Three years ago, Luminal co-founder Joe Fioti was working on chip design at Intel when he came to a realization. While he was working on making the best chips he could, the more important bottleneck was in software.
“You can make the best hardware on earth, but if it’s hard for developers to use, they’re just not going to use it,” he told me.
Now, he’s started a company that focuses entirely on that problem. On Monday, Luminal announced $5.3 million in seed funding, in a round led by Felicis Ventures with angel investment from Paul Graham, Guillermo Rauch, and Ben Porterfield.
Fioti’s co-founders, Jake Stevens and Matthew Gunton, come from Apple and Amazon, respectively, and the company was part of Y Combinator’s Summer 2025 batch.
Luminal’s core business is simple: The company sells compute, just like neo-cloud companies like CoreWeave or Lambda Labs. But where those companies focus on GPUs, Luminal has focused on optimization techniques that let the company squeeze more compute out of the infrastructure it has. In particular, the company focuses on optimizing the compiler that sits between written code and the GPU hardware — the same developer systems that caused Fioti so many headaches in his previous job.
At the moment, the industry’s leading compiler is Nvidia’s CUDA system — an underrated element in the company’s runaway success. But many elements of CUDA are open source, and Luminal is betting that, with many in the industry still scrambling for GPUs, there will be a lot of value to be gained in building out the rest of the stack.
It’s part of a growing cohort of inference-optimization startups, which have grown more valuable as companies look for faster and cheaper ways to run their models. Inference providers like BaseTen and Together AI have long specialized in optimization, and smaller companies like Tensormesh and Clarifai are now popping up to focus on more specific technical tricks.
Luminal and other members of the cohort will face stiff competition from optimization teams at major labs, which have the benefit of optimizing for a single family of models. Working for clients, Luminal has to adapt to whatever model comes their way. But even with the risk of being out-gunned by the hyperscalers, Fioti says the market is growing fast enough that he’s not worried.
“It is always going to be possible to spend six months hand tuning a model architecture on a given hardware, and you’re probably going to beat any sorts of, any sort of compiler performance,” Fioti says. “But our big bet is that anything short of that, the all-purpose use case is still very economically valuable.”
