# Neurophos

> Neurophos builds photonic compute for AI inference. Its Optical Processing Unit (OPU) compresses a rack of AI compute into the size and power draw of a single GPU — 50× faster and 30× more energy efficient than an NVIDIA B200 on sustained prefill inference. Backed by a $110M Series A (announced January 2026). Offices in Austin, TX and Sunnyvale, CA.

Key facts:

- Tulkas T100 OPU (Optical Processing Unit): 0.47 ExaOPS (dense FP4/INT4 GEMM), 235 TOPS/W, 2 kW peak power (1 kW average), 768 GB HBM at 20 TB/s, 2 GB L2 cache. Programmable via Triton and JAX.
- Tulkas T100 Rack: full rack-scale photonic compute system built on the T100 OPU — 33.84 ExaOPS (dense FP4/INT4 GEMM), 191 TOPS/W including rack and networking overhead, 176.6 kW peak (155 kW average), 55.3 TB HBM at 1.44 PB/s, 144 GB L2 cache.
- Timeline: first systems target early 2028; production ramps mid 2028. Production capacity is allocated ahead of delivery.
- How it works: photonic tensor cores clock at 56 GHz (~30× NVIDIA Blackwell) and operate natively at low-bit precision operands (INT4/FP4) — the same precisions used by modern inference workloads — with summation performed in effectively full-precision accumulators.
- Why inference: the ratio of input to output tokens is roughly 15:1 and growing; ~93% of tokens are prefill, which is compute-bound. The OPU is built to speed up the compute core itself.
- Important distinction: VARDA and MANWE are future generations on the Neurophos scaling roadmap, NOT the upcoming product. Roadmap-scale figures (e.g. 420× Blackwell effective speed, 1,400 TOPS/W) describe that roadmap trajectory — the shipping product is the Tulkas T100.
- Manufacturing: built on conventional CMOS production lines; each wafer delivers ~50× the compute of GPU wafers, on standard fab equipment.

## Pages

- [Home](https://www.neurophos.com/): what the OPU is, the manufacturing and power constraints it solves, and the T100 vs NVIDIA B200 comparison
- [Product](https://www.neurophos.com/product): full Tulkas T100 Rack and T100 OPU specifications
- [The Breakthrough](https://www.neurophos.com/The%20Breakthrough): why optical compute failed for 60 years, what Neurophos changed (metamaterial optical transistors ~10,000× smaller), and the scaling roadmap
- [Team](https://www.neurophos.com/team): founders and leadership — veterans of NVIDIA, Apple, Samsung, Intel, AMD, Meta, ARM, Micron, Mellanox
- [News](https://www.neurophos.com/news): press coverage (TechCrunch, Tom's Hardware, The Register, Bloomberg) and company press releases
- [Videos](https://www.neurophos.com/videos): technical talks, podcast appearances, and coverage
- [Whitepapers](https://www.neurophos.com/whitepapers): technical whitepapers and research
- [Careers](https://www.neurophos.com/careers): open roles in Austin, TX and Sunnyvale, CA

## Contact

- General: info@neurophos.com
- Partnerships: partner@neurophos.com
- Press: neurophos@nextpr.com