AI infrastructure startup Infinity has announced a $15 million funding round at a $100 million valuation. The round was led by Touring Capital and Principal VC, with participation from researchers at OpenAI and Anthropic. The company is building software that lets AI models run efficiently on chips other than Nvidia’s, which is a harder problem than it sounds.
Nvidia’s dominance in AI computing isn’t just about hardware. Its CUDA software platform is the real moat. Major AI frameworks like PyTorch and TensorFlow are built on top of CUDA, which means developers writing Python apps automatically get Nvidia optimization baked in. That has made switching to alternative chips expensive and technically painful for most companies.
Infinity is trying to fix that. It is building what it calls a universal inference library, a software layer that can run AI models on almost any chip architecture, from GPUs to phone chips to Systolic Arrays to SRAM. The goal is to give developers the same ease of use they get with CUDA, but without being locked into Nvidia hardware.
The company was founded by Jeremy Nixon, a former Google Brain researcher and the creator of AGI House, a well-known hacker community network. Nixon told TechCrunch he started Infinity because he became obsessed with the concept of “automated invention”, the idea that AI systems can act as a kind of meta-technology that generates other technologies. He had already built a machine learning algorithm called Omega that could create new machine learning algorithms and test them automatically in a feedback loop. That experience led him to ask whether the same approach could work for chip software.
The answer, apparently, is yes. Infinity built an AI research agent called Ignition specifically to write the low-level kernel code that chips need to run AI workloads. Kernels are the software closest to the hardware, and writing them is highly specialized work that most AI startups simply cannot afford to do in-house.
Ignition does the following automatically:
- Writes kernel code for a given chip architecture
- Tests and debugs that code
- Measures hardware performance
- Rewrites the code if it falls short, then repeats the cycle
The system is self-optimizing and, according to Nixon, works across different chip designs regardless of proprietary constraints. Humans still provide high-level direction, but the agent handles most of the repetitive technical work. In at least one case study, Infinity says Ignition compressed a process that would have taken months or years into hours or days.
Infinity’s pricing model is also unusual. It does not charge an upfront license fee. Instead, it takes a percentage of the performance gains and cost savings it delivers, measured in tokens per second. That structure aligns the company’s revenue directly with how much value it actually creates for customers.
The startup already has at least one notable customer: D-Matrix, an AI chip company positioning itself as a challenger to Nvidia. Infinity is also in talks with other chip makers and cloud providers, though Nixon did not name them.
Infinity is one of several startups now trying to reduce the AI industry’s dependence on Nvidia’s software stack. The broader trend matters because Nvidia’s CUDA advantage has effectively slowed adoption of competing chips, even when those chips have strong hardware specs. If companies like Infinity can lower the software barrier, it becomes much easier for chip makers to compete on the merits of their silicon rather than on ecosystem lock-in.
The company currently has 26 employees across engineering, design, and operations. It launched last year, which makes the $100 million valuation and the caliber of its backers a notable early signal of investor appetite for this kind of infrastructure bet.
