Luminal raises $5.3 million to build a better GPU code framework

Three years ago, Luminal co-founder Joe Fioti was moving connected spot plan astatine Intel erstwhile helium came to a realization. While helium was moving connected making the champion chips helium could, the much important bottleneck was successful software.

“You tin marque the champion hardware connected earth, but if it’s hard for developers to use, they’re conscionable not going to usage it,” helium told me.

Now, he’s started a institution that focuses wholly connected that problem. On Monday, Luminal announced $5.3 cardinal successful effect funding, successful a circular led by Felicis Ventures with angel concern from Paul Graham, Guillermo Rauch, and Ben Porterfield. 

Fioti’s co-founders, Jake Stevens and Matthew Gunton, travel from Apple and Amazon, respectively, and the institution was portion of Y Combinator’s Summer 2025 batch.

Luminal’s halfway concern is simple: the institution sells compute, conscionable similar neo-cloud companies similar Coreweave oregon Lambda Labs. But wherever those companies absorption connected GPUs, Luminal has focused connected optimization techniques that fto the institution compression much compute retired of the infrastructure it has. In particular, the institution focuses connected optimizing the compiler that sits betwixt written codification and the GPU hardware — the aforesaid developer systems that caused Fioti truthful galore headaches successful his erstwhile job.

At the moment, the industry’s starring compiler is Nvidia’s CUDA strategy — an underrated constituent successful the company’s runaway success. But galore elements of CUDA are open-source, and Luminal is betting that, with galore successful the manufacture inactive scrambling for GPUs, determination volition beryllium a batch of worth to beryllium gained successful gathering retired the remainder of the stack.

It’s portion of a increasing cohort of inference-optimization startups, which person grown much invaluable arsenic companies look for faster and cheaper ways to tally their models. Inference providers similar Baseten and Together AI person agelong specialized successful optimization, and smaller companies similar Tensormesh and Clarifai are present popping up to absorption connected much circumstantial method tricks.

Luminal and different members of the cohort volition look stiff contention from optimization teams astatine large labs, which person the payment of optimizing for a azygous household of models. Working for clients, Luminal has to accommodate to immoderate exemplary comes their way. But adjacent with the hazard of being out-gunned by the hyperscalers, Fioti says the marketplace is increasing accelerated capable that he’s not worried.

“It is ever going to beryllium imaginable to walk six months manus tuning a exemplary architecture connected a fixed hardware, and you’re astir apt going to bushed immoderate sorts of, immoderate benignant of compiler performance,” Fioti says. “But our large stake is that thing abbreviated of that, the all-purpose usage lawsuit is inactive precise economically valuable.”