lightmatter.co

What It Is

Lightmatter is a “photonic supercomputing” company headquartered in Mountain View, California. Its official positioning is photonic chips for AI supercomputing. It is not a ChatGPT-style AI application or a developer SaaS tool; instead, it provides high-bandwidth, low-power photonic interconnect platforms for GPUs, TPUs/XPUs, switches, and AI data centers.

Core Capabilities

Its core products are the Passage photonic interconnect platform and the Guide VLSP light engine. Passage spans NPO, OBO, 2D/3D CPO, and 3D photonic interposers, focusing on solving the traditional chip I/O bottleneck caused by edge “shoreline” limitations. The L200 can deliver 32-64 Tbps of optical I/O, the L20 targets high-density NPO/OBO deployments, and the M1000 EVK discloses a 114.6 Tbps bidirectional system data rate. Guide integrates large numbers of lasers into a light engine: Guide 1 supports 51.2 Tbps I/O per module, while Guide DR uses a liquid-cooled LNIC form factor and integrates 64 high-power lasers.

AI Use Cases and Integration

Lightmatter’s value is mainly in frontier AI training and inference infrastructure: 100,000-GPU coherent fabrics, trillion-parameter MoE training, multimodal foundation model training, and inference scale-up networks. The official site cites research suggesting that 3D CPO can reduce training time for trillion-parameter MoE models by 2.7x. At the interface level, Guide products disclose CMIS 5.3 over I²C/I3C, while vClick, Corning Flexconnect, BGA, and cold plates also show that the company is aimed at hardware system integration rather than software APIs.

Pricing and Trial

The official website does not publish pricing, licensing models, or procurement thresholds. Multiple pages use entry points such as Contact sales, Request evaluation kit, early access partners, and design partnerships, indicating that its business model is closer to enterprise hardware customization and joint validation. There is no public free tier or self-service trial.

Pros and Cons

The advantages are a comprehensive technical roadmap covering optical interconnects, light sources, packaging, and serviceable connectors, along with disclosed hard metrics such as bandwidth, energy efficiency, wavelengths, and packaging compatibility. It also emphasizes rack-level validation and manufacturing partners. The drawbacks are that public information mainly comes from the vendor’s own materials, with limited third-party benchmarks, real customer deployment details, pricing, or mass-production delivery information. It is also almost impossible for ordinary AI application developers to use directly.

Who It Is For and Access from China

Lightmatter is suitable for hyperscale cloud providers, AI data centers, switch/XPU vendors, and teams working on CPO/NPO/OBO solutions. It is not suitable for users looking for ready-made AI writing, image generation, or automation tools. Access from China, payment options, local sales, and Chinese-language support have not been disclosed and should be considered unknown. Domestic teams evaluating similar directions may also want to watch NVIDIA/Mellanox, Broadcom, Marvell, Intel, and other silicon-photonics interconnect solutions.