Three years in the past, Luminal co-founder Joe Fioti was engaged on chip design at Intel when he got here to a realization. Whereas he was engaged on making the very best chips he might, the extra essential bottleneck was in software program.
“You may make the very best {hardware} on earth, but when it’s laborious for builders to make use of, they’re simply not going to make use of it,” he informed me.
Now, he’s began an organization that focuses solely on that drawback. On Monday, Luminal introduced $5.3 million in seed funding, in a spherical led by Felicis Ventures with angel funding 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 corporate was a part of Y Combinator’s Summer season 2025 batch.
Luminal’s core enterprise is easy: the corporate sells compute, similar to neo-cloud corporations like Coreweave or Lambda Labs. However the place these corporations deal with GPUs, Luminal has targeted on optimization strategies that allow the corporate squeeze extra compute out of the infrastructure it has. Particularly, the corporate focuses on optimizing the compiler that sits between written code and the GPU {hardware} — the identical developer methods that triggered Fioti so many complications in his earlier job.
In the intervening time, the trade’s main compiler is Nvidia’s CUDA system — an underrated factor within the firm’s runaway success. However many parts of CUDA are open-source, and Luminal is betting that, with many within the trade nonetheless scrambling for GPUs, there will probably be loads of worth to be gained in constructing out the remainder of the stack.
It’s a part of a rising cohort of inference-optimization startups, which have grown extra helpful as corporations search for quicker and cheaper methods to run their fashions. Inference suppliers like Baseten and Collectively AI have lengthy specialised in optimization, and smaller corporations like Tensormesh and Clarifai at the moment are popping as much as deal with extra particular technical methods.
Luminal and different members of the cohort will face stiff competitors from optimization groups at main labs, which take pleasure in optimizing for a single household of fashions. Working for purchasers, Luminal has to adapt to no matter mannequin comes their approach. However even with the danger of being out-gunned by the hyperscalers, Fioti says the market is rising quick sufficient that he’s not apprehensive.
“It’s at all times going to be attainable to spend six months hand tuning a mannequin structure on a given {hardware}, and also you’re in all probability going to beat any types of, any form of compiler efficiency,” Fioti says. “However our huge wager is that something in need of that, the all-purpose use case continues to be very economically helpful.”
