PicoJool, a startup developing optical interconnects for AI infrastructure, has closed a $27.5 million Series A led by Socratic Partners, with Hudson River Trading participating. The round brings total capital to $39.5 million, following a $12 million seed in November backed by Playground Global, the venture firm associated with former Intel CEO Pat Gelsinger.
The company is advancing Vertical Cavity Surface Emitting Laser (VCSEL) technology—semiconductor laser diodes that transmit data via light signals through optical cables. The pitch: replace copper wire interconnects within data centers with a faster, lower-power alternative.
The economics are straightforward. AI chip clusters—often housing hundreds or thousands of GPUs—must move massive amounts of data between processors to train and run inference on large language models. Copper interconnects consume significant power and hit bandwidth ceilings. PicoJool's 100G, 200G, and configurable MicroVCSEL modules (50G and 64G variants) aim to compress that problem. Founder and CEO Al Yuen said the company has demonstrated 200G per lane performance and claims its architecture can aggregate bandwidth to 3.2 terabits per second while cutting energy costs per bit moved—a direct lever on data center operating expense.
The capital will fund customer qualification, team expansion in the U.S. and Taiwan, and progression from chip-level components to finished AOC and NPO modules. Yuen framed the goal plainly: "Our current focus is execution." The company is working with WIN Semi on manufacturing.
The timing is real. Hyperscalers building out clusters for frontier AI models are openly worried about interconnect bandwidth and power budgets. But PicoJool faces steep incumbent moats. Mellanox (now Nvidia), Broadcom, and internal custom silicon teams at Meta and Google already own these interfaces. Winning requires not just better specs but willingness from customers to redesign clusters around a new supplier—a multi-year qualification cycle with no guarantee of design wins.
