Infinity Raises $15M for a ‘CUDA for Everything’ AI Layer
Jeremy Nixon's startup wants AI agents to write the low-level code that lets any chip run any model.
In Brief
- Infinity raised $15M at a $100M valuation to build chip-agnostic inference software.
- Its Ignition agent writes and self-optimizes the kernels Nvidia’s CUDA normally handles.
- Customers already include AI-chip maker D-Matrix, a direct Nvidia challenger.
AI infrastructure startup Infinity announced a $15 million raise at a $100 million valuation on Monday, with investors including Touring Capital, Principal VC, and researchers from OpenAI and Anthropic. The bet is unglamorous but enormous: software that lets AI models run well on any chip, not just Nvidia’s. TechCrunch reports the round values a company trying to chip away at the most entrenched moat in silicon.
Nvidia’s dominance is not only about its physical GPUs. It is also CUDA, the software layer that lets those chips act as general-purpose AI accelerators and that frameworks like PyTorch and TensorFlow are built on top of. Most application startups lack the expertise—or the incentive—to write their own low-level kernels and port to alternative hardware. Infinity is attempting to build a universal inference library that runs across SRAM, GPUs, phone chips, and systolic arrays.
The company was launched last year by Jeremy Nixon, a former Google Brain researcher and the creator of the AGI House community. Nixon told TechCrunch he was driven by “automated invention”—the belief that “AI systems can actually be a meta technology” capable of generating and evaluating their own improvements. He had previously built a machine-learning algorithm, Omega, that essentially invented and tested new algorithms in a feedback loop.
An agent that writes the kernels
Infinity’s research agent, Ignition, is designed to write the low-level code needed for AI inference on Nvidia-alternative chips. It tests, debugs, and measures hardware performance, then rewrites the code to improve it—a self-optimizing loop that Nixon says adapts to different proprietary architectures. In one case study, the agent compressed what could have been a months-long engineering effort into hours or days.
Humans stay in the loop for high-level direction while the agent handles the tedious grunt work, and Infinity does not charge an upfront license fee. Instead, it takes a cut of the performance gains and cost savings it measures in tokens per second. That aligns the startup’s revenue with the efficiency it actually delivers. The approach parallels the broader industry shift highlighted in recent AI-compute partnership deals.
Customers already include D-Matrix, an AI-chip maker and would-be Nvidia challenger, and Nixon says Infinity is in talks with other large chip and cloud companies. With 26 employees spanning engineering, design, and operations, the startup is part of a wider wave betting that Nvidia’s software lock-in is finally contestable.
Why Infinity’s software stack matters more than the chip
The reason CUDA is so defensive is that developers default to whatever their frameworks support. Infinity is essentially trying to recreate a CUDA-level stack that is vendor-neutral, so an app written once can target many chips. If it works, it lowers the switching cost that keeps buyers tethered to Nvidia.
The valuation context is notable: $100M for a 26-person company building developer tooling is rich, but consistent with the premium investors place on AI-infrastructure leverage. The same appetite showed up in Databricks’ $188B valuation and similar bets on the pick-and-shovel layer of the AI boom.
Infinity’s real test is whether Ignition’s generated kernels match hand-tuned CUDA on speed and stability at scale. Until then, “CUDA for everything” is a compelling pitch, not yet a proven replacement.
FAQ
How much did Infinity raise and at what valuation?
Infinity raised $15 million at a $100 million valuation, with backing from Touring Capital, Principal VC, and researchers from OpenAI and Anthropic, per TechCrunch.
What does the Ignition agent actually do?
Ignition writes the low-level kernel code needed for AI inference on non-Nvidia chips, then tests, debugs, and rewrites it to improve performance in a self-optimizing loop.
Who are Infinity’s early customers?
The company says customers include AI-chip maker D-Matrix, an Nvidia challenger, and that it is in talks with other major chip and cloud firms.
