Infinity raises $15M to build a CUDA alternative for any chip
TL;DR
- Infinity closed a $15M seed at a $100M valuation on Monday, backed by Touring Capital, Principal VC, and researchers from OpenAI and Anthropic.
- Its Ignition agent writes low-level inference code for non-Nvidia chips including SRAM, phone chips, and systolic arrays, then rewrites itself to improve speed.
- Instead of a license fee, Infinity takes a cut of performance gains measured in tokens per second; AI chip maker D-Matrix is a customer.
The interesting thing about Infinity's $15 million seed, reported by TechCrunch on Monday, is not the dollar figure. It is the pricing model. Ex Google Brain researcher Jeremy Nixon's inference startup does not charge an upfront license fee for its software. It takes a cut of the performance gains and cost savings it delivers, measured in tokens per second. The round closed at a $100 million valuation, with Touring Capital and Principal VC alongside individual researchers from OpenAI and Anthropic.
The product is called Ignition, and the pitch is ambitious. Nvidia's software moat around AI is CUDA, the low-level kernel software that makes its GPUs easier to program than the alternatives. Infinity is trying to build the CUDA-equivalent for everything else, targeting SRAM, phone chips, systolic arrays, and non-Nvidia GPUs. What makes the approach worth watching is that Ignition is itself an AI research agent. According to TechCrunch, it writes the low-level code, tests and debugs it on the hardware, measures how fast the chip actually runs, and rewrites itself when it finds a slower path.
The commercial hook is D-Matrix, one of the more credible would-be Nvidia challengers on the chip side, listed as a customer, with Infinity saying it is in discussions with other major chip and cloud companies. Nixon, previously behind an ML algorithm called Omega and the AGI House hacker community, told the outlet he was drawn to the space out of an obsession with the idea of automated invention. The company has 26 people across design, operations, and engineering.
The honest caveat is that the reporting does not give you the numbers that actually matter here. There are no tokens-per-second benchmarks against hand-tuned CUDA, no detail on how a cut of measured performance gains gets baselined or audited when a vendor's paycheck depends on the delta, and no view on whether Ignition covers only inference kernels or reaches into the higher-level CUDA libraries that also lock developers in.
If it does hold up, every Nvidia alternative suddenly has a shared software layer it did not have to build alone, and the cloud providers eyeing inference margin relief get something they can actually plug in.
Originally reported by techcrunch.com
Read the original article →Original headline: Infinity raises $15M seed to build a CUDA alternative that runs AI models on any chip