POET Technologies has secured a $50 million initial order from Lumilens for optical engines aimed at easing one of the toughest constraints in artificial intelligence data centres: moving vast volumes of data faster and with less power than copper wiring can sustain.
The agreement, announced on May 14, 2026, covers optical engines built on POET’s Electrical-Optical Interposer platform and forms part of a broader supply and joint development arrangement that could rise above $500 million in cumulative purchases over five years. The companies are targeting next-generation AI infrastructure, where graphics processors, accelerators and switching systems increasingly require high-bandwidth, low-latency interconnects inside and between data centre racks.
POET, listed on Nasdaq, said Lumilens placed the initial purchase order for EOI-based engines as the first phase of a commercial framework tied to a new class of wafer-level photonic integration. The platform is designed to combine electronic and optical components in a compact architecture, reducing the need for conventional assembly methods that add cost, power draw and alignment complexity.
Lumilens, an optical interconnect company focused on AI networking, is expected to work with POET on products for 800G and 1.6T modules, as well as future near-package and co-packaged optics. Engineering samples are planned for late 2026, with volume production aimed at hyperscale deployments in 2027, subject to successful qualification and manufacturing scale-up.
The deal arrives as AI data centre operators confront rising pressure from bandwidth, energy and thermal limits. Copper links remain widely used for short-reach connections because of their cost and latency advantages, but their efficiency deteriorates as data rates climb and cable distances lengthen. Optical links, which transmit information using light, are increasingly being considered for high-speed connections where copper becomes too power-hungry or physically constrained.
POET’s technology is intended to address that shift by manufacturing optical engines at wafer level, a method closer to semiconductor production than traditional optical module assembly. The company says its interposer approach can reduce alignment steps, improve scalability and support higher-volume production. For AI systems, where thousands of accelerators must exchange data continuously, even modest improvements in power use and bandwidth density can affect operating costs and cluster performance.
As part of the arrangement, POET granted Lumilens a warrant to buy up to 22,921,408 common shares at $8.25 each over nine years. A smaller portion is immediately exercisable, while the balance is tied to future purchase payments by Lumilens. The structure gives Lumilens a financial incentive to expand orders, while offering POET a path to deeper commercial alignment if the technology reaches production scale.
The announcement also changes the tone around POET after a volatile stretch for the company. Late in April, Marvell Semiconductor cancelled purchase orders linked to Celestial AI after acquiring that business, citing alleged confidentiality breaches related to POET’s disclosures. POET’s shares fell sharply after that development, exposing investor concern over customer concentration, disclosure practices and the timetable for converting design wins into durable revenue.
The Lumilens order therefore carries importance beyond its dollar value. It gives POET a fresh commercial anchor at a time when investors are scrutinising whether its optical interposer platform can move from promise to repeatable production. The company has also been strengthening its operating bench, including the appointment of Sandeep Kumar as chief operating officer, a move aimed at improving manufacturing execution as customer programmes advance.
Still, the path is not without risk. Optical interconnect adoption in AI data centres is not uniform. Copper continues to dominate many short-reach applications, and large cloud operators often qualify new networking components slowly because reliability failures can disrupt expensive AI clusters. Competing suppliers in silicon photonics, co-packaged optics, active electrical cables and advanced switch architectures are also pursuing the same bottleneck.
The agreement, announced on May 14, 2026, covers optical engines built on POET’s Electrical-Optical Interposer platform and forms part of a broader supply and joint development arrangement that could rise above $500 million in cumulative purchases over five years. The companies are targeting next-generation AI infrastructure, where graphics processors, accelerators and switching systems increasingly require high-bandwidth, low-latency interconnects inside and between data centre racks.
POET, listed on Nasdaq, said Lumilens placed the initial purchase order for EOI-based engines as the first phase of a commercial framework tied to a new class of wafer-level photonic integration. The platform is designed to combine electronic and optical components in a compact architecture, reducing the need for conventional assembly methods that add cost, power draw and alignment complexity.
Lumilens, an optical interconnect company focused on AI networking, is expected to work with POET on products for 800G and 1.6T modules, as well as future near-package and co-packaged optics. Engineering samples are planned for late 2026, with volume production aimed at hyperscale deployments in 2027, subject to successful qualification and manufacturing scale-up.
The deal arrives as AI data centre operators confront rising pressure from bandwidth, energy and thermal limits. Copper links remain widely used for short-reach connections because of their cost and latency advantages, but their efficiency deteriorates as data rates climb and cable distances lengthen. Optical links, which transmit information using light, are increasingly being considered for high-speed connections where copper becomes too power-hungry or physically constrained.
POET’s technology is intended to address that shift by manufacturing optical engines at wafer level, a method closer to semiconductor production than traditional optical module assembly. The company says its interposer approach can reduce alignment steps, improve scalability and support higher-volume production. For AI systems, where thousands of accelerators must exchange data continuously, even modest improvements in power use and bandwidth density can affect operating costs and cluster performance.
As part of the arrangement, POET granted Lumilens a warrant to buy up to 22,921,408 common shares at $8.25 each over nine years. A smaller portion is immediately exercisable, while the balance is tied to future purchase payments by Lumilens. The structure gives Lumilens a financial incentive to expand orders, while offering POET a path to deeper commercial alignment if the technology reaches production scale.
The announcement also changes the tone around POET after a volatile stretch for the company. Late in April, Marvell Semiconductor cancelled purchase orders linked to Celestial AI after acquiring that business, citing alleged confidentiality breaches related to POET’s disclosures. POET’s shares fell sharply after that development, exposing investor concern over customer concentration, disclosure practices and the timetable for converting design wins into durable revenue.
The Lumilens order therefore carries importance beyond its dollar value. It gives POET a fresh commercial anchor at a time when investors are scrutinising whether its optical interposer platform can move from promise to repeatable production. The company has also been strengthening its operating bench, including the appointment of Sandeep Kumar as chief operating officer, a move aimed at improving manufacturing execution as customer programmes advance.
Still, the path is not without risk. Optical interconnect adoption in AI data centres is not uniform. Copper continues to dominate many short-reach applications, and large cloud operators often qualify new networking components slowly because reliability failures can disrupt expensive AI clusters. Competing suppliers in silicon photonics, co-packaged optics, active electrical cables and advanced switch architectures are also pursuing the same bottleneck.
Topics
Technology