Neurophos Raises $110M in Series A for Photonic AI Chip Technology

Austin has a way of birthing companies that do not ask permission from physics. Neurophos is one of those companies. Founded in 2020 by Patrick Bowen and Andrew Traverso, spun out of Duke University and Metacept, this team took decades of metamaterials research once aimed at bending electromagnetic waves and pointed it straight at the most expensive problem in modern computing: power. Too much heat, too much draw, too many limits pretending to be roadmaps.

On January 22, 2026, Neurophos locked in a $110 million Series A, oversubscribed, led by Gates Frontier with participation from M12, Aramco Ventures, Bosch Ventures, MetaVC Partners, Carbon Direct Capital, DNX Ventures, Alumni Ventures, Gaingels, and others who understand that electrons are starting to feel like legacy infrastructure. Total funding now stands at $118 million. The message from the cap table is clear. This is not a science project anymore. This is infrastructure.

Neurophos is building optical processing units designed as drop-in replacements for GPUs, aimed squarely at AI inference. Not someday hardware. Datacenter hardware. Their photonic chips use metamaterial optical modulators that are roughly ten thousand times smaller than traditional photonic components. Light does the math. Power stays low. Heat stops running the meeting. Their test silicon is already demonstrating orders of magnitude gains in performance per watt, and yes, hyperscalers are paying attention even if names stay off the record.

Patrick Bowen brings the theory, the commercial scar tissue from companies like Kymeta and Lumotive, and the calm confidence of someone who has been living inside this physics for over a decade. Andrew Traverso brings the experimental rigor and validation muscle that turns whiteboard math into silicon that tapes out. Hod Finkelstein joined as Chief Technology Officer to industrialize it all, pulling experience from Meta Reality Labs, Illumina, lidar, photonics, and systems that ship, not just publish.

What makes Neurophos dangerous is not the speed of light line everyone wants to quote. It is the timing. AI inference is overtaking training. Data centers are hitting grid ceilings. Power has become the new silicon. And Neurophos is offering compute that scales without asking utilities for permission slips.