SambaNova Systems

SambaNova Systems is an American artificial intelligence infrastructure company that designs its own processors, systems and cloud service for AI inference — the stage of AI where trained models actually respond to requests. Founded in 2017 in Palo Alto and valued at US$11 billion after a US$1 billion financing round in July 2026, it is one of the best-funded independent challengers to Nvidia's dominance of AI compute.^[1][2]

History

SambaNova was co-founded by Stanford professor Kunle Olukotun, a pioneer of chip multithreading, together with industry executives including Rodrigo Liang, now chief executive. The company's premise was that general-purpose GPUs were a poor fit for the shape of AI workloads, and that a dataflow architecture — where the chip's structure follows the flow of data through the neural network — could deliver far better performance per watt.^[1]

The startup raised successive large rounds from investors including SoftBank, and in February 2026 announced more than US$350 million in new funding together with a strategic partnership with Intel, struck after earlier buyout talks with Intel had reportedly failed. In July 2026 SambaNova completed the first close of a US$1 billion financing led by General Atlantic at an US$11 billion valuation, bringing cumulative funding to roughly US$2.5 billion. Management has said an initial public offering is under active consideration.^[1][2]

Key Concepts and Technology

SambaNova's stack has three layers. At the bottom is the Reconfigurable Dataflow Unit (RDU), its custom processor, now in its fourth generation with the SN40L chip and the SambaRack SN40-16 system optimised for low-power inference. Above that is software that compiles models onto the dataflow hardware, and at the top is a cloud-style service that sells inference capacity by the token.^[1]

The company's recent strategy is squarely aimed at inference economics: as model usage grows, the cost and latency of serving responses becomes the dominant expense, and SambaRack systems are positioned as an alternative that produces more tokens per dollar and per megawatt than GPU clusters. Deployments include on-premises systems at JPMorgan Chase, United States Department of Energy national laboratories and Saudi Aramco, alongside neocloud operators that resell its capacity.^[1][3]

Applications and Impact

SambaNova's customers use its platform for enterprise search, document processing, financial analysis, and serving large language models to internal users at scale. Its business model matters as much as its chips: by selling complete racks plus a hosted inference endpoint, it competes with Nvidia not only on silicon but on the whole delivery of AI as a utility. The company is frequently grouped with Groq, Cerebras and CoreWeave in the emerging AI neocloud category — infrastructure firms whose entire reason to exist is AI workloads.^[3]

>See Also

• Nvidia
• Groq
• Cerebras Systems
• AI Chip Export Controls
• AI Data Centres in Malaysia