SambaNova Systems, Inc. is an American company specializing in artificial intelligence infrastructure, based in Palo Alto, California. It designs a computing architecture dedicated to neural networks and the software needed to operate it. Its positioning is built on tight integration between hardware and software: rather than selling only a chip, SambaNova offers an environment for training, adapting or running models. Its official website is sambanova.ai.
Origins rooted in Stanford research
Founded in 2017, SambaNova brings together three co-founders with complementary backgrounds: Rodrigo Liang, a former semiconductor executive, and two Stanford researchers, Kunle Olukotun and Christopher Ré. Liang brings his experience in processor development and bringing processors into production; the researchers contribute their work on computer architectures and machine learning. The company was built around a question: how can computing be adapted to the needs of models, rather than forcing those models to fit a general-purpose architecture?
This ambition attracted several institutional and corporate investors, including SoftBank Vision Fund 2, Intel Capital and GV. Their backing supported the development of an integrated offering in a sector where chip design, software and commercialization require substantial resources. SambaNova has gradually broadened its messaging, initially focused on computing systems, to encompass enterprise uses of generative AI.
From specialized chips to inference services
SambaNova’s technological foundation is the Reconfigurable Dataflow Unit, or RDU. This processor organizes processing around data flows and a reconfigurable architecture. With its SambaFlow software, the company aims to optimize model execution and data movement, a critical factor in performance and energy consumption. Its DataScale systems combine these components into infrastructure designed for AI workloads.
The offering has also expanded around inference, meaning the use of an already trained model to produce a response. Launched in 2024, SambaNova Cloud provides access to models via an API, without installing the corresponding equipment on the customer’s premises. In particular, the company offers the ability to run open-weight models, including those in Meta’s Llama family. Depending on the offering, it also serves organizations looking to deploy dedicated infrastructure and gain greater control over their data. Its commercial promise is as much about access to computing as it is about reducing integration complexity.
What now?
For SambaNova, the challenge is to turn its hardware differentiation into sustained adoption. Against the GPU ecosystem and other inference specialists, speed alone is not enough: usage costs, availability, software compatibility and ease of operation will influence customers’ choices. The advancement of open models may favor its approach, but requires continuous adaptation. Its ability to document performance under real-world conditions and support deployments will be crucial to expanding its role in enterprise AI.