SambaNova Secures $350M Funding with Intel Backing to Challenge Nvidia in AI Inference

AI infrastructure provider SambaNova Systems has secured $350 million in Series F funding, with Intel Capital joining Vista Equity Partners, Cambium Capital, and other investors. This strategic investment terminates acquisition speculation while establishing a multi-year technical collaboration between SambaNova and Intel. The partnership will integrate Intel Xeon CPUs with SambaNova's proprietary Reconfigurable Dataflow Units (RDUs), positioning the architecture as a cost-efficient alternative to Nvidia GPUs for generative AI inference workloads.

SambaNova will deploy this capital to accelerate production of its fifth-generation SN50 accelerator, scheduled for release later this year with SoftBank as an early adopter. The SN50 represents a substantial upgrade over its predecessor, delivering 2.5× higher FP16 performance (1.6 petaFLOPS) and 5× higher FP8 performance (3.2 petaFLOPS). Its unique three-tier memory architecture remains intact, combining 432MB of on-chip SRAM, 64GB of HBM2E memory (1.8TB/s bandwidth), and configurable DDR5 memory ranging from 256GB to 2TB.

While spec comparisons show the SN50 delivering approximately 64% of Nvidia Blackwell's dense FP8 compute and less than a quarter of its memory bandwidth, SambaNova's dataflow architecture enables superior real-world efficiency. By overlapping computation and communication operations, the technology reduces data movement overhead. Benchmark results demonstrate the SN40L (previous generation) serving the 230B-parameter MiniMax M2 model at 378 tokens/second—over 100 tokens/second faster than competing GPU-based systems. The SN50 aims to extend this advantage with claims of 5× higher per-user generation speed versus Nvidia's B200.

For enterprises managing multiple AI models, SambaNova's architecture addresses critical operational challenges. Its memory hierarchy enables sub-second model switching and efficient key-value cache management, significantly improving rack-level utilization. CEO Rodrigo Liang emphasized the economic imperative: "As organizations deploy customized models, traditional racks run inefficiently. Our 2025 engineering focus ensured per-rack economics where service providers can profitably serve tokens."

The SN50 scales across 256 accelerators—3.5× more than Nvidia's NVL72 racks—using a switched fabric delivering 2.2TB/s bidirectional chip-to-chip bandwidth. Despite lower density (16 air-cooled RDUs per rack at 15-30kW), this design prioritizes operational flexibility over pure density. Enterprises evaluating inference solutions should note SambaNova's approach reduces dependency on scarce HBM supply chains while providing deterministic performance for mixed-model environments. With Intel's manufacturing scale and customer access now augmenting SambaNova's technology, organizations gain a validated alternative for cost-sensitive AI deployments.

Learn more about SambaNova's architecture | Intel AI solutions