PicoJool's $27.5 million Series A pays for customer qualification of its low-power AI optics

"Reaching 200G per lane proved that the performance is there. Our focus now is execution," PicoJool chief executive Al Yuen said ^[6]. He went on: "This funding allows us to expand our teams in the U.S. and Taiwan, qualify products with customers, and move a complete portfolio from chip-level VCSELs and microVCSELs into active optical cable (AOC) and near-packaged optics (NPO) modules." ^[7] A funding round does not change what an optics engineer at a cloud operator does next with a PicoJool sample cable, and Yuen says the company is at the same step.

The pitch is about copper. SiliconANGLE reported that copper interconnects cannot keep up with the bandwidth demands of the largest AI clusters and use vast amounts of energy ^[15]. Gelsinger framed it as a system problem. "The network is the AI, and connectivity is becoming one of the defining constraints on performance," he said ^[16].

Most of the low-power case comes from the smaller parts. According to DatacenterDynamics, the first product line, unveiled this summer, targets scale-up links and supports more than 37 gigahertz of bandwidth ^[4]. The same report said microVCSELs use far less power than conventional wire-bonded lasers ^[5]. PicoJool says its MicroVCSEL platform aggregates parallel lanes to reach up to 3.2 terabits per second ^[13]. Built from its 50G lanes, that would take 64 of them ^[17]. Neither report includes an energy-per-bit figure or a customer name. Without the first, an operator cannot set "far less power" against a copper cable or a rival module in a rack power budget.

For a buyer, sourcing is the stronger argument. DatacenterDynamics reported that gallium arsenide VCSELs, unlike co-packaged optics, are far from supply-chain constrained ^[11]. PicoJool has company in that bet. Coherent and Lumentum, both former employers of Yuen, are looking at VCSELs for AI interconnect, and a Scottish startup is developing PCSELs, which it argues offer better beam quality and higher power output ^[14].

Operators buy the part that passes their own tests and shows up in volume on the promised date.

A 2x2 sorts the decision. One axis is whether the supplier's laser supply is constrained. The other is whether the part has passed qualification in your own racks. Open supply and a qualified part is a purchase order. A constrained part that has passed testing calls for reserving allocation early. Constrained and unproven belongs on a watch list. On the public record, PicoJool sits in the fourth box: open supply, qualification still ahead ^[11][7]. I'd give that box a sample request and a test slot, and hold off on designing a rack around it. Waiting risks a later place in the queue if the parts pass; committing now ties a rack design to one startup's schedule.