SambaNova Systems

Two of the founders are Stanford professors: Kunle Olukotun, one of the pioneers of multicore processors, and Christopher Ré from the machine learning side. Their chip, the RDU (Reconfigurable Dataflow Unit), is reconfigured so that a model's compute graph is laid out directly in hardware, with data streaming between units instead of round-tripping through shared memory.

SambaNova first sold whole rack-scale systems to labs and government agencies, then — like Groq and Cerebras — pivoted toward selling inference as a cloud service. That says something about the market: selling tokens per second is easier than persuading anyone to leave CUDA.