POET Technologies (NASDAQ: POET) and optical interconnect startup Lumilens announced a $50 million supply and joint development agreement on May 14 in San Jose, California, targeting the copper-wire bottleneck that is now slowing AI infrastructure globally — a constraint that affects every hyperscale data center operator planning GPU clusters for 2027 and beyond. Under the deal, the two companies will co-develop wafer-level photonic modules that replace electrical signals with light, promising lower power draw and higher bandwidth density for the AI facilities that train and run the models consumers use daily.
A $50M Opening Order Within a Deal Worth Up to $500M
Lumilens placed an initial $50 million purchase order with POET for optical engines built on an Electrical-Optical Interposer (EOI) platform — a new class of component that integrates photonic and electronic devices at the wafer scale. The agreement establishes a commercial framework that could scale to more than $500 million in cumulative purchases from POET over five years. To align both companies to the long-term upside, POET granted Lumilens a nine-year warrant to purchase up to 22,921,408 common shares at an exercise price of $8.25 per share.
"GPU interconnects are emerging as the defining bottleneck for scaling AI, and addressing it requires rethinking the full optical stack — silicon, photonics, and packaging — together," said Ankur Singla, CEO and founder of Lumilens.
The roadmap calls for engineering samples by late 2026 and a production ramp timed to hyperscaler customer deployments in 2027. Lumilens, founded in 2024 and headquartered in Belmont, California, has raised approximately $138.8 million from investors including Mayfield and Spark Capital.
Copper Hits a Physical Wall at 200 Gb/s Per Lane
The urgency behind the deal reflects a measurable constraint. At 200 Gb/s per lane, the physics of copper become the limiting factor: passive copper cables can no longer reliably span even a single server rack without signal degradation and excessive heat. As AI clusters grow from thousands to hundreds of thousands of GPUs, that limitation compounds into a system-level architecture problem, not just a cable-swap.
Industry tracking data published at OFC 2026 shows interconnect speeds are now moving from 800G toward 1.6T and beyond — speeds at which electrical interconnects become non-linear in their power demands. Hyperscalers are expected to spend over $761 billion on data centers and AI infrastructure in 2026, a 70% year-over-year acceleration, with networking infrastructure absorbing a growing share of that outlay. Optical fiber is now the preferred solution for connections beyond three meters, with the optical interconnect market in AI data centers projected to grow from $3.75 billion in 2025 to $18.36 billion by 2033.
Nvidia, which plans an eightfold increase in the maximum number of GPUs per system — from 72 to 576 — by 2027, has acknowledged that optical components currently consume about 10 percent of a data center's compute budget and that progress on photonics reliability and manufacturing cost is required before that share can fall.
What POET and Lumilens Are Building Together
The joint platform is built around POET's alignment-free wafer-scale manufacturing — an approach designed to eliminate the single largest cost and yield constraint in conventional optical engine production, which requires precise manual fiber alignment. Lumilens contributes silicon photonics chipsets, mixed-signal ICs, and advanced optical manufacturing capabilities.
The roadmap spans the full spectrum of AI interconnect needs: 800G and 1.6T pluggable transceivers for operators upgrading existing infrastructure, Near-Package Optics (NPO) for the next generation of GPU rack designs, and Co-Packaged Optics (CPO) for the longer-term vision of integrating optical engines directly onto switch ASICs.
POET Arrives Carrying Baggage From a Turbulent Month
The Lumilens agreement comes three weeks after one of POET's most disruptive partner separations on record. On April 23, 2026, Marvell Semiconductor — which had acquired Celestial AI — cancelled all outstanding purchase orders it had inherited from that acquisition, citing POET's alleged public disclosure of purchase-order and shipping details in violation of confidentiality obligations. POET's stock fell 46% on April 28, its sharpest single-session drop on record.
Analysts at Simply Wall St noted that the episode reveals how Marvell is willing to sacrifice a supplier relationship to enforce contract discipline — a signal with implications for other photonics vendors seeking hyperscaler business.
Short seller Wolfpack Research also disclosed a short position in POET in April, branding it an "obvious stock promotion" and warning that POET's cash accumulation through share dilution could trigger Passive Foreign Investment Company (PFIC) classification — a designation that would impose punitive IRS tax treatment on U.S. shareholders. POET CFO Thomas Mika dismissed the report as "a big nothing burger" and described short sellers as "maggots," while the company announced plans to redomicile in the United States to eliminate the foreign-corporation classification that underpins the PFIC risk.
Wolfpack told Stocktwits: "A year from now, the market will render its verdict. Management is financially insulated regardless of the outcome." POET did not respond to a request for additional comment by publication time.
Investors should also note that POET reported only $1.1 million in revenue over the last twelve months and remains unprofitable, while its market capitalization reached approximately $2.2 billion in the days following the Lumilens announcement — implying a price-to-sales multiple above 2,000x. The company carries $430 million in cash and is building production capacity in Malaysia. Revenue from the Lumilens order is conditional on the successful development, qualification, and manufacturing scale-up of the jointly designed modules.
A Crowded Field Racing Toward the Same Deadline
POET and Lumilens are not alone. Ayar Labs has integrated its TeraPHY optical engines into chip design services with Global Unichip and is demonstrating near-die latency of 5 to 20 nanoseconds. Broadcom's Tomahawk 6 CPO switch — now in full volume production — has reduced power consumption to below 3.8 picojoules per bit, compared to 12–15 pJ/bit for conventional copper systems. Nvidia's Quantum-X Photonics InfiniBand switch reached commercial availability in early 2026, with Spectrum-X Photonics Ethernet expected in the second half of the year.
The engineering tradeoffs in co-packaged optics remain real. Yole Group's Photonics Packaging 2026 analysis highlights that even minimal misalignment in photonic assembly can affect performance and yield. A failed optical engine in a CPO system may require replacing equipment worth tens of thousands of dollars, unlike a pluggable transceiver that can be swapped in minutes. Network World reports that Cisco demoed a CPO switch in 2023 but has yet to make a formal commercial announcement — a signal that at least one major networking vendor remains cautious about serviceability risk. Broad CPO adoption for scale-up GPU interconnects is not expected until 2028 to 2030 as supply chains and standards mature, according to Wevolver.
The AI Bottleneck Shift Affects Every Cloud Service You Use
The practical stakes extend beyond Wall Street. Data center operators that cannot scale interconnect bandwidth efficiently face a choice: slow the expansion of AI capacity, or increase power consumption per GPU — costs that feed through to cloud pricing and AI service availability. The transition to optical interconnects is a prerequisite for the next generation of AI services, from real-time translation and medical imaging analysis to the AI coding assistants that tens of millions of developers now use daily.
The POET-Lumilens engineering samples expected in late 2026 will provide the first independent signal of whether wafer-scale alignment-free manufacturing can hold up under production conditions — and whether the $500 million commercial framework the two companies have put in place will ever be fully drawn down.
ⓒ 2026 TECHTIMES.com All rights reserved. Do not reproduce without permission.
