THE OPU

The Optical Processing Unit — compressing a rack of AI compute into the size and power draw of a single GPU.
SCROLL
HEAD TO HEAD
T100 OPU VS NVIDIA B200
Sustained prefill inference at INT8. Measured on public workloads. NVIDIA B200 datasheet values.
CLOCK SPEEDGHz
T100 OPU
56 GHz
NVIDIA B200
1.84 GHz
30×ADVANTAGE
ENERGY EFFICIENCYTOPS/W
T100 OPU
235 TOPS/W
NVIDIA B200
7 TOPS/W
34×ADVANTAGE
COMPUTE DENSITYrelative
T100 OPU
420×
NVIDIA B200
420×ADVANTAGE
THE PROBLEM

The constraints
are now physical.

AI demand growth has outstripped the growth of the systems that support it.

2020 DEMAND / CAPACITY 2030 AI WORKLOADS THE GAP INFRA
TWO CONSTRAINTS: MANUFACTURING AND POWER

Neurophos
Solves Both

50× COMPUTE PER WAFER. 30× PERFORMANCE PER WATT.
MANUFACTURING

How much
can you manufacture?

Every expansion plan hits the same constraint: compute per wafer.

SAME WAFER OUTPUT. FAR MORE AI.

More compute
per fab.

Fabs process a fixed number of wafers per day. Today, each wafer produces a certain amount of compute. The OPU breaks that constraint.

By compressing rack-level performance into a single package, each wafer delivers 50× the compute, on conventional production lines.

The OPU upgrades every fab into a hyperfab.

POWER

How much
can you power?

Demand keeps scaling with larger models, longer prompts, and agentic workloads. Power infrastructure is not scaling that fast.

SAME POWER. FAR MORE AI.

More compute
per gigawatt.

Power is becoming the binding constraint on AI deployment. The OPU breaks that constraint.

It delivers 30× more performance per watt on real inference workloads, compressing what would traditionally require a GPU rack into the power envelope of a single package.

TW compute at GW scale.

“Moore's Law has slowed, but AI can't afford to wait.

Our breakthrough in photonics unlocks an entirely new dimension of scaling by packing massive optical parallelism on a single chip.”

DR. PATRICK BOWEN
CEO · CO-FOUNDER
NEUROPHOS
SCALE
THE BREAKTHROUGH, IN ONE FIGURE


SMALLER

1 MILLION PHOTONIC ELEMENTS IN 5×5 MM. THIS USED TO BE THE SIZE OF A TABLE.
FOR ENGINEERS

The Questions You're
Actually Asking.

Decode is memory-bandwidth limited, not compute-limited. The industry is making strong progress there. We focus on prefill — the compute-bound bottleneck that represents 93% of inference tokens.

Photonic tensor cores operate natively at low-bit precision operands (INT4/FP4) — the same precisions used by modern inference workloads — and perform summation with effectively full precision accumulators.

Yes. Manufactured on standard CMOS. No exotic materials. No custom fabs.

Multi-chip module with electronic companion die for memory and I/O. Standard server rack integration — no rewrites to the surrounding stack.

Optical cores consume energy only for I/O — matrix multiplication becomes nearly free in the photonic domain. Efficiency improves quadratically with array size.

Tulkas T100 ships in 2028. Production capacity is allocated ahead of delivery — contact us to discuss timelines.

CONTACT

Start the
conversation.

Production capacity is allocated ahead of delivery. T100 OPU ships in 2028.

View Careers
Photonic die on fingertip

10,000×
SMALLER.

THE FAB MATH

1 million photonic elements fit in 3% of a reticle chip the size of a postage stamp. What used to fill a meter-squared optical bench now rides on the tip of a finger — on conventional CMOS lines, no exotic materials, no bespoke fabs.

≈ 33 million elements would fit in a single postage-stamp die.

CONVENTIONAL
1 m²
OPTICAL BENCH
NEUROPHOS
25 mm²
PHOTONIC DIE

Each wafer that leaves the fab now carries the compute formerly spread across a server rack. The economic unit of AI shifts from the rack to the die.

THE POWER
ENVELOPE.

ENERGY PER TOKEN

Optical cores consume energy only for I/O. Matrix multiplication, once the dominant cost of inference, becomes nearly free in the photonic domain. Efficiency improves quadratically with array size.

B200
~9
TOPS / W
T100 OPU
~300
TOPS / W

THE
BREAKTHROUGH.

PHOTONIC TENSOR CORES

A new class of optical compute fabric that performs matrix multiplication in the photonic domain. Compact, CMOS-compatible, and scaled on standard silicon — unlocking efficiency and density previously thought decades away.

DENSITY
10,000×
VS. BENCHTOP OPTICS
EFFICIENCY
40×
TOPS / W VS. GPU

Built on standard CMOS processes. No exotic materials. No bespoke fabs. Every wafer that leaves the line carries rack-scale compute on a 25 mm² die.

CONTACT.

TALK TO US

Reach the team directly — product inquiries, technical deep-dives, and partnership conversations are all welcome.

GENERAL
info@neurophos.com
PARTNERSHIPS
partner@neurophos.com
PRESS
neurophos@nextpr.com
VIA NEXTPR — OUR PRESS AGENCY
Austin, TX · Sunnyvale, CA

BUILD
WITH US.

OPEN ROLES

We're assembling a small, deeply technical team across photonic device engineering, ASIC design, compilers, and systems. If the physics of the problem excites you, we want to hear from you.

HARDWARE
Photonic Device Engineer
Pasadena, CA
HARDWARE
Senior ASIC Designer
Pasadena, CA
SOFTWARE
Compiler Engineer
Remote · US
SYSTEMS
Inference Systems Engineer
Pasadena, CA

Don't see your role? Write to careers@neurophos.com with what you'd build here.