Infinity Raises $15M to Build AI-Chip Inference Software

Infinity is building low-level inference software for AI accelerators. Its Ignition agent is designed to write, test and optimize kernels—the hardware-specific routines that determine how efficiently a model uses a chip. The company presents this as a way for chipmakers to shorten the work required to support new models and hardware.

That promise remains a company claim rather than an independently benchmarked result across the market. TechCrunch reported that Infinity is working with d-Matrix and is in discussions with other chip and cloud companies. The same report said the startup has 26 employees and charges through a share of measured performance gains and cost savings instead of an upfront software license.

Nvidia's advantage includes CUDA and the mature software ecosystem around its hardware, not only chip performance. New accelerator vendors therefore face a software-enablement problem: developers need reliable frameworks, kernels and tooling before alternative hardware is useful in production. Infinity is betting that automated code generation and repeated hardware measurement can reduce that barrier.

For infrastructure teams, the round is a signal to watch rather than proof that a universal inference layer has arrived. The decisive evidence will be reproducible performance across multiple chips and models, the effort required to integrate generated kernels into production systems, and whether gains persist as architectures and model workloads change. The funding gives Infinity more capacity to pursue that validation, but the retrieved sources do not establish broad third-party benchmarks or general availability across every accelerator.