The Information reported that Nvidia is nearing a deal to buy AI startup Inception for at least $500 million, The Decoder said on 20 November. Three people familiar with the talks were its sources. Inception is two years old, based in the Bay Area, and has more than 30 employees. The deal reportedly includes performance bonuses for them. It builds language models using diffusion, a method from image generation. Its latest release is Mercury 3.
Neither company commented at first, so neither has confirmed why Nvidia wants the deal. The Decoder states, without naming a source, that the aim is to strengthen Nvidia's open-source Nemotron models and its robot models. It says that would help Nvidia compete with OpenAI and Anthropic, which both build their own inference chips. Co-founders Stefano Ermon, Aditya Grover and Volodymyr Kuleshov, all professors, are expected to join Nvidia as full-time staff. No closing date is given.
Most chatbots commit to each small piece of text, called a token, before writing the next. A diffusion model drafts the whole answer and revises it in passes. The reported deal is for that second way of generating text, not for chips. Buyers can ask vendors how their models generate text. The sources give no figures on speed or cost.