Infinity, an Inference Startup, Secures $15M Investment from Touring Capital and Experts at OpenAI and Anthropic
Infinity, an AI infrastructure firm, revealed on Monday that it has successfully raised $15 million at a valuation of $100 million, with investments coming from Touring Capital, Principal VC, and researchers associated with OpenAI and Anthropic.
The company is focused on developing software aimed at improving how AI chips execute AI models. Nvidia’s success as a leading player is attributed not only to its high-performance chips but also to its CUDA software (Compute Unified Device Architecture). This software enables Nvidia’s GPUs, initially intended for graphics processing, to function as general-purpose CPUs. Major AI frameworks like PyTorch and TensorFlow utilize CUDA, allowing developers to create applications in widely-used programming languages such as Python, ensuring these applications inherently operate on Nvidia hardware.
Many startups specializing in app development lack the expertise or resources to create their own kernels—low-level software that governs chip functionality—and adapt their applications for various AI chips. Consequently, Infinity aims to produce kernel software that serves as an alternative to CUDA and can operate across diverse chip types, including SRAM, GPUs, mobile processors, and Systolic Arrays. Infinity represents a growing number of startups that are working strategically to diminish Nvidia’s market hold.
Infinity seeks to create a universal inference library that can function on all chip types, which would enable the automation of replicating cutting-edge research findings.
The company was founded last year by Jeremy Nixon, a former Google Brain researcher and the founder of the AGI House hacker network. Nixon explained his motivations for establishing this venture stemmed from his fascination with “automated invention”—the conviction that AI can act as a catalyzing technology. He developed a machine learning algorithm named Omega, capable of generating new algorithms and assessing them in a self-improving feedback loop.
This achievement led him to consider its applications in hardware, convinced that automated systems could also generate essential low-level codes, such as kernels, to enhance chip performance.
Infinity’s AI research agent, Ignition, is designed to write the necessary low-level code for AI inference on chips that are alternatives to Nvidia. It tests, debugs, and monitors hardware performance with the code, making automatic adjustments to optimize results. This self-optimizing system continually learns and adapts to various chip architectures, according to Nixon. The end product is what Infinity claims to be comparable to a CUDA-level software stack.
Among its clients is D-Matrix, an AI chip manufacturer that aims to rival Nvidia, and the company is in discussions with several major chip and cloud firms.
While the AI agent handles the bulk of the work, human experts provide overarching guidance. A case study illustrated that the agent can perform tasks significantly quicker than a human alone, compressing a process that could take years into a matter of hours or days. Instead of charging an upfront licensing fee, Infinity takes a share of the performance improvements and cost efficiencies, gauged by variations in tokens per second.
Currently, Infinity employs a team of 26, with roles spanning design, operations, and engineering.
