Infinity raises $15M to break CUDA's grip on AI chips

Infinity raised $15 million to build a software layer that lets AI models deploy on new silicon without per-chip porting work, per techstartups.com's July 20 funding roundup. The pitch is narrow: the reason nobody has cracked NVIDIA's position isn't fabrication capacity, it's the compiler and runtime surface every model has to pass through to run on a given chip. Build that layer once and abstract it across vendors, and porting a model to a new chip stops being a multi-quarter engineering slog.

That matters because of where AI margins sit. This category runs around 52% gross margins, and a real slice of that spread is the CUDA lock-in tax, the cost every team pays to rewrite kernels and tooling just to move off one vendor. A working multi-silicon abstraction doesn't just crack the door for AMD or custom silicon, it changes what inference costs on paper.

The gap in the story

Techstartups.com's roundup doesn't say which silicon vendors Infinity is targeting first, or whether any exist as design partners yet. $15 million is seed-stage money for a problem NVIDIA has spent a decade defending with CUDA's software moat, not a signal the abstraction already runs in production.

Watch whether Infinity ships a working port to a non-NVIDIA chip in the next round of coverage, or whether this stays a funding announcement with no deployed model behind it. That's the difference between a real crack in the moat and a pitch deck.