SambaNova launches new AI accelerator, teams with Intel to power inferencing and agentic

SambaNova’s SN50 accelerator marks the company’s first‑generation push into pure‑inference silicon, a market that industry observers have long viewed as the next growth frontier after training. According to Tom’s Hardware, the dual‑chiplet processor is built on SambaNova’s proprietary Reconfigurable Data Unit (RDU) architecture and pairs a three‑tier memory hierarchy—SRAM, HBM and DDR5—to keep multiple models resident for “hot‑swapping” without incurring latency penalties. The firm claims the SN50 delivers five‑times the compute per accelerator and four‑times the networking bandwidth of its predecessor, while consuming roughly one‑third the power of Nvidia’s B200 GPU. Those efficiency gains translate, the company says, into a three‑fold lower total cost of ownership for inference workloads, a metric that could be decisive for enterprises seeking to run large language‑model agents at scale.

The hardware announcement is tightly coupled with a multi‑year strategic partnership with Intel, which will supply Xeon CPUs as the host platform for SambaNova’s rack‑scale solutions. Tom’s Hardware notes that the collaboration is aimed at “building large‑scale AI inference infrastructure around Intel Xeon platforms and SambaNova AI accelerators,” positioning the SN50 as the compute engine within Intel‑based servers rather than as a stand‑alone card. Each 20 kW SambaRack SN50 chassis houses 16 RDU processors, and the design supports inter‑rack networking that can link up to 256 accelerators, delivering multi‑terabyte‑per‑second bandwidth across the system. By anchoring the accelerator to Xeon, SambaNova hopes to tap Intel’s extensive ecosystem of enterprise customers and government contracts, a move that could accelerate adoption in data centers that already standardize on Intel silicon.

Beyond the Intel tie‑up, SambaNova has secured a deployment agreement with SoftBank, which will install SN50‑powered racks in the Japanese conglomerate’s data‑center portfolio. The SoftBank deal, reported by Tom’s Hardware, underscores the company’s strategy of pairing its hardware with large‑scale operators that can showcase the low‑latency, agentic inference capabilities the SN50 is designed for. Rodrigo Liang, SambaNova’s co‑founder and CEO, emphasized that “AI is no longer a contest to build the biggest model” and that the real competition now centers on “who can light up entire data centers with AI agents that answer instantly, never stall, and do it at a cost that turns AI from an experiment into the most profitable engine in the cloud.” The SoftBank partnership provides a real‑world testbed for that claim, offering a path to demonstrate the accelerator’s performance in voice‑assistant, recommendation and other latency‑sensitive services.

Financially, the SN50 rollout arrives as SambaNova continues to raise capital for mass production. VentureBeat reported a $676 million financing round that valued the startup at over $5 billion, while earlier fundraising efforts of $450 million and $250 million have been earmarked for expanding both hardware and software stacks. The infusion of cash is intended to “expand its customer base—particularly in the datacenter,” according to the VentureBeat article, and to scale manufacturing of the SN50 and its associated SambaRack solutions. If the efficiency metrics hold up in production, the accelerator could undercut Nvidia’s B200 on both performance per watt and total cost of ownership, a proposition that may attract cost‑conscious cloud providers and enterprises wary of the premium pricing of GPU‑centric inference.

Analysts who have examined the announcement caution that the “five times more compute” claim is relative and not benchmarked against a specific competitor beyond the B200, as Tom’s Hardware points out. Moreover, the performance advantage hinges on the RDU’s ability to keep multiple models resident in memory—a capability that will be tested as models grow larger and more complex. Nonetheless, the combination of a purpose‑built inference chip, Intel Xeon integration, and a high‑profile SoftBank deployment gives SambaNova a differentiated value proposition that could reshape the economics of AI agent serving. Whether the SN50 can deliver on its promise of “turning AI from an experiment into the most profitable engine in the cloud” will depend on real‑world latency measurements and the willingness of data‑center operators to adopt a non‑GPU architecture at scale.