Luminal raises $25M to build the compiler stack that unlocks heterogeneous AI compute across chip vendors

Luminal raises $25M to commoditize AI chip software

Luminal is betting that the real bottleneck in the AI chip race isn't hardware — it's software. Joe Fioti, the company's co-founder and CEO, describes Luminal as a software-defined accelerator company: its compiler stack is designed to unlock heterogeneous compute across chip vendors, so that the growing field of AI accelerator startups isn't bottlenecked by the one company capable of building the accompanying software layer. That company, Fioti notes, is worth $5 trillion now.

Luminal has raised $25M to scale that stack.

“Luminal is a software-defined accelerator company. There's actually a lot of accelerator companies that can make the hardware component and only one that can make the software component — and they're worth $5 trillion now. We raised $25 million. On our cloud, we're doing something like 100 trillion tokens a month and we're ramping a lot of that on Trainium, getting pretty phenomenal price performance even against Blackwell Ultra.”

Throughput over speed

Fioti's core commercial argument is that inference customers almost universally prioritize cost over raw speed. Beyond a performance threshold, faster doesn't win — cheaper does. That shapes Luminal's chip preferences. The company runs roughly 100 trillion tokens per month in inference throughput on its own cloud, and a significant portion of that runs on AWS Trainium. Fioti says Luminal's Trainium performance holds up favorably against Blackwell Ultra on price-performance.

Go-to-market

Luminal operates on two commercial tracks: licensing its compiler stack directly to neo-clouds, foundation model labs, and inference providers; and running its own hosted inference cloud for smaller customers who want Luminal to manage the hardware. Larger customers that own their own silicon license directly.

Chip partnerships

One of Luminal's first compiler backend partnerships was with Positron AI, whose Atlas chips are now ramping in Luminal's cloud. Fioti says the compiler-to-chip timeline is actually an advantage — bringing up a compiler backend is faster than taping out a chip, so Luminal enters the relationship late in a chip vendor's design cycle, with a completed design in hand. Positron support came together within a few months. Fioti says Luminal has conversations underway with other emerging chip vendors, including Etched and Cerebras, though no additional partnerships beyond Positron are confirmed in the transcript.

Luminal's underlying thesis is that the starting gun has already fired on a race to commoditize compute, and that software is what will determine which hardware actually wins commercial workloads.