Nvidia is planning an investment in d-Matrix, a startup designing AI inference chips that compete with its own, according to The Information. The two companies are building a system in which d-Matrix chips run a small model that predicts a larger model’s answer, and Nvidia GPUs then check and accept those guesses, a technique that speeds up responses. The deal would be Nvidia’s first with an inference chip startup since its $20B Groq licensing agreement in December, and it follows multibillion-dollar stakes in Intel and Marvell made partly to get their chips working alongside Nvidia GPUs. Rivals that couldn’t displace Nvidia are increasingly building to run with its hardware, which also helps sell Nvidia’s networking gear.
