SambaNova and Super Micro target 10,000 SN50 chips for AI inference build

SambaNova Systems and Super Micro Computer are teaming up to build AI inference systems stuffed with 10,000 of SambaNova's SN50 chips. The two companies haven't said when that hardware actually lands, so the timeline stays a mystery for now.

The SN50 is the newest generation of SambaNova's Reconfigurable Dataflow Unit, the company's in-house name for its custom AI processor. It launched in February 2026 and is aimed squarely at inference — the part where a trained model answers a question, writes text or spits out an image for a user. SambaNova claims the chip can handle models with up to 10 trillion parameters.

That parameter claim is the setup. The punch is aimed at Nvidia: SambaNova says the SN50 runs up to 5x faster and costs up to 3x less to own than Nvidia's B200 GPUs. Big talk, and the Super Micro build is how SambaNova gets to prove it outside a lab.

The hardware math gets wild fast. SambaRack SN50 systems hold 8 to 16 RDUs per rack and pull roughly 20 kW per rack on average, scaling to clusters of as many as 256 RDUs tied together by multi-terabyte-per-second interconnects. Crunch the numbers and 10,000 chips works out to somewhere between 625 and 1,250 racks depending on configuration — a rough sketch, not a disclosed build plan. The upside: the racks are designed to slot into existing data centers without liquid cooling.

Shipping for SN50 systems is set for the second half of 2026. SoftBank is lined up for the first deployments in Japan, and SambaNova also has a multi-year collaboration with Intel on integrating with Xeon processors. Oracle Cloud Infrastructure is on the list as well, in a partnership aimed at spreading high-performance inference across cloud environments.

The push comes with fresh cash behind it. SambaNova raised more than $350 million in a Series E round in February 2026, with Intel Capital among the participants.