Luminal builds a compiler that makes AI models run more efficiently on existing hardware. We led its $25 million Series A because better software can unlock more useful computation from chips companies have already bought. Meeting demand for AI is not just a matter of building more data centers. It also means getting more out of the ones we have.
Founders Joe Fioti, Matthew Gunton, and Jake Stevens are approaching that problem through compilers: software that translates a model into the instructions a chip executes. In the interview, Joe describes working on AI hardware at Intel and seeing capabilities go unused because the software did not take advantage of them. Luminal started with the idea that the hardware was already capable of more.
A model’s mathematics does not dictate a single way to run it. There are different ways to move data, combine operations, and schedule work. Those choices determine how much time a chip spends computing rather than waiting. Luminal searches for efficient implementations instead of requiring engineers to optimize every model and chip combination by hand.
The founders describe this as applying more computation to compilation itself. Their approach uses equality saturation, a technique for representing many equivalent versions of a computation and searching among them. The goal is to do more work before a model runs so that serving it repeatedly costs less.
Customers can use Luminal’s hosted inference service or license its software to run on their own infrastructure. That gives the team two ways to deliver the same result: help customers serve models without managing the hardware themselves, or improve the performance of hardware they already operate.
The opportunity also extends beyond Nvidia. A chip is only useful if developers can run their models on it. Luminal is building software for GPUs and specialized AI chips, with the ambition of making it easier to choose hardware based on the workload rather than the software ecosystem it requires. As data centers combine different kinds of processors, compiling and coordinating work across them becomes a larger problem.
The founders chose Standard Capital because they wanted to keep building rather than step away for a long fundraising process. In the interview, they describe a straightforward process and a community of founders who help one another. That is the relationship we want: less time spent raising money, more time solving the problem.
Luminal’s ambition is to make inference faster and cheaper by improving the software between a model and the hardware running it. The chips are already there. Better software can make them do more.
Learn more at luminal.com.