Celestial AI is developing optical interconnect technology designed to improve data movement in AI and data center systems. At the core of its innovation is the Photonic Fabric, which uses light instead of electrical signals to transmit data between processors, memory, and servers. This approach could offer faster and more efficient data movement across disaggregated systems, potentially improving scalability, a key factor in the next generation of AI workloads.
The rise of Celestial AI
Since securing a $56 million funding in February 2022, led by Koch Disruptive Technologies and joined by Temasek’s Xora Innovation, The Engine (MIT), and Broadcom, Celestial AI has attracted backing from major industry players. The company followed with a $100 million Series B in June 2023, bringing in investors including Samsung Catalyst and Porsche. Most recently, in March 2025, Celestial AI closed a $250 million Series C1 round led by Fidelity, with participation from BlackRock, Tiger Global, and Lip-Bu Tan, pushing its valuation to $2.5 billion.
Backed by growing momentum, Celestial AI’s Photonic Fabric is emerging as a critical enabler for AI infrastructure, addressing the demand for ultra-high memory bandwidth and low-latency interconnects that traditional electrical systems can no longer meet.
How does the technology work?
Celestial AI’s Photonic Fabric is an optical interconnect system designed to accelerate and optimize data movement within AI and high-performance computing (HPC) environments. Unlike traditional electrical connections that rely on copper wiring and consume more power, Photonic Fabric transmits data using light, enabling faster, more energy-efficient communication between chips and memory modules.
The technology integrates optical components such as lasers, modulators, and photodetectors into or alongside processors, including GPUs and AI accelerators, allowing high-bandwidth, low-latency communication. This design supports memory-to-memory transfers over longer distances while minimizing energy loss.
A previous article on the Taara chip explored a similar innovation, highlighting the industry’s broader shift toward photonic solutions in AI hardware (see article here: Connected by Light: The Patents Behind Taara’s Photonic Chip). This report focuses on Celestial AI’s patenting activity, with particular attention to the patents underpinning its Photonic Fabric technology.
Celestial AI: Patenting Activity
Celestial AI’s patent activity surged in 2022, driven in part by several key developments. The company’s $56 million Series A funding round, led by Koch Disruptive Technologies, likely provided both the resources and strategic incentive to accelerate its patent filings. In the same year, Celestial AI also expanded its collaboration with POET Technologies to advance and supply its cutting-edge light source module built around the “Optical Interposer.”
The Celestial AI and POET Technologies partnership began with a development and supply agreement for multi-laser external light source modules. Celestial AI launched a co-development program to integrate POET’s light engines into its AI accelerator products. By April 2023, POET introduced its Starlight™ packaged light source for AI applications, and Celestial AI placed an advance order for initial production units. Built on POET’s LightBar™ optical interposer, the modules were designed to meet Celestial’s Photonic Fabric™ goals for scalability, cost, and performance.
In October 2024, Celestial AI expanded its capabilities through the acquisition of Rockley Photonics, adding more than 200 issued and pending patents covering optoelectronic packaging, electro-absorption modulators, and optical switching, strengthening its patent portfolio and position in the photonic computing market.
Celestial AI: Top Jurisdictions
Celestial AI’s patent filings are primarily concentrated in the United States, where its core market and base of operations is located. The company also has filings under the PCT route (WO), in Europe, and in Asia, particularly in South Korea, Japan, Singapore, which are important hubs for semiconductors and advanced manufacturing.
Celestial AI: Top Law Firms
Celestial AI’s primary legal representation is handled by U.S.-based law firms, including Lewis Roca Rothgerber Christie (now merged with UK-based Womble Bond Dickinson), Ray Quinney & Nebeker, and Steven & Lee.
The company also works with several international law firms, including France’s Alatis, Japan’s Sugimura & Partners, South Korea’s Yoon & Lee, and Australia’s Spruson & Ferguson.
Celestial AI: Top Technology Areas
The majority of Celestial AI’s patents fall under the classification G02B (optical elements), G02F (optical devices and arrangements), G06N (computing arrangements), and H01L (Semiconductor devices).
This reflects the company’s focus on integrating advanced photonics with computing hardware, covering everything from optical interconnects and modulators to system-level AI processing architectures and chip packaging techniques.
Patents behind Celestial AI’s Photonic Fabric
Celestial AI is steadily expanding its patent footprint in photonic interconnect technology, with filings on optical data transfer and system memory interconnection. These filings reflect the company’s goal and efforts in becoming one of the global leaders in computer architecture advancements.
Application of light-based memory fabric to improve computing efficiency
Electronic computing systems excel in the rapid processing of data, but they are often restricted by slow and complex memory access. Current solutions to this problem are highly restrictive, requiring the memory system to be placed very close to the device using complex, expensive packaging. This also creates a technical problem as well, due to issues like thermal constraints, signal degradation, and limited scalability of memory size and capacity.
U.S. Patent No. 12,271,595 describes a photonic memory fabric, a network of light-based interconnections that links the processor and memory systems. This system replaces traditional electrical wirings with optical signals, enabling the fast and reliable transfer of data to-and-from optical routers across longer distances. By eliminating the need for memory systems to be physically close to the processor, it overcomes the bandwidth, signal integrity, and thermal limitations of existing electronic computing systems, while also improving overall computing performance.
The patent, titled “Photonic Memory Fabric for System Memory Interconnection,” was filed on September 6, 2022, and granted on April 8, 2025. The inventors are David Lazovsky, Philip Winterbottom, and Martinus Bos.
Loss and delay balance in photonic computing systems
Photonic computing systems offer an alternative path for computer advancement amidst the stagnancy of conventional von Neumann architectures, by utilizing light for data transmission and processing. However, the use of light can cause imbalance and inaccuracies in matrix calculations due to signal loss or timing delays along different light paths.
U.S. Patent No. 11,509,397 addresses this challenge by proposing a photonic circuit that systematically balances loss and delay across all optical channels regardless of distance. The light input to the circuit is split into multiple signals and then regulated via a two-stage optical modulation before being converted back into electronic signals. This design ensures that all optical paths have uniform signal loss and minimized time delays via symmetric routing. This prevents light imbalance in photonic computing, leading to accurate and consistent matrix calculations.
The patent, titled “Balanced Photonic Architectures for Matrix Computations,” was filed on December 17, 2021, and published on November 22, 2022. The inventors are Yangjin Ma, Nikolaos Pleros, David Lazovsky, George Giamougiannis, Apostolos Tsakyridis, Angelina Totovic, Martinus Bos, and Philip Winterbottom.
Single-wavelength photonic circuit design with improved scalability
Neuromorphic computing, a system that mimics how our brain functions, continues to face significant challenges. Existing systems use simplified models that result in inaccurate neuron encoding, while more complex architectures often lack scalability when applied to large datasets. These limitations have slowed the development of large-scale, reliable photonic networks needed to advance next-generation computing.
To address this, the invention outlined in U.S. Patent. No. 11,817,903 proposes a refined photonic circuit design which can utilize linear neural networks while remaining scalable and precise. The system splits a singular light carrying input signals into multiple carrier signals, then sends these into a photonic circuit which adjusts the strength and direction of the light wave to represent numerical values.
The output signals are then combined and further processed for further calculations. This solution makes it possible to perform precise calculations using light, while also ensuring high scalability for large photonic neural networks.
The patent, titled “Coherent Photonic Computing Architectures,” was filed on August 6, 2021, and published on November 14, 2022. The inventors are Nikolaos Pleros, David Lazovsky, George Giamougiannis, Apostolos Tsakyridis, and Angelina Totovic.
All featured patents were represented by Ray Quinney & Nebeker, with Hans Troesch, Micah Goldsmith, Paul Taylor et al. named on the application.
