Infinity.inc, founded by Google Brain researcher Jeremy Nixon, closed a $15 million seed round at a $100 million valuation on July 20, 2026, backed by Touring Capital and Principal VC, with participation from OpenAI and Anthropic researchers and chip-industry executives. The San Francisco startup is building Ignition, an autonomous AI agent that writes, tests, and optimizes low-level compute kernels for new AI silicon—work that historically takes engineering teams months or years.
Ignition compresses the inference software development cycle to days through recursive self-improvement: real-world performance data from deployed kernels feeds back into training the next generation of code, while human engineers steer architecture rather than hand-tuning. In public benchmarks, Ignition improved inference throughput on Qwen3-8B from ~1,400 tokens/sec to >20,000 tokens/sec (14x improvement) in one day, and reached 92% of d-Matrix's Corsair chip's theoretical peak performance within 10 hours of first accessing hardware.
For AI architects evaluating infrastructure economics, Infinity addresses a fundamental NVIDIA moat: decades of CUDA development. Inference is expected to represent two-thirds of AI compute spending in 2026, yet NVIDIA's advantage persists because alternative chips lack optimized software stacks. Infinity is already generating millions in annual recurring revenue from chip design partnerships and is in talks with other major semiconductor companies—a signal that non-NVIDIA alternatives may finally have a software path to market viability.