Infinity raises $15 million to make any AI chip inference-ready

An early stage research company called Infinity has raised $15 million in seed funding to attack a problem that has quietly protected Nvidia for two decades, which is that a chip is only as good as the software written to run models on it. The round, announced on 20 July, values the San Francisco firm at $100 million post money and drew significant participation from Touring Capital, alongside Principal Venture Partners, chip industry executives and researchers from OpenAI and Anthropic. Infinity was founded in August 2025 by Jeremy Nixon, a former Google Brain researcher. The product at the center of it is Ignition, an autonomous agent that writes, tests and tunes the low level compute kernels that decide how much of a processor's ardvertised performance a model can actually reach. Engineers have traditionally hand tuned those kernels for every combination of architecture and model, a grind that runs for months or years. It is the main reason capable silicon from AMD, Qualcomm and AWS has struggled to convert raw specifications into deployed inference. Nixon argues the industry has lived under an artificial constraint because only one vendor spent twenty years building that software layer. The numbers on offer are eye catching. In a design partnership with d-Matrix, Infinity says its agents reached up to 92 percent of the theoretical peak of the Corsair accelerator within ten hours of first accessing the hardware, and had Qwen3, Qwen3.5 and Gemma4 running end to end inside ten days. On a Qwen3-8B model, throughput reportedly climbed from about 1,400 tokens per second to more than 20,000 in a day, beating the vLLM framework by roughly 34 percent. Nixon has acknowledged that test used a single H100 and has not been independently validated. The commercial model is unusual. Instead of licensing fees, Infinity takes a share of the compute savings it produces, roughly a fifth against an agreed baseline, and already generates millions of dollars in annual recurring revenue. The new capital funds hiring, the research platform and talks with further chipmakers, timed against forecasts that inference will absorb about two thirds of all AI compute spending this year. For the Gulf this is an operational headache rather than an abstraction. Saudi Arabia's HUMAIN and the UAE's G42 were cleared by Washington in late 2025 to buy up to 35,000 Blackwell class chips each, yet both have spread their bets. HUMAIN has committed $1.5 billion with Groq for inference infrastructure, including a Dammam site running 19,000 of its processing units, signed with Qualcomm for up to 200 megawatts of systems, and added AMD to dilute single vendor risk. G42 is building Stargate UAE inside a five gigawatt campus pitched partly on low latency inference for the region. Those non Nvidia deployments all need mature software to justify themselves, and regional buyers have little appetite for waiting. Whether Infinity can generalize across that many architectures at scale remains unproven. But if the claims hold, the moat around CUDA looks less like a wall and more like a head start.