Lightmatter Passage

Learning Objectives

• Understand what photonic interconnects are and why they matter for AI data centers
• Identify the key Passage products (M1000 and L200) and their specifications
• Evaluate the potential impact of photonics on AI infrastructure power consumption and performance

What Is Lightmatter Passage?

Lightmatter Passage is a platform of photonic interconnect chips that use light instead of electricity to connect AI processors in data centers. In today's largest AI training clusters — thousands of GPUs working together — the wires connecting those chips consume nearly 30% of total data center energy. Copper electrical interconnects degrade over distance, create heat, and become a bottleneck as clusters grow larger.

Passage solves this by replacing copper with silicon photonics: microscopic waveguides etched into silicon that carry data as pulses of light. The result is dramatically higher bandwidth, lower power consumption, and the ability to connect thousands of GPUs in a single communication domain without signal degradation.

Passage Products

Passage M1000 ("Photonic Superchip")

The M1000 is a 3D photonic interposer that sits underneath the GPU or AI accelerator, providing optical connectivity across the entire die area.

Passage L200 / L200X

The world's first 3D co-packaged optics (CPO) product, designed for next-generation AI chip packages:

Why This Matters for AI

The interconnect problem is becoming the dominant bottleneck in AI scaling:

• Power: Interconnects consume nearly 30% of data center energy. Passage achieves 4.6 pJ/bit versus approximately 15 pJ/bit for legacy copper — a roughly 3.5 times power reduction
• Bandwidth: The M1000 delivers 114 Tbps optical bandwidth, enabling 5 to 10 times more I/O than chip-to-chip electrical interconnects
• Distance: Optical signals travel much farther than copper without degradation, enabling physically larger GPU clusters
• GPU utilization: Faster interconnects mean GPUs spend less time waiting for data and more time computing — directly improving training speed and cost-efficiency

As Lightmatter CEO Nick Harris stated: "Over the next few years, all of the GPUs in the world that are designed for AI training and inference are going to be built on Lightmatter's Passage."

Manufacturing Partners

Lightmatter has assembled a serious supply chain for production:

Passage vs. Competitors

Lightmatter's key differentiator: The M1000 is a 3D photonic interposer that sits beneath the processor, providing "edgeless I/O" across the entire die area. Competitors place optical chiplets in limited space beside the die — which Lightmatter's CEO calls a "generation 2.5" approach. The M1000 also includes built-in solid-state optical circuit switching, which no competitor offers.

Company Details

Strengths

• Fundamental technology shift — photonics addresses the interconnect bottleneck that copper cannot solve as AI clusters scale to thousands of GPUs
• 3.5 times power reduction over copper interconnects — critical as data center power consumption becomes a limiting factor for AI scaling
• 114 Tbps bandwidth (M1000) — far exceeds competing photonic and electrical interconnect solutions
• Built-in optical switching — the M1000 includes solid-state optical circuit switching on-chip, a unique capability
• Strong supply chain — manufacturing partnerships with GlobalFoundries, Amkor, GUC, and ASE for production at scale

Limitations and Considerations

• Pre-revenue — no confirmed customer deployments as of March 2026; the M1000 was scheduled for summer 2025 sampling but shipping status is unconfirmed
• No named customers — the GUC partnership signals hyperscaler interest, but specific customers have not been disclosed
• NVIDIA relationship unclear — NVIDIA has its own photonics roadmap (co-packaged optics by 2026) and could be a collaborator or competitor
• Small team — approximately 165 employees building a product that needs to integrate into the world's largest chip packages
• Technology risk — silicon photonics at this scale is unprecedented; manufacturing yields and reliability at production volumes are unproven

Key Takeaways

• Lightmatter Passage uses light instead of electricity to connect AI processors — delivering 114 Tbps bandwidth and 3.5 times power reduction over copper interconnects
• Addresses a critical bottleneck: interconnects consume nearly 30% of data center energy and become the limiting factor as AI clusters scale
• The Passage M1000 is a 4,000+ square millimeter photonic superchip with 34 integrated chiplets and built-in optical circuit switching — the most ambitious photonic interconnect ever built
• Valued at $4.4 billion with $850 million raised; manufacturing partnerships in place with GlobalFoundries and Amkor; customer deployments expected but not yet confirmed