer deployments.

• Supermicro: Integration, testing, delivery, support and on-site service.
• SambaNova: Model deployment, tuning and customer assistance.
• Initial markets: NeoClouds and sovereign clouds.
• Expansion: Enterprises running AI workloads in on-premises data centres.

The initial go-to-market focus is NeoClouds, which the announcement describes as cloud service providers, and sovereign clouds. Enterprise deployments in customers’ own on-premises facilities are planned as the offering expands.

No launch date, named customer, pricing or volume beyond the 10,000-chip plan appears in the announcement. The market logic is straightforward: add inference hardware to constrained sites while retaining existing GPUs. The commercial proof remains pending. Those claims require shipped systems, workload benchmarks and disclosed pricing.