Supermicro will build and sell AI systems that pair SambaNova inference chips with Nvidia GPUs, according to The Next Web. The collaboration initially plans for 10,000 SN50 chips, with the companies announcing the partnership on Tuesday.
Key takeaways
- Supermicro and SambaNova are partnering to build new AI inference racks.
- The initial plan covers 10,000 SambaNova SN50 chips.
- Nvidia GPUs handle the reading stage while SN50 chips handle writing.
- Operators can fit the 20 kW racks into existing data centres without waiting for new builds.
How do these new inference chips work?
Running an AI model involves two distinct stages. The system first reads the prompt before writing the answer piece by piece. In the newly designed racks, Nvidia GPUs handle the reading while SambaNova SN50 chips take care of the writing.
Rodrigo Liang, co-founder and chief executive of SambaNova, stated that agentic AI requires fast responses. “Agentic AI is the next wave of inference, and it has to run everywhere,” he said. Supermicro will assemble, test, deliver, and support these systems, while SambaNova engineers will tune AI models for the SN50.
Where will the supermicro AI systems go first?
The companies plan to target neoclouds and sovereign clouds initially. Sovereign clouds keep data and computing inside a single country. After those initial deployments, the partners plan to sell to companies running AI workloads in private data centres.
SambaNova points out that the SN50 is its fifth-generation Reconfigurable Dataflow Unit. A SambaRack SN50 draws an average of 20 kW. This power profile lets operators add capacity in current data centres without waiting for new facilities constrained by power and cooling limits.
What happens next?
The partners plan to deliver much more inference capacity in 2027. This agreement follows recent deals for SambaNova, including an Oracle announcement at AI World on October 27 to deploy SN50 chips in cloud data centres, and an agreement with Emergys on October 29. SambaNova raised funding at an $11 billion valuation in July and previously worked with Foxconn and Intel in June.
Frequently asked questions
Supermicro is pairing SambaNova SN50 inference chips with Nvidia GPUs in its new ready-made data centre systems.
The collaboration initially plans for 10,000 SambaNova SN50 chips, according to the company's announcement.
Nvidia GPUs handle the reading stage of an AI model prompt, while SambaNova SN50 chips handle writing the answer piece by piece.
A SambaRack SN50 draws an average of 20 kW, allowing operators to add capacity in existing data centres.
Supermicro will put the systems together, test them, deliver them, and provide customer support including on-site service.
Based on reporting by The Next Web. Read the original article
