Infinity raises $15M to build CUDA alternative for AI chips

Infinity, an AI infrastructure startup, announced on Monday that it raised $15 million at a $100 million valuation. The round included Touring Capital, Principal VC, and researchers from companies including OpenAI and Anthropic.

The startup is building software that lets AI chips run AI models more efficiently. Nvidia’s dominance in AI hardware rests not just on its high-performance chips but on CUDA (Compute Unified Device Architecture), the software layer that lets its GPUs — originally built for graphics — act as general-purpose processors. The two leading AI development frameworks, PyTorch and TensorFlow, are both built on top of CUDA, so applications written in Python using those frameworks run on Nvidia chips by default.

Most application-layer startups lack the resources to write their own kernels — the low-level software that operates chips — and port their products to other AI chips. Infinity is trying to build CUDA-alternative kernel software designed to work across chip types, including SRAM, GPUs, phone chips, and systolic arrays, part of a broader wave of startups chipping away at Nvidia’s market position product by product. The company describes its goal as a universal inference library that would let any chip automatically replicate state-of-the-art research results.

Infinity was founded last year by Jeremy Nixon, a former Google Brain researcher and creator of the hacker network community AGI House. Nixon told TechCrunch he started the company because he was obsessed with the idea of automated invention — the belief that, in his words, “AI systems can actually be a meta technology.” He said he had previously invented Omega, a machine-learning algorithm that creates new machine-learning algorithms and automatically evaluates them in a feedback loop; that experience led him to apply the same automated approach to hardware, including the low-level code needed to run chips efficiently.

Infinity’s AI research agent, Ignition, is intended to write the low-level code needed for AI inference on Nvidia-alternative chips. It tests, debugs, and measures hardware performance, then automatically rewrites the code to improve it — a self-optimizing system that Nixon says adapts to different chip architectures, including proprietary ones. Infinity says the result is a CUDA-level software stack. Customers include AI chip maker D-Matrix, and the company is in talks with other large chip and cloud companies, according to Nixon.

Humans stay in the loop, setting high-level direction while the agent handles more of the routine work. In one case study, Nixon said, the agent worked much faster than a human alone, cutting a process that could have taken years or months down to hours or days. Infinity doesn’t charge an upfront license fee; instead, it takes a cut of the performance gains and cost savings it generates, measured in tokens per second. The company currently has 26 employees across design, operations, and engineering.

Why it matters

Infinity is part of a growing crowd of startups betting that software, not just silicon, can loosen Nvidia’s grip on AI infrastructure — a dynamic that could lower costs for any company running inference at scale, regardless of which chip vendor it picks.