Problem

AI-chip companies can design silicon with impressive theoretical performance yet struggle to make it useful for production inference because the required kernels, runtimes, and model integrations are difficult to build and optimize for every new architecture. Infinity’s investor describes this as a software-tooling bottleneck: alternative chips remain functionally sidelined against NVIDIA’s mature CUDA ecosystem, while hardware teams may spend months or years recreating the systems engineering needed to support modern models. The economic pain is delayed time-to-market, delayed time-to-revenue, and underutilized hardware. The clearest use case is taking a new accelerator—such as d-Matrix’s Corsair—from hardware access to production-quality large-language-model inference quickly.

Product / Service

Infinity provides an automated software layer for AI-chip inference. Its autonomous agent, Ignition, takes a chip’s hardware specifications, reverse-engineers optimization strategies, writes and tests low-level kernels, and generates a complete inference library for the target silicon. The system is designed to work across GPUs, NPUs, TPUs, custom ASICs, and edge accelerators, in whatever programming language the customer’s stack requires, and across text, vision, video, and audio models. Infinity claims a registry of more than 1,000 optimized kernels and roughly 10-times faster time-to-market than manual kernel development.

The delivery model is primarily B2B infrastructure for chip companies, with design partnerships and a cloud deployment that exposes the resulting inference capability. In the d-Matrix engagement, Ignition reached up to 92% of the chip’s empirical compute peak within 10 hours and had Qwen3, Qwen3.5, and Gemma4 running end-to-end within 10 days. The benefit is that hardware vendors can ship competitive, production-ready inference software without building a large specialized kernel-engineering organization from scratch; Infinity’s d-Matrix Cloud is an early live example.

Market

Infinity operates in AI infrastructure, specifically accelerator enablement, inference software, compiler/runtime tooling, and kernel optimization—the software layer between AI silicon and deployed models. Its central competitive reference point is NVIDIA’s CUDA ecosystem, whose advantage comes from decades of hardware-software co-optimization. Other substitutes include chip vendors’ internally built toolchains and general inference stacks such as vLLM; Infinity says its generated engine surpassed vLLM on traditional hardware. Public materials do not identify a single direct startup rival, so the competitive set is best understood as the incumbent CUDA stack, in-house enablement teams, and adjacent compiler and inference frameworks.

Infinity is an early-stage company founded in August 2025. As of July 2026, it had raised $15 million in seed funding at a $100 million post-money valuation, had Touring Capital and Principal VC participation plus strategic angel backing, and reported a live design partnership with d-Matrix alongside active discussions with other major chip companies. Its d-Matrix results and internal-preview cloud provide meaningful technical and commercial validation, but the cited materials do not disclose revenue; it should therefore be viewed as early commercial traction rather than a proven scaled-revenue business.