MicroLEDs Get A Connector

Three hundred and thirty-five light-emitting diodes, each carrying up to 3 Gbps, now terminate in a ferrule an engineer can unplug with two fingers. Avicena showed the arrangement at ECOC 2026 in Málaga between 21 and 23 September and called it the world's first connectorized microLED optical interconnect. The claim is narrow, and it is the right claim. The company launched its evaluation kit in March and began shipping one-terabit kits in August. What it added this month was a connector — and in optical interconnect, the connector is usually where the technology dies.

The product is the LightBundle eKit. Avicena announced on 18 August 2026 that it had begun shipping a version with 335 independent microLED channels feeding a 335-element integrated photodiode array through a multicore fibre bundle, for up to one terabit per second of aggregate raw throughput. The March 2026 version of the kit ran 320 channels at up to 3.5 Gbps each — 256 active, 64 spare — for up to 896 Gbps, with a raw bit error rate better than 10⁻⁹ at 512 Gbps with no forward error correction at all. That kit offered five-metre and ten-metre optical fibre connectivity. The BER figure is the one worth holding onto: raw performance below 10⁻⁹ without FEC means the link has margin to spend, and a connector is a thing that spends margin.

The September announcement adds an industry-standard MPO form factor connector built around a ferrule optimised for multicore fibre bundles. Avicena's own framing of why that matters is operational rather than optical: a detachable optical interface, the company says, can simplify system assembly, allow electronic and optical subsystems to be tested independently, improve field serviceability, and support flexible routing between packages, boards, trays and racks. None of that is a performance claim. All of it is what a buyer asks about before a part goes into a rack.

The other launches at ECOC this year were laser-based. Lumentum introduced an external laser module delivering eight simultaneous DWDM wavelengths in the 1311 nm and 1331 nm bands at up to 24 dBm — roughly 250 mW — per wavelength at the fibre, from a single pluggable module drawing roughly 12 W, aimed squarely at co-packaged optics. Coherent presented PhotonLink, an integrated optics platform for CPO, near-package optics and chip-to-chip connectivity. The Lumentum approach assumes what the whole CPO roadmap assumes: that the light source is a laser sitting outside the package, and that the hard problems are wavelength accuracy, thermal management and where the fibre lands.

Avicena's bet is that for the shortest optical links — inside a rack, between a switch ASIC and the boards around it, under ten metres — a laser is the wrong component. A laser needs threshold current, temperature control and wavelength stability. A microLED needs none of those things. It turns on, it turns off, it does not much care what temperature it is, and it is made by the million in the GaN micro-emitter ecosystem built for displays, which is where Avicena said it was drawing its high-speed emitters as far back as the platform's unveiling in June 2021. The penalty is speed: a single microLED tops out in the low gigabits per second, roughly two orders of magnitude below what a modulated laser can do. Avicena's answer is to stop treating channel rate as the figure of merit and go wide instead — hundreds of slow channels in parallel, each one dumb, none of them tuned.

The system-level numbers follow from that choice. Avicena has claimed shoreline density above 1 Tbps per millimetre at sub-pJ/bit over more than ten metres since March 2025, when the modular platform was announced.

Note the date on that one. Pezeshki has since handed the chief executive's job to Chisari, who was named in September 2025. The engineering claim has not changed; the company around it has.

There is a second consequence of the LED choice that the company does not spell out but that follows from its own description of the platform. Avicena pairs its microLED transmitters with silicon photodetectors and purpose-built silicon transceivers. Silicon stops absorbing beyond about 1.1 µm, so the emission must sit at a wavelength silicon can actually detect — the visible and very near infrared, not the 1310 and 1550 nm telecom bands. That takes germanium and indium phosphide out of the receive path entirely and puts detection back into standard CMOS. It also makes the fibre a short-reach component rather than a long-haul one, which is exactly what the application wants. Avicena has never published a wavelength in nanometres, and it is worth not guessing one.

A detachable interface sounds like packaging. It is not. Every optical connection that comes apart and goes back together costs insertion loss, and a multicore fibre connector has to hold angular alignment as well as lateral alignment, because a rotational error puts each core onto the wrong detector rather than merely dimming it. On a 335-channel bundle there is no useful notion of a partly working link: either the map from emitter array to core to detector array holds, or a block of channels goes dark.

That is why the outside analyst quoted in Avicena's own release is talking about manufacturing rather than physics.

Choosing the MPO form factor is a deliberate piece of conservatism. MPO is what data centres already use, already clean, already test and already stock. A new optical technology that also demands a new connector standard asks a buyer to take two risks at once, and buyers of infrastructure do not do that.

The eKit is an evaluation platform, not a shipping interconnect. Avicena has published no insertion loss, return loss or mating-cycle durability figures for the new ferrule, and those three numbers will decide whether the connectorized version keeps the raw BER that the earlier kit demonstrated at 512 Gbps. Nor is there a published figure for energy per bit measured through the connector rather than through a fixed fibre.

The supply chain, at least, is further along than the product. Avicena has worked with TSMC on photodetector arrays since April 2025 and with ams OSRAM on high-volume production of chip-to-chip interconnects since March 2023, and its Series B in May 2025 was led by Tiger Global with SK hynix participating — a memory maker investing in the wiring between memory and compute.

For two years microLED interconnect has been an interesting curve on a conference slide: lower energy per bit than anything else, at rates nobody in a data centre cared about, over distances nobody was routing optically. The wide-and-dumb architecture answered the rate objection. Shipping evaluation kits answered the "does it work outside your lab" objection. A standard connector answers the last one that mattered to the people who actually buy racks.

What it does not answer is whether the co-packaged optics roadmap — external lasers, DWDM, fibre attached at the package — leaves any room for a second physical layer underneath it. Avicena is not competing with Lumentum's 24 dBm laser module for the same socket. It is arguing that below ten metres there should not be a laser in the socket at all. ECOC 2026 was the first year that argument arrived with a connector you can unplug.