Three years ago, Joe Fioti, now co-founder and CEO of Luminal, was deep in chip design at Intel. Despite working on cutting-edge hardware, he realized the real bottleneck wasn’t in silicon—it 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 recalled. That insight led him to found Luminal, a startup focused on solving one of AI’s most persistent challenges: making GPU-powered computing more accessible and efficient. On Monday, Luminal announced a $5.3 million seed round, led by Felicis Ventures, with angel investments from notable figures including Paul Graham, Guillermo Rauch, and Ben Porterfield. Fioti’s co-founders, Jake Stevens and Matthew Gunton, bring deep experience from Apple and Amazon, respectively. The company also joined Y Combinator’s Summer 2025 batch, marking a strong start in the competitive AI infrastructure space. Luminal operates like a cloud compute provider—offering access to powerful GPU resources similar to companies like Coreweave or Lambda Labs. But its real differentiator lies in optimization. Instead of just selling raw compute, Luminal focuses on improving how code runs on that hardware, particularly by enhancing the compiler layer that translates developer-written code into instructions for GPUs. This is where the company’s core mission lies. While Nvidia’s CUDA platform dominates the market, many of its components are open-source, creating an opening for innovation. Luminal is building a more efficient, flexible, and developer-friendly alternative to the current stack, aiming to unlock better performance across a wide range of models and use cases. The move comes amid a surge in inference-optimization startups. Companies like Baseten and Together AI have long specialized in speeding up model deployment, while newer players such as Tensormesh and Clarifai are emerging with niche technical solutions. Luminal joins this growing ecosystem, betting that there’s still significant value in creating general-purpose optimization tools that work across diverse models and frameworks. Still, the company faces challenges. Large AI labs like Google, Meta, and OpenAI have in-house teams that can fine-tune models and hardware for specific architectures, giving them a performance edge. Luminal, by contrast, must support a wide variety of models and client needs, making it harder to achieve the same level of hand-optimized results. But Fioti remains confident. “It is always going to be possible to spend six months hand-tuning a model on a given hardware, and you’re probably going to beat any sort of compiler performance,” he said. “But our big bet is that anything short of that—most real-world use cases—the all-purpose, general optimization approach is still very economically valuable.” With demand for faster, cheaper inference growing, Luminal believes it’s positioned to meet the needs of developers who can’t afford to wait months for custom tuning.