Infinity, a burgeoning AI infrastructure company, announced a significant funding round this week, securing $15 million at an impressive $100 million valuation. The investment comes from a consortium of backers including Touring Capital, Principal VC, and notably, individual researchers from leading AI organisations such as OpenAI and Anthropic. This financial injection is set to fuel Infinity's ambitious goal of developing universal software that can make AI chips more versatile and efficient, directly challenging the entrenched market position of giants like Nvidia.
At the core of Infinity's mission is the creation of software designed to simplify the process of running AI models on various types of AI chips. Nvidia's long-standing dominance in the AI hardware sector has been largely attributed not just to its high-performance GPUs, but also to its proprietary CUDA software. CUDA provides a crucial interface, allowing developers to build AI applications using popular frameworks like PyTorch and TensorFlow, which then run seamlessly on Nvidia chips. Infinity aims to develop a 'CUDA-alternative' kernel software that is compatible with a wider array of chip architectures, including SRAM, standard GPUs, phone chips, and Systolic Arrays.
The startup's innovative approach involves an AI research agent named Ignition. This agent is designed to automatically write the low-level code, or 'kernels', necessary for AI inference on a multitude of chips, particularly those not produced by Nvidia. Ignition tests, debugs, and measures hardware performance with its generated code, automatically rewriting and optimising it for improved efficiency. This self-optimising system continuously learns and adapts to different chip designs, irrespective of their proprietary nature, striving to create a software stack that Infinity claims rivals the capabilities of CUDA.
Infinity was founded last year by Jeremy Nixon, a former Google Brain researcher and the creator of the AGI House hacker network. Nixon's motivation stems from his concept of 'automated invention', where AI systems become a 'meta technology' capable of generating other technologies. His prior success with an algorithm called Omega, which created and evaluated new machine learning algorithms, led him to apply this philosophy to hardware, believing automated systems could also generate the low-level code required for optimal chip performance.
The company's business model is also forward-thinking, eschewing upfront licence fees. Instead, Infinity takes a share of the performance gains and cost savings achieved by its customers, measured in 'tokens per second'. Early adopters already include AI chip maker D-Matrix, a potential challenger to Nvidia, and Infinity is reportedly in discussions with other major chip and cloud companies. With a current team of 26 employees across design, operations, and engineering, Infinity represents a new wave of startups aiming to incrementally disrupt Nvidia's significant market share in AI infrastructure.