Infinity's $15M: your token price is a software problem

When your AI bill goes up, the instinct is to blame the hardware — Nvidia's margin, the GPU shortage, the price of memory. That's half the story. A meaningful chunk of what you pay per token is decided by the software layer between the model and the silicon, and almost nobody outside a handful of labs can touch it. A seed-stage company just raised $15 million to hand that layer to an agent.

What actually happened

On July 20, Infinity announced a $15 million seed round at a $100 million post-money valuation, led in part by Touring Capital with Principal Venture Partners and angels from OpenAI, Anthropic, and chipmakers. The founder is Jeremy Nixon, previously a Google Brain researcher and co-founder of AGI House.

The product is called Ignition, and the pitch is a CUDA alternative that isn't hand-written. Per SiliconANGLE, Ignition is an AI agent that generates the low-level stack for running models on a given processor — kernels, compilers, profilers, debuggers, and the orchestration between them — using decompilers to work against proprietary instruction sets. Humans specify which tools to build; the agent implements them. Infinity claims Ignition raised Qwen3-8B throughput 34% over vLLM on a single Nvidia H100 after one day of work, and hit roughly 92% of theoretical peak on d-Matrix's Corsair accelerator within 10 hours. SiliconANGLE notes those numbers are the company's own and haven't been independently validated. Treat them as directional, not gospel. The business model is the tell: Infinity typically takes about 20% of the compute savings it produces, and already books millions in ARR from chip-design partnerships.

Why it matters for your business

You are not going to write inference kernels. That's not the takeaway. The takeaway is that "our AI costs too much" has more than one fix, and the expensive one — buy more compute, or accept the vendor's price — is the one you'll reach for by default.

Two practical moves. First, before you scale up hardware or jump to a bigger instance, check whether your serving stack is leaving throughput on the floor. If you're self-hosting an open-weight model, the gap between a naive deployment and a tuned one on the same GPU is real money, and it's the cheapest optimization available. Second, watch what this round implies about the next 18 months: if the software that makes a chip usable can be generated instead of hand-built by a scarce team of specialists, the moat around any single vendor's ecosystem gets thinner, and more silicon becomes viable for inference. That's good for your bill — but only if you built so you can move. A stack hard-wired to one provider's SDK can't collect on a price war it isn't allowed to enter. Keep your model layer swappable and your serving layer boring, and cheaper compute becomes something you can actually take advantage of.

Sources: TechCrunch, SiliconANGLE.