Three years back, Luminal’s co-founder Joe Fioti was engaged in chip development at Intel when he had an important insight. Despite his efforts to create top-tier chips, he realized the real challenge lay within the software.
“You can build the most advanced hardware available, but if developers find it difficult to work with, they simply won’t adopt it,” he explained to me.
Today, he has launched a startup dedicated to addressing this very issue. On Monday, Luminal revealed it had secured $5.3 million in seed funding, with Felicis Ventures leading the round and additional backing from angel investors Paul Graham, Guillermo Rauch, and Ben Porterfield.
Fioti’s fellow founders, Jake Stevens and Matthew Gunton, previously worked at Apple and Amazon, and the company participated in Y Combinator’s Summer 2025 program.
Luminal’s main offering is straightforward: it provides computing resources, similar to new cloud providers such as Coreweave or Lambda Labs. However, unlike those firms that concentrate on GPUs, Luminal emphasizes optimization strategies to maximize the computing power of its existing infrastructure. Specifically, the company is improving the compiler that bridges the gap between developers’ code and GPU hardware—the same systems that once frustrated Fioti in his earlier career.
Currently, Nvidia’s CUDA platform is the dominant compiler in the field—an often overlooked factor in Nvidia’s massive success. Yet, with many aspects of CUDA being open-source, Luminal is wagering that as the demand for GPUs continues, there’s significant opportunity in enhancing the rest of the technology stack.
This approach places Luminal among a rising group of startups focused on inference optimization, which have become increasingly valuable as organizations seek more efficient and cost-effective ways to operate their models. Companies like Baseten and Together AI have long specialized in this area, and newer entrants such as Tensormesh and Clarifai are emerging to tackle more specialized technical challenges.
Luminal and its peers will have to contend with strong competition from optimization teams at major research labs, who benefit from focusing on a single model family. In contrast, Luminal must tailor its solutions to a variety of client models. Despite the possibility of being outpaced by larger players, Fioti believes the rapidly expanding market leaves plenty of room for growth.
“You could always spend half a year meticulously tuning a model’s architecture for specific hardware, and you’d likely outperform any compiler,” Fioti remarks. “But our main belief is that for anything less than that, a general-purpose solution still holds substantial economic value.”
