POET Technologies has a real commercial opening in AI optical networking, but the Lumilens deal still has to survive development, qualification and scale.
POET Technologies did not just announce another AI partnership. It announced the kind of purchase order that can change how investors see a small photonics company, especially one that reported only $503,389 of first-quarter revenue the same day.
The company and Lumilens said on May 14 that they entered a supply and joint-development agreement for Electrical-Optical Interposer-based optical engines used in AI networks. Lumilens placed an initial $50 million purchase order, and POET said the broader supplier relationship could exceed $500 million in cumulative purchases over five years if the program develops as planned.
That last part matters. The market likes the headline number, and Investing.com reported that POET shares rose 16% after the announcement. But the story is not simply that POET found a buyer. The more useful question is whether this is the moment its optical interposer technology starts moving from promise to commercial adoption.
AI infrastructure has been a story about GPUs for years, but the next constraint is increasingly the network around them. Training and running large models requires huge numbers of chips to move data quickly, reliably and with tolerable power use. If the interconnect cannot keep up, adding more compute becomes less efficient.
That is where POET wants to fit. Its Optical Interposer platform is designed to integrate photonic and electronic devices at chip scale using semiconductor manufacturing techniques. In this deal, POET and Lumilens are building around an Electrical-Optical Interposer, or EOI, which the companies say can support wafer-level optical engine production and reduce reliance on active alignment, one of the cost and throughput pressure points in optical module manufacturing.
The roadmap is ambitious. POET and Lumilens said they plan to start with 800G and 1.6T pluggable transceivers, then move toward near-package optics and co-packaged optics. Engineering samples from the joint development program are expected in late 2026, with production ramp aligned to hyperscaler deployments in 2027.
For founders and operators, this is the part worth watching. AI infrastructure is creating new openings for component companies that can remove practical bottlenecks, not just produce clever lab results. If POET can help optical engines scale more like semiconductors and less like painstakingly assembled specialty hardware, the commercial prize is obvious.
The deal also gives Lumilens a powerful incentive
The structure is not clean enough to treat as ordinary revenue visibility. POET granted Lumilens a warrant to buy up to 22,921,408 common shares at $8.25 per share. The warrant is immediately exercisable for 2,292,140 shares, while the rest vests in tranches based on cumulative payments by Lumilens toward future purchase orders totaling up to $500 million. It runs for nine years.
That is not a small side note. It means Lumilens is not only a customer and development partner, but also a potential equity holder with upside if POET's share price stays above the exercise price. The arrangement may help align both parties around the long-term rollout. It may also leave existing shareholders asking how much value is being shared away to secure the order.
This is common enough in strategic commercial relationships that it should not be dismissed automatically. Smaller hardware companies often need to give partners a reason to commit engineering time, qualification resources and market access. But it changes the investor reading of the deal. A $50 million order attached to warrants is still meaningful, but it is not the same as a simple cash purchase from a customer with no other incentive.
Lumilens itself is another part of the diligence. POET describes the company as an emerging leader in scale-up and scale-out optical interconnects for AI workloads, designing near-package optics, co-packaged optics and pluggable optical transceivers for hyperscale AI data centers. That sounds relevant, but it also means the market will need evidence of manufacturing depth, customer access and execution capacity as the program moves through sampling and qualification.
Commercialization is still the hard part
POET's own first-quarter results show why the announcement landed so strongly. The company reported NRE and product revenue of $503,389 for the quarter ended March 31, up from $166,760 a year earlier, but still tiny compared with the scale implied by the Lumilens framework. It also reported a net loss of $12.3 million, or 8 cents per share.
That gap between current revenue and possible future orders is exactly why the stock moved, and also why the risk remains high. The purchase orders and associated revenue depend on successful module development, qualification and manufacturing scale-up. Those are not administrative steps. In AI data centers, optical components have to meet demanding standards for performance, yield, power, reliability and supply consistency before hyperscalers build around them.
POET has other validation points, including partnerships with LITEON and Lessengers, and Dr. Suresh Venkatesan has been positioning the company around AI and hyperscale data-center demand. Lumilens founder and CEO Ankur Singla is also framing the problem around the full optical stack, including silicon, photonics and packaging. The message is clear: the companies want to attack the interconnect problem before it becomes an even more expensive constraint for AI deployments.
The next proof will not come from the stock chart. It will come from engineering samples in late 2026, customer qualification work and evidence that POET and Lumilens can turn a development roadmap into repeatable production. If that happens, the $500 million figure starts to look like a commercial pipeline. If it does not, this will be remembered as another AI infrastructure announcement that reached the market before the factory floor.
Also read: llama.cpp gives RDNA3 users a sharper local AI path • Data center builders now have a backyard problem • ChatGPT Has Made College Grades A Weaker Hiring Signal
