Infinity raises $15M seed to make any AI chip inference-ready

Infinity is going after one of the nastiest bottlenecks in the AI stack: the years-long grind of building optimized inference software for new accelerators. By positioning Ignition as an autonomous AI agent that writes and tunes low-level kernels, Infinity is trying to turn something that looked like a permanent NVIDIA moat into an automated, vendor-agnostic service. If this works at scale, it chips away at CUDA’s lock-in and opens the door for a far more diverse ecosystem of AI hardware.

Strategically, that matters for the race to AGI because compute is the binding constraint. More viable chips, brought online faster, mean more aggregate training and inference capacity for frontier models. The fact that Infinity is already partnered with d‑Matrix and backed by chip executives and researchers from OpenAI and Anthropic signals that leading labs and hardware makers see software enablement as a first-order problem, not an afterthought. Longer term, AI agents optimizing the very code that runs AI models is a concrete step toward recursive improvement loops in the infrastructure layer.