Neurophos OPU

Neurophos announced its Optical Processing Unit (OPU), a photonic AI chip achieving 0.47 ExaOPS at 300 TOPS/W, claiming 10,000x size reduction over traditional photonic tensor cores. The company says a single tray of 8 OPUs outperforms a 3,000 lbs GPU rack while consuming only 10 kW, with evaluations starting in 2026 and production in 2028.

top of page THE OPU Compressing a rack into the size and power draw of a single GPU. 0.47 ExaOPS The most efficient AI chip on the planet 300 TOPS/W THE BREAKTHROUGH Photonic Tensor Cores Are Enormous Neurophos Made Them 10,000x Smaller The elements required for ExaOPS speeds occupy 1m x 1m. Demonstrated. Protected by 300 patents. 30 YEAR LEAP A LEAP EQUIVALENT TO 30 YEARS OF TRANSISTOR SCALING THE WORLD DEMANDS TERAWATT COMPUTE This requires more power than the United States produces today With GPUs With OPUs - 1000+ nuclear plants - $20 Trillion for GPUs - 10 nuclear plants - 100x fewer XPUs 10 kW 2 ExaFLOPS Single tray. 8 OPUs. Faster than a 3,000 lbs GPU rack. TERAWATT COMPUTE. GIGAWATT POWER. Evaluations begin 2026. First systems targeted for early 2028. Production ramp begins mid 2028. bottom of page