Luminal Searches Trillions of Options to Find a Fast Way to Run AI
As chips grow more specialized and harder to program, the startup bets that trying many options automatically beats writing rules by hand.

Standard Capital · Luminal Raises $25M Series A to Break AI's Hardware Bottleneck, Led by Standard CapitalFor years, chips got faster almost on their own as transistors shrank and used less power. That steady trend, known as Moore's law, has largely stalled.
To keep improving, companies built more specialized processors, from graphics chips repurposed for AI to custom circuits built only for inference, the work of answering a query. That hardware can do more math, but it leaves far more work for software to schedule operations and fetch data from memory at the right moment.
A compiler does that translation. It takes a model written in a popular research tool called PyTorch, which makes the math easy to write, and generates the detailed instructions a particular chip needs.
Traditionally, compiler engineers hand-write rules of thumb, called heuristics, to choose among possible implementations. Joe, a co-founder of the compiler startup Luminal, says his team took the opposite lesson from AI research.
In AI, the so-called bitter lesson holds that systems improve most when designers add more computing power and rely less on human expertise, an idea seen in chess, Go, protein folding, and language models. Luminal applies the same idea to compilers, spending extra computing power while a program is being prepared to search through options, rather than asking people to anticipate every shortcut.
The trick is a technique called equality saturation, which Luminal did not invent. It packs trillions of possible ways to run the same model on the same chip into one compact structure, and the compiler then walks through that space to find a fast implementation, instead of requiring engineers to write each trick by hand.
Luminal says it is among the first companies to use the approach at industrial scale. The claim has historical echoes: better database software in the 1990s, including work at Oracle, and better use of existing networking hardware helped the internet scale without entirely new machines.
If that history repeats for AI, the payoff is practical. As chips become more specialized and harder to program, how fully they are used may depend less on the silicon itself than on how thoroughly software searches for a fast way to run it.
Questions people ask
What does equality saturation actually do?
In plain terms, it is a way to store many equivalent versions of a calculation at once. Instead of picking one rewrite and discarding the rest, the compiler keeps all the alternatives together and keeps adding new equivalents, then chooses the cheapest complete version at the end.
Does searching trillions of options slow down the AI for users?
No, the search happens once, before the model serves users. That upfront work can be heavy, but it pays off because the resulting program may then answer more queries faster while using the same chip and power.
Is this only useful for non-Nvidia chips?
It can help both. Nvidia chips also run faster with better implementations, but the gap can matter more for alternative chips whose existing software is less mature. Better compilation can therefore make a non-Nvidia chip practical for work that once required Nvidia.

