SambaNova SN50 chips to ship in Supermicro racks
The GPU-plus-RDU split separates prefill from decode workloads instead of adding more of the same hardware, a design that could reshape how inference clusters are procured. Data centre operators and inference platform teams should compare SN50 rack economics against GPU-only deployments when planning 2027 capacity. The partnership gives SambaNova a distribution channel through Supermicro rather than direct sales.
SambaNova announced a partnership with Supermicro to build disaggregated inference systems that combine NVIDIA GPUs with SN50 RDUs, starting with 10,000 chips. In the split architecture, NVIDIA GPUs handle prompt processing while SN50 chips generate the answer. Supermicro will build, sell and support the systems, which run on standard power and cooling in existing data centers. The company claims the setup adds more inference capacity in 2027 than a GPU-only build-out schedule would allow.
SambaNova SN50 chips are coming to @Supermicro racks!
Supermicro will build, sell and support disaggregated inference systems that pair NVIDIA GPUs with SN50 RDUs, starting with 10,000 chips. GPUs read the prompt, SN50 writes the answer fast.
Standard power, standard cooling, existing data centers. More inference capacity in 2027 than the build-out schedule would otherwise allow.
🔗Full press release here: https://t.co/klTYqBsR8y
