Oxmiq raises $35M to develop AI chip architecture aimed at lowering training and inference costs

The startup says its architecture is designed to run CUDA programs without Nvidia hardware, adding a potential new route for AI chip competition beyond incumbent GPU suppliers.

Summary

Oxmiq has raised $35 million to develop chip and system architectures aimed at cutting the cost of AI training and inference, while also targeting compatibility with CUDA programs without relying on Nvidia hardware. The company has said it plans a unified chip that combines GPU, CPU and tensor engine functions with a chiplet-and-memory compute fabric, positioning the effort as a broader rethink of AI infrastructure rather than a conventional standalone GPU push. The update adds a potentially important competitive angle because CUDA has long been central to Nvidia's AI ecosystem, and the ability to run CUDA workloads on alternative hardware could widen access to AI computing and challenge incumbent dominance if Oxmiq can execute on its design goals.

Terms & Concepts
  • CUDA: A software platform and programming model widely used to run accelerated computing workloads, especially on AI chips.
  • chiplet-and-memory compute fabric: A design approach that links smaller chip components and memory into a broader computing system.
  • inference: The stage where a trained AI model is used to generate predictions or responses.