Infinity said on Monday it has closed a $15 million seed round at a $100 million valuation to commercialize software that aims to make “any” AI chip inference-ready, reducing the need for manual kernel engineering. Investors include Touring Capital, Principal VC, executives from major chip firms, and researchers formerly at OpenAI and Anthropic.
## The CUDA bottleneck
Nvidia’s dominance in data-centre AI acceleration rests not just on raw silicon performance but on CUDA, the nearly two-decade-old software layer that underpins major frameworks such as PyTorch and TensorFlow. That integration means models written in Python typically run on Nvidia hardware by default, helping the company capture an estimated 80% share of the data-centre AI accelerator market. Competing chips from AMD, Qualcomm and cloud vendors can match Nvidia on paper, but the work of porting models—writing and optimizing low-level kernels—remains a costly barrier.
## Infinity’s approach: an agent that writes kernels
Infinity’s product, called Ignition, is an AI agent designed to generate, test and iteratively rewrite kernels so models can run efficiently on new silicon. According to the company, human engineers set goals and constraints while Ignition performs the repetitive development and tuning, ingesting performance feedback to improve generated code automatically.
## Early results, as reported by the company
The startup reported striking internal benchmarks. Working with the chip maker d-Matrix, Infinity says Ignition reached 92% of a new chip’s peak performance within 10 hours of first interacting with the hardware, and that three frontier models were running end to end on the device within 10 days. In another internal test, the company claims a model’s throughput rose from roughly 1,400 to more than 20,000 tokens per second in a single day—an improvement it says would surpass vLLM, a widely used open-source inference framework. Those figures are self-reported and have not been
