PicoJool Raises $27.5M Series A to Scale AI Data Center Optics

PicoJool has raised $27.5 million in Series A funding to commercialize optical connectivity technology for increasingly dense AI data centers. Socratic Partners led the round, with participation from Hudson River Trading. The Palo Alto company's total funding now stands at $39.5 million, following a $12 million seed round led by Playground Global.

The round targets a problem that raw processor performance cannot solve: moving data quickly and efficiently between accelerators. As AI clusters grow, the links connecting GPUs and other processors increasingly determine how much of that expensive compute stays productive.

PicoJool plans to spend the capital on expanding its teams and facilities in the United States and Taiwan, moving its laser products through customer qualification, and building the manufacturing and support capacity that hyperscalers and data center operators demand. The company has already begun early sampling of its chip-level products, making this fundraise less about proving a laboratory result and more about turning that result into a dependable supply chain.

The Optical Bottleneck

Training and running large AI models requires thousands of processors to exchange enormous volumes of data. Adding more accelerators increases theoretical computing power, but it also raises pressure on the network connecting them. If those links cannot deliver enough bandwidth at acceptable latency and energy use, processors spend more time waiting and less time calculating.

PicoJool attacks that constraint with vertical-cavity surface-emitting lasers, better known as VCSELs. These tiny semiconductor lasers emit light vertically from a chip surface and already serve optical communications and sensing markets. PicoJool applies their cost and manufacturing characteristics to the far more demanding links inside scale-up AI systems.

The company's current portfolio includes high-speed 100G and 200G VCSEL products alongside lower-power MicroVCSEL configurations. The architecture treats the optical link as part of compute economics: every reduction in energy per transmitted bit affects power density, cooling requirements, and the number of accelerators a system can use effectively.

From 200G Per Lane to a Full Portfolio

The centerpiece is a 200G-per-lane VCSEL that the company says exceeds 37 GHz of bandwidth. PicoJool is also developing quad-100G, quad-200G, and 32-by-50G NRZ MicroVCSEL configurations, giving system designers several routes toward 1.6-terabit connectivity and, eventually, 3.2-terabit links.

Those options matter because AI data centers are not converging on one optical design. Pluggable transceivers remain common, while active optical cables move more electronics into the cable assembly. Near-packaged optics place optical components closer to the processor, and co-packaged optics integrate them even more tightly. Each architecture trades off reach, serviceability, density, heat, and power consumption differently.

PicoJool intends to address multiple layers of that market, moving from chip-level VCSELs and MicroVCSELs into active optical cable and near-packaged optics modules. Its MicroVCSEL platform uses many parallel optical channels to build aggregate bandwidth while cutting the energy required to move each bit. The company is also targeting 50G NRZ optical compute interconnects and 64G NRZ/PAM4 configurations for PCIe 6.0-based near-packaged optics.

The Series A follows the company's June unveiling of its 200G VCSEL and MicroVCSEL family. According to the funding announcement, customer interest has shifted the focus from performance demonstrations toward qualification and deployment.

The Funding Is Really About Manufacturing

Advanced optical components face a difficult transition between a strong device benchmark and a product that ships at hyperscale volumes. Yield, packaging, thermal behavior, reliability testing, and integration with existing systems can all become limiting factors. That is why a meaningful portion of PicoJool's plan centers on operations and production rather than another headline speed milestone.

The company is working with WIN Semiconductor and other gallium-arsenide foundries to prepare its 200G VCSELs for production. WIN describes itself as the world's first pure-play six-inch GaAs foundry and says its manufacturing services include optoelectronic device fabrication for optical communications. Its broader optoelectronics portfolio covers epitaxial growth, wafer processing, device characterization, and reliability verification.

That manufacturing relationship gives PicoJool access to an established compound-semiconductor production base. It does not remove execution risk. The products still need to meet customer reliability requirements, integrate into emerging optical architectures, and reach attractive economics at volume — precisely the tasks the new capital is intended to fund.

A Crowded Race

PicoJool enters a crowded contest involving laser suppliers, silicon photonics companies, networking vendors, and accelerator manufacturers. The market is moving fast because the cost of inefficient data movement rises with every generation of AI cluster. Faster links alone will not determine the winners; power per bit, packaging complexity, manufacturing yield, and the ability to qualify with hyperscale customers will matter equally.

The involvement of Hudson River Trading adds a strategic investor that operates large-scale computing systems. Playground Global general partner and former Intel CEO Pat Gelsinger remains a prominent backer from the seed round. Socratic Partners' lead investment places the new funding behind a specific transition: from a company demonstrating high-speed VCSEL performance to one attempting to supply a complete optical connectivity platform.

PicoJool's next milestones will be commercial rather than theoretical. Early samples must become qualified products, foundry relationships must support repeatable output, and prospective hyperscaler work must turn into deployments. The Series A gives the company room to cross that gap; whether its VCSEL approach becomes a meaningful part of AI data center architecture depends on how effectively it converts bandwidth claims into reliable, power-efficient links at production scale.