Infinity.inc Raises $15 Million to Build AI Inference Software for Any Chip
Infinity.inc, an early-stage AI infrastructure research company, has closed a $15 million seed round at a $100 million post-money valuation. Touring Capital led the participation, joined by Principal Venture Partners, executives from major chip companies, researchers from OpenAI and Anthropic, and a group of angel investors.
The company’s core product is Ignition, an autonomous AI agent that writes, tests, and optimizes the low-level compute kernels that determine how efficiently a chip runs AI models. The pitch is straightforward: NVIDIA’s roughly 80% share of the data-center AI accelerator market rests less on raw silicon performance than on two decades of CUDA software development. Rival chipmakers with competitive hardware, including AMD, Qualcomm, and AWS, have struggled to close that gap because building a comparable inference stack by hand takes years. Infinity is betting that an AI agent can compress that timeline into days.
Founder and CEO Jeremy Nixon, a former Google Brain researcher and co-founder of the AGI House network, frames the opportunity as one about access rather than raw chip design. He argues the next phase of AI competition will be decided by whichever company can make any chip run state-of-the-art models efficiently, not just whichever company builds the fastest silicon.
Early results with d-Matrix
Infinity’s first live design partnership is with d-Matrix, whose Corsair chip now runs production inference through the Infinity d-Matrix Cloud. According to the company, Infinity’s agents reached 92% of Corsair’s theoretical peak performance within 10 hours of first accessing the hardware, running tensor-parallel matrix multiplications across all 32 of its compute units. Within 10 days, three frontier models — Qwen3, Qwen3.5, and Gemma4 — were running end-to-end on the chip, with every layer written from scratch.
On a separate benchmark, Infinity says it pushed inference throughput on a Qwen3-8B model from roughly 1,400 tokens per second to more than 20,000, a roughly 14x improvement achieved in a single day of automated kernel generation, and one that beat the widely used vLLM framework by more than 34%.
The system relies on what Infinity calls recursive self-improvement: real-world performance data from deployed kernels feeds back into training the next generation of generated code, with human engineers steering architecture rather than hand-tuning kernels directly.
Business model and next steps
Rather than charging chip partners upfront licensing fees, Infinity says it captures a share of the performance gains and cost savings its software generates — a structure meant to align its incentives with the hardware companies it works with. The company is already generating multiple millions of dollars in annual recurring revenue from its chip design partnerships and says it is in active discussions with other major chip makers beyond d-Matrix.
The new funding will go toward scaling Ignition, growing Infinity’s engineering team, and deepening its existing chip partnerships. The company is currently hiring research engineers for hardware enablement work and engineers to maintain its Infy inference library.
With inference workloads projected to account for roughly two-thirds of AI compute spending in 2026, the question Infinity is racing to answer is whether an AI agent can genuinely substitute for the years of manual kernel engineering that have kept NVIDIA’s software moat intact — and whether chip makers outside the CUDA ecosystem are willing to bet their production stacks on it.