Photonic AI hardware is now in outside testers' hands

Kopin is moving its headquarters staff out of 125 North Drive in Westborough, Massachusetts, and into 75 North Drive, because it needs the old building's floor space for manufacturing. Optics.org reported the plan in its November 12 business round-up. The expansion could add up to 80 jobs over five years. The state is backing it with an Investment Tax Credit of about $1.46 million under the Economic Development Incentive Program and an $800,000 grant from the Massachusetts Manufacturing Innovation Initiative.

The microLEDs made there will go to two markets. One is color microdisplays for next-generation soldier and pilot systems. The other is Neural I/o, a Kopin interconnect that sends data between AI processors as light so that moving it takes less energy. The second market is the one that ties this week's photonics news together, and it raises a practical question for anyone who buys or runs AI infrastructure: which of these energy claims can be tested now, and which are still only promises?

MicroLEDs move from screens to data links

Until recently, a microLED company was a display company. Photonics Spectra's November editorial says micro-LED became an unexpected buzzword at this fall's European Conference on Optical Communications. It names Avicena, ams OSRAM, Credo, Ennostar and Kopin, working with partner Fabric.AI, among the companies pushing the technology into optical communications. Kopin's move to free up production space puts physical capacity behind that shift, and it comes with state money attached.

There's a reason buyers are paying attention. Sajan Saini, education director at MIT's Initiative for Knowledge and Innovation in Manufacturing, told Optics & Photonics News that some experts estimate bringing optics right up to the chip can cut a data center's power use by 30%, and possibly by as much as 50%. Saini was describing co-packaged optics in general, not microLED links in particular. Still, his estimate shows how large the savings could be for every company trying to move data with light instead of copper.

Kopin CEO Michael Murray described the plan in terms of engineering talent, equipment and domestic production capacity. Talent may be the hardest of those to add. The new roles cover engineering, advanced manufacturing and semiconductor processing. In the same OPN feature, writer Stewart Wills reports that the biggest US shortage is in technicians and other middle-skilled workers. He cites a November 2023 study in Applied Optics that estimated demand for 42,000 new middle-skilled photonics workers by 2030. At current graduation rates, meeting that demand would take about 100 more community college programs. Eighty jobs over five years is a small share of that number. For operations leaders hiring photonics technicians in Massachusetts, though, it means one more employer competing for the same small group of trained people.

Two photonic chips outsiders can now test

The same round-up carried the item that matters most to infrastructure buyers. Q.ANT, a photonic computing developer based in Stuttgart, and IONOS, a cloud provider for small and mid-sized businesses, have opened a pilot that puts Q.ANT's second-generation Native Processing Unit on a bare metal server in a German IONOS data center. Applicants approved through the IONOS website can run their own workloads there, using their own data and software. When Q.ANT released its open-source SDK in September, developers could only build and simulate applications. Now they can run them on a physical processor.

Q.ANT says its processor computes complex nonlinear functions directly in light, so developers can build smaller models that run faster. Founder and CEO Michael Förtsch said the AI conversation is turning from what models can do to what they cost to run. IONOS CEO Achim Weiss said the processor runs under German jurisdiction and that the pilot's results will show which applications benefit most and shape the companies' next development work. That makes the pilot a joint research effort as much as a product launch. For European teams whose data has to stay under EU rules, the hosting location removes one obstacle to testing with real production data instead of samples.

PsiQuantum is offering something similar on the networking side. On November 3, Optics.org reported that the quantum computing company had launched an optical circuit switch built on the silicon photonics platform it has spent nearly a decade developing for quantum computers. The first unit is an 8×8 demonstration switch, and PsiQuantum is sending samples to selected prospective customers.

According to PsiQuantum, all three switch results beat the targets it set for the device. The barium titanate technology behind its larger quantum switch isn't needed for this demonstration unit or for the first products, so the company keeps it in reserve for faster future products. Commercial products are due to be announced in the coming months.

This month, two photonic answers to AI's power problem reached hardware that outside teams can actually test: PsiQuantum's switch samples and Q.ANT's processor pilot. Neither is a production purchase yet. An 8×8 switch is a demonstration-scale part, and Q.ANT's pilot only admits approved applicants. Both offers are mainly chances for buyers to gather evidence. For most infrastructure teams, that evidence will be worth more than any purchase made this quarter.

AI help for scarce optical engineers

Opdo, an AI-native optical engineering startup in Los Angeles, raised $2.3 million in pre-seed funding from investors including Angular Ventures. Co-founder and CEO Stephan Kuehr's pitch is that hardware teams have optics problems waiting in line and too few engineers to work on them. Opdo combines AI agents with its own optical specialists. Its software works inside Zemax and CODE V, the design tools engineers already use. Test data from manufactured parts goes back into its models to improve the next design. Customers can also order prototypes and finished components through Opdo. The company says this approach has cut development and manufacturing cycles from months to weeks on projects that include laser and imaging systems, micro-optics and nanostructured surfaces.

The OPN feature explains why the timing makes sense. Saini says photonics training needs a "fast-reskilling" curriculum rebuilt about every two years. He gives the example of co-packaged optics, where a mid-career fiber-optics engineer has to match terahertz-scale optical signals to gigahertz-scale electrical ones, work most haven't done since college. An EU-funded industry survey run by the Phortify education network found that employers rate graduates' technical training well but want stronger data and digital skills, including experience using AI at work.

There's a fair objection. The months-to-weeks figure comes from Opdo's own customer work, not an outside benchmark, and the company's model still depends on its own engineers. For a product team with a waiting list of optics projects, the realistic offer is a design service that doesn't require changing tools. It doesn't replace the engineers that team would otherwise need to hire.

Indoor light instead of battery swaps

The week's largest raise was for edge devices, not data centers. RadianX, previously called California Perovskite Technology, raised $9 million in combined venture and project financing, co-led by SOSV and Twynam. Pacific Union Technology Precision Manufacturing (PUT) is separately funding a dedicated production line in its facility. CEO Vivian Qu describes that line, RadianX's first automated 6 MW line, as a design the company can repeat at larger scale elsewhere. The modules are as thin as 0.7 mm, come in custom sizes and output voltages, and are designed to be built into the surfaces of sensors, wearables and other connected devices.

Perovskite keeps producing useful power in dim and indoor light, where silicon performs poorly. Chief operating officer Josh Douglas says the aim is to convert light the devices already receive into power, keeping them running longer and reducing or eliminating batteries. That matters most for operators with large fleets of battery-powered sensors in lit indoor spaces, where battery replacement is a recurring maintenance job. For those teams, the question to ask is how much power a module produces in the actual light at the installation site.

Next test: Photonics West in January

SPIE Photonics West runs from January 30 to February 4, 2027, at the Moscone Center in San Francisco. SPIE expects more than 4,500 presentations and more than 1,400 exhibitors, many of whom will launch products during the week. PsiQuantum's Pete Shadbolt is on the list of plenary and Hot Topics speakers. The Embedded Vision Summit, a computer vision and physical AI conference, joins the event for the first time.

Two things will show whether this read holds up. One is whether PsiQuantum's promised commercial switches arrive with specs close to the demonstration unit's. The other is whether Q.ANT and IONOS publish which workloads ran better on the processor. That second result is the most useful one for a data center buyer: a named list of workloads where computing with light paid off.