San Jose, California, May 14, 2026, 05:04 (PDT)
- Lumilens kicked things off with an initial order—$50 million—for POET optical engines.
- The companies put the potential value of the supply deal at more than $500 million across five years.
- Engineering samples won’t surface before late 2026; production is targeted for 2027.
POET Technologies announced Thursday it has entered into a supply and joint-development deal with Lumilens, landing an initial $50 million order for optical engines meant for AI data-center networks. Over the next five years, the companies say cumulative orders could top $500 million.
Timing is key. As AI clusters shuttle ever-larger volumes between GPUs, memory, and switches, the spotlight has shifted to optical interconnects—hardware that uses light, not electricity, to move data—drawing growing attention from chipmakers and network equipment suppliers.
POET and Lumilens are taking aim at an Electrical-Optical Interposer, or EOI—a wafer-level platform designed to pack electronic and photonic features more closely together. The goal: shrink optical engines down to chip-like form factors, making them simpler to reproduce and manufacture at scale.
Investors wasted no time. POET shares jumped 16% after the news, according to Investing.com. The most recent market data pegged POET at $14.37—a roughly 4.5% increase compared to the previous close—in a trade that crossed at 11:48 UTC.
Lumilens, which counts Mayfield and Spark Capital among its backers, is building optical links aimed at both scale-up and scale-out AI infrastructure. Scale-up interconnects tie together processors within a single, close-knit compute cluster. For larger data-center networks, scale-out links bridge multiple clusters.
Ankur Singla, who founded Lumilens and serves as CEO, flagged GPU interconnects as a real hurdle for AI, saying the challenge spans “silicon, photonics, and packaging.” According to Singla, the collaboration leans on Lumilens’ strengths in chipsets and manufacturing, pairing them with POET’s approach to wafer-level photonic integration. GlobeNewswire
POET Chairman and CEO Suresh Venkatesan described the platform as a way to introduce “semiconductor-style manufacturing discipline” to the optical engine landscape. That’s the main sell here: shift away from labor-intensive optical assembly by pushing more of the workflow toward wafer-scale production. GlobeNewswire
Lumilens picked up a nine-year warrant that lets it buy as many as 22,921,408 POET common shares for $8.25 apiece. Right away, 2,292,140 of those shares are available to exercise. The rest are set to vest in phases, depending on how much Lumilens ends up spending on future purchase orders—capped at $500 million.
The lineup stretches from 800G and 1.6T pluggable transceivers all the way to near-package optics and co-packaged optics. With co-packaged optics, light-based data moves in closer to the switch or processor—designed to trim both power consumption and latency across data centers.
The agreement comes on the heels of a tougher episode in April. On April 27, POET reported that Marvell Semiconductor had canceled every Celestial-related purchase order with the company after Marvell bought Celestial AI, pointing to supposed violations around confidential purchase-order and shipping details. POET, for its part, said it would keep concentrating on other clients, highlighting a different order valued at roughly $5 million.
POET remains dwarfed by the markets it’s targeting. For the fourth quarter of 2025, the company posted just $341,202 in product and non-recurring engineering revenue, offset by a net loss of $42.7 million. It also noted raising $225 million during the quarter and an additional $150 million in January.
Execution is the sticking point here. Revenue from the Lumilens purchase orders hinges on development milestones—successful builds, module qualification, and scaling up manufacturing. The warrant could mean dilution if exercised, but that mostly kicks in if future purchase payments come through.
For now, POET’s got a new customer commitment to point to and its pitch looks sturdier after the Celestial AI stumble. The tougher tests are still ahead: samples expected late 2026, then customer qualification, and finally, seeing if they can actually kick off that production ramp in 2027.
