Infinity builds Autoresearch, software that makes a given AI accelerator ready for faster inference. It takes a hardware target and a serving workflow, then runs automated experiments over kernels, memory layout, and serving settings until throughput improves. Chip partners such as d-Matrix have used the approach to bring models up on new silicon. This launch applies it with HP to hardware already in the field: the ZGX Fury, a rack-mountable, office-deployable AI station on the NVIDIA DGX Station platform and the GB300 Grace Blackwell Ultra Desktop Superchip, built so several people can share one box for local inference. Infinity and HP report that Autoresearch doubled concurrent workloads and doubled token throughput for NVIDIA Nemotron 3 Super, a 120.6B model, on the tested GB300 configuration, with no hardware change. The test used HP's existing vLLM benchmark, compared against an 82-concurrent-user baseline, and held a defined responsiveness threshold, with quality checks on NVIDIA SPEED-Bench, MMLU, and WikiText. The optimizations span the inference stack, drawing on techniques used in vLLM, TensorRT-LLM, and SGLang plus Infinity's own work. The company says the system was given HP's baseline, ran 216 experiments, and reached the 2x result without speculative decoding, with the first gain in under 24 hours. The same launch thread reports 2.89x on Qwen 3.8 Max across eight B300s, 2.07x on Deepseek v4 Flash on one B300, and 2.19x on GLM 5.3 Flash across four B300s. Jeremy Nixon, founder and CEO, is a former Google Brain researcher and a Harvard graduate. He started Infinity in San Francisco in August 2025. In July 2026 the company raised $15 million in seed funding at a $100 million post-money valuation, with participation from Touring Capital, Principal VC, and researchers from OpenAI and Anthropic. Earlier published results include 92% of theoretical peak on d-Matrix Corsair within 10 hours of first hardware access, a later 20x tokens-per-second gain on Qwen 3 versus Infinity's own first Corsair implementation, and an H100 run in which its infy system delivered up to 34.3% more tokens per second than vLLM on Qwen3-8B under matched settings. The HP figure is that same measurement loop on a GB300 workstation, where twice the concurrency means more simultaneous users on one box.