Milan Lustig's Opt32 raised $5M to build full-stack compute infrastructure for autonomous robots and drones

Opt32

Milan Lustig dropped out of Harvard as a freshman to build Opt32, a full-stack compute infrastructure company targeting physical autonomy. The company raised a $5M seed round co-led by BoxGroup roughly two months ago and currently operates as a three-person team, all technical co-founders who have been friends since high school.

The pitch is that robots, drones, and autonomous vehicles have a fundamental compute problem that neither cloud-scale GPUs nor off-the-shelf accelerators solve well. A battery-powered robot can't carry a server-grade GPU, and general-purpose chips aren't optimized for the low-latency, constrained-compute environment that on-device machine learning demands. Opt32 builds model optimization software that gets robotics companies' perception models running faster, or fits more capable models onto cheaper hardware, across existing back ends including NVIDIA GPUs and Qualcomm accelerators.

“At Opt32, what we're trying to do is essentially build modern full stack compute infrastructure for physical autonomy — from software, so compilers, down to chips, custom accelerators to run on device machine learning and things like robots, drones, cars, autonomous defense systems. We raised our seed round about two months ago. We raised $5,000,000 co-led by BoxGroup and by Venture.”

Stack and roadmap

Lustig describes a deliberate hardware progression. The company starts at the software layer, which is already generating revenue through design partners, with near-zero operating costs. From there the plan moves to single-board computers built around existing accelerator chips, then FPGA implementations, then a smaller prototype tape-out, with a full advanced-process-node ASIC run targeted one to two years out pending further capital raises. The software-first approach lets Opt32 generate cash before the capital-intensive chip work begins.

The entry market is robotics, though Lustig is specifically bullish on manufacturing automation, where he sees a skills gap in trades like welding that robots could help close. He expects meaningful consumer-facing autonomy, including cooking robots and civic applications like street cleaning, within roughly three years.

On the question of whether edge compute can simply be replaced by a rack of GPUs in a server closet with lightweight robots feeding in camera data over WiFi, Lustig's answer is that it depends. Some use cases can tolerate the latency; others can't. Opt32's software already handles split workloads, running latency-sensitive tasks on device and heavier inference in the cloud.

The team is three people and actively hiring. The longer capital need is real, but the software layer buys time before the ASIC spend becomes unavoidable.