What Neurophos does

Neurophos is a privately held semiconductor company founded in 2020 and headquartered in Austin, Texas, spun out of Duke University and the metamaterials incubator Metacept Inc. The company develops Optical Processing Units (OPUs) — photonic AI accelerator chips built on a proprietary metamaterial-based compute-in-memory (CIM) architecture in which micron-scale optical modulators, integrated onto a standard CMOS process, enable more than one million optical processing elements on a single chip. Its flagship products are the Tulkas T100 OPU (0.47 ExaOPS for MAC/GEMM at 2 kW peak, 768 GB on-chip HBM, 56 GHz) and the Tulkas T100 Server, a single-tray system integrating eight OPUs to deliver 2 ExaOPS at 10 kW, with a software stack supporting Triton and JAX.

The company is pre-revenue and targeting hyperscale data center operators, enterprise AI infrastructure buyers, and green AI data center operators as primary customer segments. Its go-to-market relies on direct enterprise sales, with production capacity allocated in advance via a "Secure Allocation" form reviewed on a rolling basis; customer evaluations are scheduled to begin in 2026, with first systems targeted for early 2028 and production ramp in mid-2028. Future revenue is expected from hardware sales of the Tulkas T100 OPU and T100 Server, with potential subscription or licensing components tied to an early-access developer hardware program.

The business has raised $118 million to date across a pre-seed from MetaVC Partners, a $7.2 million seed round in December 2023 led by Gates Frontier, and a $110 million oversubscribed Series A in January 2026 led by Gates Frontier with participation from Microsoft M12, Carbon Direct Capital, Aramco Ventures, Bosch Ventures, Tectonic Ventures, Space Capital, and others. The IP portfolio is anchored by 300 patents covering the metamaterial optical modulator breakthrough and an exclusive license to MetaVC Partners' metamaterials IP for optical computing. A pilot partnership with Terakraft to evaluate OPUs in a renewable-powered data center in Sauda, Norway, with deployments planned for 2027, marks its first commercial channel activity outside the United States.

Name

Neurophos

Legal name

Neurophos Inc.

Website

https://neurophos.com

Company type

Private

Founded year

2020

Operating status

Operating

Headcount range

11–50 employees

Short description

Neurophos is a pre-revenue semiconductor company developing metamaterial-based Optical Processing Units (OPUs) for AI inference acceleration in hyperscale data centers. Founded in 2020 and headquartered in Austin, Texas, it targets hyperscalers and data center operators seeking drop-in GPU replacements.

Ownership category

akta.pro rank

Product category

AI Inference Accelerators

NAICS

Semiconductor and Other Electronic Component Manufacturing (33441)

SIC

Semiconductors & Related Devices (3674)

akta.pro primary industry

Analog & Mixed-Signal ICs (HDAHAAAD)

Keywords

• Photonic AI chips
• Optical processing units
• AI inference accelerators
• Metamaterial optical modulators
• Compute-in-memory architecture

Headquarters

HQ city

Austin

HQ country

United States

HQ region

North America

Offices2 records

Markets served

GTM type

B2B

Offering type

Hardware or Manufacturing

Cost components

Technology or R&D, Personnel, Supply Chain, Infrastructure, Operations, Marketing or Sales

Revenue model

1. Hardware sales (planned): Pre-revenue; future revenue from sales of Tulkas T100 OPU modules and full T100 server systems to data center customers. Production capacity is allocated in advance via Secure Allocation process; first systems targeted for early 2028 with production ramp mid-2028.
2. Early access / developer hardware (planned): Plans to deliver early-access developer hardware as part of $110M Series A funding, alongside full software stack, suggesting potential future subscription or licensing components tied to the developer ecosystem.

Pricing tiers

Go-to-market motion3 records

Distribution channels3 records

Marketing channels6 records

Core offering

Neurophos designs and sells metamaterial-based Optical Processing Units (OPUs) for AI inference acceleration. Its flagship Tulkas T100 OPU integrates more than one million optical processing elements using micron-scale metamaterial optical modulators to deliver 0.47 ExaOPS per chip, and the Tulkas T100 Server packages eight OPUs in a single tray to deliver 2 ExaOPS for hyperscale data center inference workloads.

Product overview

Neurophos offers a unified photonic AI accelerator product line centered on the Optical Processing Unit (OPU). The core product is the Tulkas T100 OPU chip, which performs AI inference via metamaterial-based optical Compute-in-Memory matrix multiplication. Eight T100 OPUs are integrated into the Tulkas T100 Server, a single tray that delivers 2 ExaOPS at 10kW. The platform combines the Tulkas T100 OPU and Tulkas T100 Server with a supporting software stack (Triton, JAX) and is positioned as a drop-in replacement for GPUs in AI data centers. Customer evaluations begin 2026 with first systems targeted for early 2028 and production ramp beginning mid-2028.

Differentiator

Problem solved

Functional benefit

Brands

• Tulkas T100: Neurophos's flagship Optical Processing Unit (OPU) product line, offered as both a single OPU and a T100 server configuration targeting AI inference workloads in data centers.

Products and services

• Tulkas T100 OPU Photonic AI inference chip (Optical Processing Unit) integrating more than one million optical processing elements via micron-scale metamaterial optical modulators in a compute-in-memory architecture, delivering 0.47 ExaOPS per chip for hyperscale data center AI inference.
• Tulkas T100 Server Single-tray inference server integrating eight Tulkas T100 OPUs to deliver 2 ExaOPS of photonic AI inference throughput for hyperscale data center buyers.

Quantifiable outcome

• Up to 100x the performance and energy efficiency of current leading GPUs
• +5 more outcomes

Segments3 records

Ideal customer profiles2 records

Technology focussed Yes

API detail

Has API

No

API docs

API detail

Core technology

AI maturity

App detail

Integration2 records

AI capability2 records

Feature5 records

Partnerships

Four partnerships are on record, tiered core and flagship.

• TerakraftcoreStrategic or Co-development Partner · 21 February 2026Partnership announced February 2026 to pilot Neurophos' optical processing units within Terakraft's green AI data centre in Sauda, Norway. Pilot deployments planned for 2027. Agreement covers integration and evaluation activities to compare optical processing approaches with existing GPU-driven systems and explore gains in compute density and energy efficiency for AI workloads. Demonstrates photonic accelerators hosted in renewable-powered data centers.
• Duke UniversityflagshipStrategic or Co-development PartnerNeurophos is a spinout from Duke University and Metacept Inc. Underlying chip technology builds on research from Professor David R. Smith's Duke metamaterials lab, including prior work used to create an invisibility cloak.
• Metacept Inc.flagshipStrategic or Co-development PartnerIncubator that spun out Neurophos, led by David R. Smith of Duke University. Co-founder Metacept is Alexander Hayes. Neurophos was founded as a spinout from Metacept.
• Silicon CatalystcoreStrategic or Co-development PartnerWorld's only semiconductor-focused incubator + accelerator. Neurophos joined Silicon Catalyst incubator program in December 2023 to accelerate product development; provides privileged access to services, expertise, and IP to support startups from idea through prototype to volume production.

Scale indicators12 records

Recent moves2 records

Expansion highlights6 records

Broad incumbents

• SambaNova Systems: SambaNova builds enterprise AI accelerator systems (RDU) for training and inference, serving hyperscaler and enterprise customers. Broad incumbent in the AI inference accelerator category where Neurophos competes for workload share.
• NVIDIA: NVIDIA is the dominant incumbent in AI compute, with its GPU platform (H100, B200, Rubin) and CUDA software ecosystem representing the primary incumbent Neurophos is positioning against as a "drop-in replacement." All Neurophos performance claims are benchmarked against NVIDIA products.
• Cerebras Systems: Cerebras builds wafer-scale AI processors (WSE) for training and inference, targeting hyperscalers and large enterprises. Broad incumbent in the AI accelerator space competing for the same inference workloads and customer budgets that Neurophos is targeting.
• Groq: Groq builds LPU (Language Processing Unit) inference accelerators targeting low-latency AI inference for hyperscalers and enterprises. Broad incumbent in the AI inference accelerator market where Neurophos is positioning, though using deterministic systolic architecture rather than photonics.

Emerging players

• Mythic: Mythic builds analog compute-in-memory AI inference chips using flash memory cells to perform matrix multiplication. Comparable to Neurophos's CIM architecture approach for energy-efficient AI inference, though using electronic analog rather than photonic compute.
• Condor Computing: Condor Computing develops AI inference accelerator chips led by President Ty Garibay (a Neurophos advisor). Comparable as a deep-tech AI inference chip startup targeting similar hyperscaler workloads, though using conventional silicon rather than photonics.

Direct peers

• Lightmatter: Lightmatter is the most direct competitor — a US-based photonic AI processor company building silicon-photonic AI accelerators with compute-in-memory architecture. Both target hyperscaler AI inference with drop-in GPU replacement positioning, and both are explicitly named as best-funded US photonic AI startups in the Jon Peddie Research Q3 2025 report.
• Luminous Computing: Luminous Computing is developing a photonic AI processor for large-scale AI workloads, founded by Marcus Gomez and Mitchell Nahmias. Direct competitor in photonic AI inference accelerators with similar hyperscaler customer targeting.
• Lightelligence: Lightelligence develops photonic computing chips for AI workloads using optical matrix multiplication, founded out of MIT research. Direct competitor to Neurophos in the photonic AI accelerator space, targeting similar inference use cases for data center customers.
• Celestial AI: Celestial AI builds a Photonic Fabric architecture combining silicon photonics for memory and compute, targeting AI accelerator interconnect and in-package optical memory. Directly comparable to Neurophos as a photonic AI hardware company pursuing data center inference workloads.

Market position

Strengths5 records

Weaknesses5 records

Competitive moat5 records

Key risks6 records

Key highlights7 records

Customer concentration

• LinkedInlinkedin.com/company/neurophos

Revenue estimate

Valuation estimate

Number of profiles

Profiles13 records

Funding overview

Funding rounds6 records

Investors18 records

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

M&A

Investments

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Frequently asked questions about Neurophos