Building What Comes Next
AI is transforming everything. But the infrastructure powering it is approaching physical limits. Soon, computing will consume more power than entire countries. At Lightmatter, we’re building the photonic foundation for the next era of computing—interconnects, lasers, and eventually compute itself.
Energy Crisis
AI’s gigascale inflection
In Texas alone, the interconnection queue for large loads has reached 233 GW—nearly triple the state’s record peak demand.
That figure far exceeds the state’s total generation capacity of 103 GW. ERCOT now projects data center demand could reach 78 GW by 2030, driven by a new class of “gigascale” campuses.
Individual AI campuses now require up to 2 GW each, pushing many operators to bypass the grid entirely and rely on on-site, behind-the-meter generation to power the next generation of frontier AI.
233 GW
Total large-load connection requests currently in the ERCOT queue
103 GW
Total installed generation capacity available to the Texas grid
78 GW
Revised 2030 projection for Texas data center demand
10x
Increase in rack power density (from 10 kW to 100+ kW) for AI clusters
Bottlenecks
Two walls, one solution
The computing industry faces two fundamental constraints. Both have the same answer.
The Interconnect Wall
GPUs spend 40-60% of their time waiting for data. As AI clusters grow from thousands to hundreds of thousands of chips, interconnect bandwidth becomes the dominant constraint.
The number of connections grows quadratically with cluster size. Traditional electrical interconnects can’t scale—they’re limited by the chip’s physical perimeter, by cable weight, by power consumption. Interconnect is all you need.
The Silicon Wall
Moore’s Law no longer holds. Transistors have shrunk to near-atomic scales. Quantum tunneling, leakage currents, and heat dissipation create fundamental physical barriers. We’re approaching the final nodes—the last meaningful process shrinks in silicon’s 60-year history.
The industry is responding the only way it can: building bigger chips. Die area is growing 13–18% annually. By 2040, processors will reach wafer-scale—40,000 mm². But bigger chips need even more bandwidth, compounding the interconnect crisis. We need a new approach to computation itself.
Solution
Photonics from interconnect to compute
Light has extraordinary properties. Photons travel at nature’s speed limit, cross paths without generating heat, and allow multiple wavelengths to share a single fiber. Unlike electrons, they avoid the resistive losses, capacitance, and inductance that limit electrical signals across a datacenter. Lightmatter is building a complete photonic stack—solving today’s interconnect challenges while laying the foundation for tomorrow’s breakthroughs in computing.
Interconnect
Passage
3D-integrated photonic circuits delivering up to 114 Tbps of bandwidth. Edgeless I/O eliminates the shoreline bottleneck, enabling massive GPU clusters to scale beyond the limits of electrical interconnects.
Explore Passage →Laser
Guide
External Light Source (ELS) technology integrates hundreds of lasers on a single chip. A software-defined, self-healing light engine designed to power next-generation photonic infrastructure.
Explore Guide →Compute
Envise
A photonic AI accelerator and the world’s first photonic processor executing production AI workloads. Matrix multiplication performed with light at the speed of physics, published in Nature in April 2025.
Receive updates →The Future
A new kind of computer
Envise is a photonic AI accelerator that performs matrix multiplication using light. Inside the device, four photonic chips manipulate 512 light beams through over 200,000 optical components.
The physics are elegant: when light passes through an optical system, interference naturally computes linear transformations. No transistors switching. No electrons moving. Just photons, lenses, and the laws of physics doing math at the speed of light.
Efficiency
262T
Ops/Watt
Latency
200ps
Per matrix-vector product
Photonics
200K+
Components
Electronics
50B
Transistors
Demonstrated workloads
ResNet
Image classification demonstrated at near-electronic precision.
BERT
Natural language processing and sentiment analysis workloads.
NanoGPT
Text generation demonstrated using the TinyShakespeare dataset.
DQN Atari
Deep reinforcement learning using DeepMind’s DQN algorithm.
Semantic Segmentation
Pixel-level segmentation of animals in the Oxford-IIIT Pet dataset.
“We show that you can build a computer that’s not based on transistors and run state-of-the-art workloads. This is an essential step towards developing post-transistor computing technologies.”
— Nicholas Harris, PhD, Founder & CEO
Universal photonic artificial intelligence acceleration
Nature Volume 640, Issue 8058 — April 2025
Read the full paper →The Journey
From MIT lab to production
A decade of breakthroughs in silicon photonics has transformed a research project into production infrastructure for AI. What began in an MIT lab is now powering the next generation of hyperscale computing.
2017
MIT $100K
Three MIT researchers, Nicholas Harris, Darius Bunandar, and Thomas Graham, win the MIT $100K Entrepreneurship Competition with a vision for photonic computing. Working in Professor Dirk Englund’s quantum photonics lab, they developed silicon photonic circuits capable of performing matrix operations with light. Lightmatter is founded.
2020
Envise at Hot Chips
Lightmatter presents silicon photonics for AI at Hot Chips, introducing Envise, the world’s first photonic AI accelerator. The demonstration signals that photonic processors can accelerate neural networks and begins to draw attention across the industry.
2021
Passage unveiled
Lightmatter introduces Passage, a photonic interconnect platform designed to overcome the limits of electrical I/O. The company’s roadmap becomes clear. Solve the interconnect bottleneck first, then transform computation itself.
2022
Passage at Hot Chips
At Hot Chips 34, Lightmatter presents Passage to the broader semiconductor community. The platform brings silicon photonics and co packaged optics to the emerging chiplet era and advances the vision for edgeless I/O architectures.
2024
$4.4B valuation
Lightmatter raises $850M and reaches a $4.4B valuation. Strategic partnerships with GlobalFoundries and Amkor Technology establish a path to high volume manufacturing. The industry increasingly recognizes that photonic interconnects are essential for scaling AI infrastructure.
2025
Production and Nature
Passage enters production and ships to customers. At SC25, Lightmatter demonstrates four racks of production hardware, including M1000 and Passage 50. The first 16 wavelength bidirectional link achieves eight times the bandwidth density of existing solutions. In April, the paper “Universal photonic artificial intelligence acceleration” is published in Nature, demonstrating photonic processors executing state of the art neural networks.
The Horizon
A decade-scale roadmap
The goal is not faster connections between isolated chips. It is a datacenter that operates as a unified system, with thousands of processors working as one machine. Today we are solving the interconnect crisis. Tomorrow comes photonic compute. The same physics that make light superior for moving data also make it ideal for processing it. We are building toward that future one layer at a time.
2026
Near-Packaged Optics
Passage near-packaged optics (NPO) systems enter production deployments. Optical connectivity moves closer to compute, improving bandwidth density and power efficiency for next-generation AI clusters.
2027
Co-Packaged Optics
Passage co-packaged optics (CPO) integrates photonic interconnects directly alongside XPUs. This architecture eliminates traditional pluggable bottlenecks and unlocks dramatically higher system bandwidth.
2028+
Photonic Interposer
Passage evolves into an active photonic interposer for chiplet-based systems. Optical connectivity becomes a native part of advanced packaging, enabling dense chiplet fabrics and new compute architectures.
2030+
Full Photonic Stack
Advancing a complete photonic computing stack integrating Envise compute, Passage interconnects, and Guide laser technology for frontier AI infrastructure.
Build the future with us
We’re looking for exceptional engineers, researchers, and leaders who want to work on problems that matter. The transistor defined the last 60 years of computing. What comes next is up to us.