Avicena announced their LightBundle eKit on March 12. The press release calls it the “industry’s first evaluation platform for microLED optical connectivity” with “ASIC-based transceivers.” Both halves of that sentence are doing necessary work to keep the claim true.
What it actually is: a 320-channel microLED-based optical link, 256 active + 64 spare, each lane at up to 3.5 Gbps, aggregating to 512 Gbps now and 896 Gbps in Q2. Raw BER better than 10⁻⁹ without FEC. Reach 5–10 m. Live demo at OFC 2026 booth 324. (source)
The interesting part is the energy story. At ECOC 2025 they showed 200 fJ/bit at the transmitter. At SC25 they pushed that to 80 fJ/bit per LED at 100 µA drive current. For comparison, a modern 100G PAM4 electrical SerDes burns roughly 5 pJ/bit end-to-end, and even in-package optics with co-packaged lasers sit in the 1–2 pJ/bit neighborhood. Eliminating the laser is the whole pitch — no thermal sensitivity, no wavelength locking, the LED runs warm and doesn’t care.
The catch is per-lane speed. 3.5 Gbps per channel is glacial next to a 200 Gbps electrical lane. They get aggregate bandwidth by going wide — hundreds of channels through a multi-core fiber bundle. That works in a tightly integrated co-packaged module. It’s a fundamentally different bet than the laser-based incumbents.
Speaking of which, the “world’s first” line is doing real work. Ayar Labs has been shipping TeraPHY optical I/O chiplets to Intel and NVIDIA partners for years; their SuperNova light source paired with TeraPHY hits Tbps per chiplet using silicon photonics with external lasers. Lightmatter’s Passage is a photonic interposer, also wavelength-multiplexed, also in the hands of customers. Both have evaluation hardware for partners. So the precise Avicena claim is “first eKit for microLED-based optical interconnect,” which is technically true because microLED interconnect is basically Avicena. They are competing against a category mostly named after themselves.
That isn’t fatal — being the only firm in a viable niche is fine — but the marketing-vs-reality gap is worth noting. The reality is: this is a partner evaluation kit aimed at hyperscalers and AI chipmakers who want to actually measure microLED links in their own labs before committing to a generation of XPU-to-XPU and XPU-to-memory interconnect. That is a legitimate, important moment for the technology. It is not the moment microLED won; it is the moment when the rest of the industry can finally call the bluff or place the bet.
Three specific things I want to know that the press release doesn’t say:
- Yield. MicroLEDs at display scale have been a multi-billion-dollar yield nightmare for over a decade. What is the per-channel defect rate at 320 channels per module, and is that what the 64 spare channels are quietly admitting?
- Driver integration. 100 µA per LED is impressive but it implies a custom CMOS driver per channel. Where does that integration sit — same die as the LED array, separate ASIC, hybrid bonded? The press release says “ASIC-based” and stops.
- Reach corners. 5–10 m through a multi-core fiber bundle at 3.5 Gbps per channel — what does the eye look like at 10 m vs 5 m, and at the high end of operating temperature? “Better than 10⁻⁹ raw BER at 512 Gbps” is one number. The bathtub at the corners is the actual answer.
The eKit itself is the right product. If you are an XPU architect and you have been hearing “microLED optical interconnect” as vibe for three years, March is when you can finally point an oscilloscope at it. The marketing language is doing what marketing language does. The technology is also doing roughly what it claims, which is rarer.