AI Infrastructure announced the completion of $15 million in financing, valued at $100 million. Participants included Touring Capital, Prince VC and researchers from OpenAI and Anthropic. The company is developing a software layer for AI reasoning in the hope of making models easier for different types of chips.

The direction of finance to the products

In the current AI industrial chain, Nvidia ' s advantage derives not only from chip performance but also from the CUDA software ecology. Developers usually prepare applications based on mainstream frameworks such as PyTorch and TensorFlow, which have been built on CUDA for a long time, so a large number of AI applications run by default on Nvidia chips.

Infinity wants to solve this dependency. The company tried to build a bottom-up software that could replace CUDA, so that more chips could more easily carry AI reasoning tasks. In its own words, the software fits different structures such as GPU, mobile chip, SRAM and Systolic Array.

Aim for cross-chip reasoning.

The goal of Infoality is to create a common reasoning library that allows different chips to run and replicate the results of the frontier AI research. For most application-level entrepreneurship companies, the preparation of their own kernels and the migration of their applications to other chips, often time-consuming and lacking in adequate engineering resources, is one of the important reasons why Nvidia software has long-term ecological excellence.

Jeremy Nixon, founder of the company, was a Google Brain researcher and founder of the AGI House community. He said that the idea of creating Infinity came from "automated invention." In his view, the AI system could generate not only models but also bottom codes to help the chips operate more efficiently.

Generate bottom code with AI proxy

One of the core products of Infinity is the AI Research Agent named Ignition. The system is used to develop the bottom code required for reasoning for non-Nvidia chips and to complete testing, debugging and performance measurements. If the result is not satisfactory, the system automatically rewrites the code to continue to improve the efficiency of the operation.

According to the company, the system has a continuous self-optimization capability and can adapt to different chip structures, including hardware with proprietary design. Infinity locates it as a software warehouse near the CUDA level.

Currently, Infinity's clients include AI chip company D-Matrix. Jeremy Nixon indicated that the company was also in contact with other large chip companies and cloud computing companies. The business model is not a pre-received software authorization fee, but is divided by performance enhancement and cost savings, and measures include the number of tokens per second.

Additional information:To date, the Infinity team has 26 people, covering design, operation and engineering positions.