Supermicro will sell AI systems pairing Nvidia GPUs with SambaNova chips

Supermicro will build and sell AI systems that combine Nvidia GPUs with SambaNova’s specialized SN50 chips, SambaNova said on November 3. The partnership puts integration and customer support behind a mixed-hardware design for running AI agents, with a phased rollout beginning with 10,000 SN50 chips rather than an immediate deployment across all customer groups.

Two processors, two parts of an answer

The design is built around inference: the work an AI model does to process a request and generate an answer. Rather than assign that entire job to one type of processor, the proposed systems divide it between Nvidia GPUs and SambaNova’s SN50 RDUs, its specialized inference processors.

SambaNova describes two stages with different demands. Prefill processes the incoming prompt and depends heavily on computing power, which the company identifies as a strength of Nvidia GPUs. Decode produces the response one token, or piece of text, at a time. That stage depends more on memory, the task SambaNova says its SN50 chips were designed to handle.

The company’s argument is therefore not that customers should replace GPUs outright. It is that each processor should handle the phase it suits best. SambaNova claims this split, called disaggregated inference, delivers faster and more efficient results than either hardware type alone. Its announcement provides no numerical comparison for that claim.

Agents are the target workload

SambaNova ties the architecture to AI agents, which repeatedly call language models while carrying out tasks. It says agents can make thousands of model calls, with individual calls taking minutes. In the company’s account, agent workloads spend most of their processing time in decode, making response generation the focus of its acceleration pitch.

The deployment pitch also concerns the buildings that house the equipment. SambaNova says SN50 processors can run large models at high speed and throughput without requiring extremely high power density per rack. It argues that this lets customers use existing data centers, rather than making unusually dense rack power a condition of deployment.

Supermicro handles the system; SambaNova tunes the models

Supermicro’s role extends beyond selling the chips. It will integrate and validate the combined systems, deliver them and provide support, including on-site service. SambaNova engineers will work alongside Supermicro on customer deployments and tune models for performance on SN50 processors. The arrangement assigns both the hardware assembly work and the model-specific work to named partners.

The first recipients will be neoclouds and sovereign clouds. Enterprises running AI in their own data centers will follow. The initial 10,000-chip batch describes the starting scale of the rollout, not a count of completed customer installations. The announcement does not specify shipment dates for those customer groups or disclose system pricing.

SambaNova expects the partnership to produce significantly more deployed high-performance inference capacity in 2027 than would otherwise be possible. That is a forecast about what the sales and deployment channel can enable, not a disclosed capacity commitment. The concrete commitment today is the division of responsibilities for getting the combined systems to customers.