Columbia Researchers Help Power Faster, More Energy-Efficient AI
Xscape Photonics, co-founded by Columbia Engineering researchers, has raised $37 million in new funding and launched a multiwavelength laser designed to move data faster and more efficiently across increasingly powerful AI systems.
As artificial intelligence systems grow larger, moving enormous amounts of data between the chips that power them has become a major engineering challenge. A startup founded with technology and expertise from Columbia Engineering is working to address that bottleneck using light.
Xscape Photonics, a semiconductor startup co-founded by Columbia photonics researchers, recently announced $37 million in new funding, bringing its total Series A investment to $81 million. The company also launched FalconX, a new laser device designed to increase the speed, capacity and energy efficiency of data transmission inside AI data centers.
Founded in 2022, Xscape grew out of research in photonics, a field that uses light to transmit and process information. The company was founded by CEO Vivek Raghunathan and four Columbia researchers: Alexander L. Gaeta, David M. Rickey Professor of Applied Physics and Materials Science and professor of electrical engineering; Yoshi Okawachi, a research scientist in electrical engineering; Keren Bergman, Charles Batchelor Professor of Electrical Engineering and scientific director of the Center for Integrated Science and Engineering; and Michal Lipson, Eugene Higgins Professor of Electrical Engineering and professor of applied physics. At Xscape, Gaeta serves as president, Okawachi as vice president of research and development, and Bergman and Lipson as members of the board of advisers.
Their work builds on decades of Columbia research into controlling light at extremely small scales. Lipson pioneered foundational technologies in silicon photonics, which uses silicon chips to manipulate and transmit information using light. Gaeta's research in nonlinear optics and integrated photonics explores how light can be generated and manipulated on chips. Bergman's research focuses on optical interconnects and photonic architectures that can move large amounts of information across computing systems with greater speed and energy efficiency.
That expertise is becoming increasingly important as AI places new demands on computing infrastructure.
Modern AI systems rely on large clusters of graphics processing units, or GPUs, and other accelerators working together. Those processors need to exchange vast quantities of information quickly. As clusters grow, the connections between processors can become a bottleneck, limiting performance and increasing energy demands.
Xscape is developing optical connections that use multiple wavelengths, or colors, of light to carry information. Instead of relying only on electrical signals traveling through traditional copper connections, photonic systems can transmit large amounts of data using light.
The company's newly announced FalconX device uses its CombX laser technology to generate as many as eight wavelengths of light on a single silicon photonics chip. Each wavelength can serve as a separate channel for transmitting information, allowing multiple streams of data to travel simultaneously. The technology is designed to support multi-terabit-per-second data movement while reducing the number of components needed to support those connections.
The approach targets what Xscape calls the "escape bandwidth" problem: the limited capacity for data to move between AI accelerators in large computing clusters. As AI models and the systems that run them become larger, improving those connections can help processors work together more effectively.
Reliability is another challenge. In a large AI cluster, the failure of a single laser can affect workloads across the network. FalconX includes built-in redundancy designed to keep the system operating if an individual laser component fails.
The technology also addresses a broader challenge facing the rapid expansion of AI: energy consumption. Moving data within and between computing systems requires significant power, and those demands increase as AI infrastructure scales. Photonic interconnects offer a potential way to move more information while improving energy efficiency, an area Columbia researchers have been exploring as part of efforts to develop more sustainable high-performance computing systems. Bergman's research, for example, focuses on nanoscale photonic networks that can move information across chips, memory and large computing systems with greater energy efficiency.
Xscape is already looking beyond eight wavelengths. Its ChromX platform is designed to scale to 16, 32 and eventually more than 128 wavelengths, creating many parallel channels for moving information through future AI data centers.
The latest $37 million investment was led by Addition, with continued participation from investors including NVIDIA and IAG Capital Partners. The financing will support the company's efforts to develop and scale its multiwavelength technology for AI infrastructure.
For Columbia researchers, Xscape's growth shows how fundamental research into the behavior of light can move from the laboratory into technologies addressing emerging engineering needs. As AI systems demand ever more computing power, advances in photonics could help build infrastructure that moves information faster and more reliably while using energy more efficiently.