Skip to main content

LightOn (France Silicon Photonics AI Inference)

Product Overview

LightOn is a French silicon photonics AI inference chip company, founded 2016, headquartered in Paris, originating from French National Center for Scientific Research (CNRS) + Ecole Polytechnique. Alfred is its flagship photonics inference platform, 2022 released, 1.2 PFLOPS FP16 compute, ~80W TDP, one of the most power-efficient AI inference chips. Known as one of the "Silicon Photonics Big Three" alongside Lightmatter (US) and Lightelligence (China).

Strategic significance: LightOn is the strategic choice for European AI compute (France/Germany/EU GDPR + AI Act compliance), with Microsoft Azure France data center + OVHcloud European Cloud already integrating LightOn inference API.

Core Specs

ItemParameter
ArchitectureLightOn Alfred (Optical Processing Unit)
ProcessHybrid: silicon photonics + standard CMOS controller
Photonic Cores32 MZI meshes (each 4x4)
Digital Cores24x ARM Cortex-A78 (scheduling + activation)
HBM32GB HBM2E (digital portion)
Memory Bandwidth819 GB/s
FP161.2 PFLOPS
FP32~0.6 PFLOPS
INT82.4 POPS
TDP~80 W
Efficiency15 TOPS/W
Form FactorPCIe Gen4 x16
InterconnectLightOn Photonic Link
Mass Production2023-H1
Unit Price~$3,000-5,000

Silicon Photonics Big Three Comparison

DimensionLightOn Alfred (France)Lightmatter Envise (US)Lightelligence Tianjic (China)
Compute FP161.2 PF2 PF1.6 PF
TDP80W65W100W
Efficiency15 TOPS/W30 TOPS/W16 TOPS/W
Mass Production2023-H12023-H22023-H2
CountryFranceUSChina
EcosystemLightOn ApplianceIdiomTianjic Stack
CustomersAzure France, OVHcloudGoogle internal, MetaChinese supercomputing centers
Price~$4K~$6K~$4K

LightOn advantage: Europe's only commercial silicon photonics (GDPR + AI Act compliant), Microsoft + OVHcloud integration. A key focus of France's national AI strategy.

LightOn Appliance Platform

ItemSpec
Form Factor1U / 2U rack server
Alfred Cards4 / 8
Total Compute4.8 PF / 9.6 PF FP16
Total HBM128GB / 256GB
Optical Interconnect1.6 TB/s intra-domain
Rack TDP320W / 640W
Price~$20K / $40K

LightOn Cloud Services:

  • LightOn Cloud API (OpenAI compatible)
  • OVHcloud AI Endpoint (European cloud)
  • Microsoft Azure France region (2024-Q4 integrated)

Software Stack LightOn Appliance

LayerToolDescription
AI frameworkLightOn AppliancePyTorch / TensorFlow compatible
LightOn-Oumiopen-source inference optimization
CompilerPhoton Compileroperators -> MZI config + ARM code
RuntimeLightOn Runtimemulti-card scheduling
APILightOn APIOpenAI compatible (chat + completions)
Model ZooModelZooMistral / LLaMA / Qwen / SDXL

LightOn ecosystem advantage: open-source Oumi library (GitHub 3K+ stars), GDPR-friendly for Europe, Microsoft Azure integration.

Vendor Information

ItemContent
CompanyLightOn SAS
FounderIgor Carron (CEO, former CNRS researcher)
Founded2016
HeadquartersParis, France
Funding$50M+ (Series A 2022 led by: BNP Paribas + European Innovation Council)
Valuation (2025)~$200M
2024 Revenue~$8M (early commercialization)
Employees~80
FabSTMicroelectronics (European fab) + CEA-Leti (silicon photonics process)
CustomersMicrosoft Azure France, OVHcloud European Cloud, French National Research Center
National ProjectsFrance AI National Strategy, EU Horizon Europe, EuroHPC
Patents80+ silicon photonics AI patents

France / European AI Strategic Significance

  • France AI National Strategy: Macron 2018 AI strategy, LightOn is France's AI compute representative
  • EU AI Act compliance: European AI regulation, data + compute must be localized (LightOn is Europe's only option)
  • EuroHPC project: European supercomputing, LightOn is a candidate
  • GDPR data protection: European data cannot leave for US/China cloud (LightOn is GDPR-native)
  • Microsoft Azure France region: Microsoft + LightOn partnership
  • OVHcloud European Cloud: France's largest cloud provider integration

Use Cases

  • European AI inference (GDPR + AI Act compliant)
  • Low-power AI inference (80W air-cooled)
  • France / European government AI (sovereign AI)
  • Microsoft Azure France integration
  • Enterprise LLM inference (Mistral 7B/8x7B optimized)
  • ❌ AI training (inference only)
  • ❌ US/China market (geographic compliance)
  • ❌ CUDA proprietary workloads

LightOn Product Line

ProductReleasedCompute FP16TDPTarget
Alfred 120221.2 PF80WCloud inference
Alfred 2 (est.)2025 H22.5 PF120Whigh-end inference
Alfred Mini2023200 TF25Wedge inference
Photon Link (est.)2026--optical interconnect IP

Key Features

  • Silicon Photonics Computing: Europe's only commercial photonic AI
  • 15 TOPS/W: H100 7x efficiency
  • 80W TDP: air-cooled deployment
  • GDPR / AI Act compliant: European sovereign AI
  • Microsoft Azure integration: France region
  • Open-source Oumi library: open ecosystem
  • Drawbacks: compute 40% below Lightmatter, 3-4 year ecosystem

France AI Ecosystem

CompanyDomain
LightOnsilicon photonics AI inference
Mistral AIopen-source LLM (7B, 8x7B, 8x22B)
Hugging FaceAI model platform (HQ Paris)
Owkinmedical AI
Aleph AlphaGerman sovereign LLM