Participated · Series C · Mar 2024

Celestial AI develops the Photonic Fabric, an optical interconnect technology platform that provides seamless chip-to-chip to server-to-server networking for AI computing systems. The Photonic Fabric is designed to deliver a generational leap in bandwidth, low latency and energy efficiency compared with copper-based interconnects. The company positions the technology as foundational for data center computing, networking and memory systems to enable next‑generation AI architectures. Celestial AI plans to accelerate commercial deployment, strengthen its volume manufacturing supply chain, scale production capabilities, and deepen strategic partnerships with foundries, including TSMC. Company leadership says the technology addresses the data‑movement limits of copper interconnects and enables system designs that can scale processors while reducing power consumption. Financially, Celestial AI has raised $520 million to date, including a final close of a $255 million Series C1 to support commercialization and supply‑chain scaling. Celestial AI is the creator of the Photonic Fabric, an optical interconnect technology platform for compute and memory. The Photonic Fabric enables disaggregation of compute and memory and, the company says, delivers over 25x greater bandwidth and memory capacity while reducing latency and power consumption by up to 10x versus existing optical interconnect alternatives and copper. The company reports successful validation of the technology in silicon and says this has driven tremendous demand from hyperscalers. Amid growing demand for generative AI and next‑generation data centers, Celestial AI positions the Photonic Fabric as a solution to constraints including utility power availability, memory capacity and high cost of operation. The company plans to use the Series C proceeds to execute multiple large‑scale customer collaborations focused on commercializing the Photonic Fabric platform. Celestial AI presents the Photonic Fabric as the foundational technology for optically scalable, disaggregated data center compute and memory to enable sustainable and profitable AI advancements. Celestial AI has developed a photonics-based optical interconnect architecture for compute-to-compute, compute-to-memory and on-chip data transmission aimed at addressing data-movement bottlenecks in AI workloads. The company says its photonic fabric scales across multiple-chip systems, is compatible with industry interconnect standards such as CXL and PCIe, and can enable memory disaggregation. Celestial claims its technology delivers 25x higher bandwidth and 10x lower latency and power consumption than optical alternatives and can reduce the electricity spent on data movement, indirectly boosting a chip's effective compute capacity. It offers the interconnect through a licensing program and reports engagement with several "tier-one" customers, including hyperscalers and processor and memory companies. Celestial also sells an AI accelerator chip, Orion, built on its photonics architecture, though the article notes engineering challenges remain for photonics-based chips at scale. Founded by David Lazovsky and Preet Virk with architect Phil Winterbottom, the company began as a portfolio company of The Engine. The company has around 100 employees and expects to grow to about 130 by the end of the year. Celestial AI builds ML/HPC accelerators based on its proprietary Photonic Fabric™ that uses integrated photonics for high-bandwidth, low-power data movement and optical access to memory and compute. Its core product line is the Orion AI accelerator family, which combines electronic compute with photonic data movement and accompanying system software for simpler, efficient mapping of data and compute. The company says this hybrid approach reduces power spent on data movement and enables redistribution of power to compute, improving performance, latency, and TCO versus electronic-only systems. Celestial AI positions its software and multi-chip scaling as competitive differentiators as AI model complexity grows and cites an Omida addressable market projection of over $70 billion in 2025. The company plans to use new capital to expand its global engineering team, advance product development, and pursue strategic supplier engagements, including Broadcom, to commercialize Orion products. No revenue or user metrics are reported in the article.

Participated · Seed · Apr 2021

Deeplite offers a platform, Neutrino, that transforms trained deep neural networks into highly compressed models tailored for constrained hardware so they can run on devices like cell phones, drones or Raspberry Pi. The product lets customers specify deployment targets and compression levels, trading small amounts of accuracy for large reductions in model size (examples cited include shrinking 200 MB to 11 MB and 50 MB to 100 KB). The company positions its software to reduce compute, memory and battery requirements for edge AI deployments. Deeplite launched officially in mid-2019 with roots in the TandemLaunch incubator and is based in Montreal. It has 15 employees and plans to double headcount by the end of the year, with hubs planned in Montreal and Toronto and flexible remote work options. The company recently closed a $6M seed investment to support product and team growth.