Neurophos

Market position

Neurophos positions the OPU as an alternative to GPU-based AI inference hardware, benchmarking itself directly against NVIDIA's B200 on clock speed, energy efficiency, and compute density, and arguing that Moore's Law slowdown makes optical parallelism a new scaling dimension.

Stated differentiators are the density of its photonic tensor cores (about 1 million photonic elements on a 25 mm² die), compatibility with conventional CMOS production lines without exotic materials or custom fabs, native low-bit precision (INT4/FP4) operation with full-precision accumulation, near-zero energy cost for optical matrix multiplication with efficiency scaling quadratically with array size, and drop-in integration as a multi-chip module in standard server racks.

Technology

The core technology is a photonic tensor core: an optical compute fabric that executes matrix multiplication in the photonic domain. Neurophos states it has packed roughly 1 million photonic elements into a 5x5 mm (25 mm²) die occupying about 3% of a reticle, which it describes as about 10,000x smaller than an equivalent benchtop optical setup on a 1 m² optical bench; by that density, around 33 million elements would fit on a single postage-stamp-sized die. The devices are said to be manufactured on standard CMOS processes with no exotic materials or bespoke fabs. Photonic tensor cores operate natively on low-bit precision operands (INT4/FP4) while accumulating with effectively full-precision accumulators. Because optical cores consume energy only for I/O, the company states matrix multiplication becomes nearly free in the photonic domain and efficiency improves quadratically with array size. The OPU is delivered as a multi-chip module with an electronic companion die handling memory and I/O, intended for standard server rack integration without changes to the surrounding software stack.

Go-to-market

Direct engagement with prospective customers and partners ahead of the T100 OPU's stated 2028 shipping date, with production capacity allocated in advance of delivery. The company solicits product inquiries, technical deep-dives, and partnership conversations through its website, and uses an external press agency (NextPR) for media relations.

Operators and builders of large-scale AI inference infrastructure that are constrained by manufacturing capacity (compute per wafer) and by available electrical power, deploying accelerators in standard server racks.

Geography

Listed locations are Austin, Texas and Sunnyvale, California, with open engineering roles based in Pasadena, California and remote in the United States.