In early 2024, Apple quietly buried a project it had been chasing for almost
ten years and somewhere between one and three billion dollars. The plan,
inherited from a 2014 acquisition called LuxVue, was to put a microLED
display on the Apple Watch — brighter pixels, longer battery, no organic
material to burn in. After a decade of yield problems, the company shelved
it. Trade press treated the story as a small obituary for a frustrating
technology. Reporters at Display Daily and The Information wrote about
microLEDs the way they write about flying cars.

The obituary was premature. While Cupertino was giving up on microLEDs as
screens, a 30-person company in Sunnyvale was busy turning them into
something else: the optical pipes inside an AI accelerator.

In 9.3 million US patent grants pulled from the USPTO bulk feed,
Avicena Tech holds 21 of the 25 grants that describe using microLED arrays
for chip-to-chip data communication, or 84% of the entire category. The next
closest filer in the niche is IBM, with one. The company has 39 grants in
total since its first issuances landed in early 2022, every one of them about
the same idea: take the same gallium-nitride microLED chips Apple was trying
to use for pixels, drop them onto a CMOS transceiver, and use them to
machine-gun bits between processors.

In May 2025, Avicena raised a $65M Series B led by Tiger Global, with SK
hynix, Lam Research, Hitachi Ventures and Cerberus Capital participating, per
Optics.org. Total raised: $120M. A month earlier, TSMC announced a
strategic partnership to build the photodetector arrays that pair with
Avicena’s transmitter chips. At ECOC 2025 in Copenhagen, Avicena demonstrated
a working link at 200 femtojoules per bit — the lowest publicly reported
energy per bit for any optical chip-to-chip link, by a wide margin.

Apple’s microLED morgue and Avicena’s NVIDIA-adjacent rocket are the same
piece of physics. The story is which problem turned out to be the bigger
market.

The 2010 paper that aged into a roadmap

The intellectual source code is older than Avicena. In 2010 and 2011, a
Strathclyde University group led by Jonathan McKendry published two papers
showing that the same indium-gallium-nitride microLED arrays being prototyped
for displays could be modulated at gigahertz frequencies and used as
visible-light data transmitters. The 2011 paper in IEEE Journal of Lightwave
Technology, “Visible-Light Communications Using a CMOS-Controlled
Micro-Light-Emitting-Diode Array,” reported 16×16 pixel arrays at 370–450
nanometers running 1 Gb/s per pixel with bit-error rates below 1×10⁻¹⁰. The
paper sat there for 14 years collecting 315 citations and going nowhere
commercial. There was no obvious customer for an interconnect that ran from
a chip to a sensor across a few centimeters.

The customer arrived in 2023. NVIDIA’s H100 launch made the bandwidth between
GPUs the line item that decides whether a training run finishes in days or
weeks. Hyperscalers started buying every fast electrical SerDes Broadcom
could build and asking what came next. Co-packaged optics, the practice of
welding the optical-to-electrical conversion onto the switch silicon, became
a 2025 product roadmap at NVIDIA, Broadcom and Marvell. The race was on.

NVIDIA and Broadcom both bet on silicon photonics, the workhorse approach
that uses tunable on-chip lasers in the 1310/1550-nanometer telecom band.
Avicena bet on a different chemistry. GaN microLEDs at 450 nanometers are
small (a few microns square), can sit directly above the CMOS that drives
them, and don’t need single-mode fiber or expensive lasers. They are also,
not incidentally, the exact device LuxVue was trying to manufacture for an
Apple Watch.

What the patents actually say

The thing the patents disclose is not “use microLEDs.” Anyone can write
that. The thing they disclose is a particular geometry. Read US 12,271,046,
issued April 2025: each transceiver site has a photodetector that is
larger than the microLED, with the microLED bonded directly on top of the
photodetector in a concentric arrangement. Light goes out through a
multimode waveguide; light coming back lands on the ring of photodetector
not occluded by the LED above it. The claim specifies the microLED area is
“11% or less” of the photodetector area, sized so the optical loss from the
LED’s footprint is under 0.5 dB. That is a compactness rule, not a marketing
phrase. It tells you why this works: you can pack a transmit/receive pair
into the same footprint a single VCSEL would occupy, then tile the result
into a 2D array.

Read the more recent US 12,332,488 (“MicroLED parallel optical
interconnects”), issued June 2025: a refractive-lens stack focuses an entire
transmitter array into a multicore fiber bundle, where each microLED maps to
a single fiber core. The patent specifies the math: first lens at the focal
length, second lens separated by the sum of the two focal lengths, multicore
fiber on the far side. This is image-relay optics borrowed from microscopy,
applied to integrated photonics. The whole package fits in a square
millimeter.

Read US 12,395,246, the August 2025 packaging patent: the transceiver IC
hangs over a hole cut through the substrate, with the optical engine pointing
down through the hole into a waveguide stuck inside the substrate itself.
You don’t route fiber on top of your AI accelerator. You route it underneath.

These three patents are the same invention seen from three angles: a tiny
optical engine that fits where a copper pad used to be. Mentally delete the
phrase “microLED” and the patents still describe a coherent piece of
engineering: a stacked transmitter/detector pair, an array, a fiber
coupling, a substrate routing trick. That is the difference between a real
technology and a keyword cluster.

Why this matters now

The bandwidth between GPUs has become the gating constraint on AI training
costs, and the constraint compounds. Inside a server, an NVIDIA NVLink
domain currently tops out at 72 GPUs because copper interconnect cannot push
much further. Beyond that domain, traffic falls onto InfiniBand or Ethernet
fabrics that consume tens of kilowatts per rack just for the optical
transceivers. The bet on co-packaged optics is that you can shrink that
power bill by an order of magnitude and double the bandwidth at the same
time. NVIDIA’s March 2025 Spectrum-X Photonics announcement claimed 3.5x
power reduction; Broadcom’s Tomahawk 6-Davisson is sampling at 102.4 Tbps
of optical switching capacity.

If Avicena’s 200 fJ/bit number holds up in volume manufacturing, that puts
its energy budget below the silicon photonics roadmap by a factor of five.
At hyperscaler scale, with millions of links per data center, a 5x energy
reduction at the optical layer is the difference between a new AI campus
needing one nuclear reactor or two. SK hynix’s investment is not portfolio
filler; SK hynix sells the HBM stacks that NVIDIA glues to its GPUs, and
faster optical I/O lets it sell more of them.

The adjacent that wasn’t a watch

There is a tidy lesson here for anyone scanning patent landscapes for early
signals. Avicena’s founder, Bardia Pezeshki, is not a display engineer. He
is a Stanford EE PhD who started at IBM Watson, founded a tunable-laser
company called Santur in 2000 (acquired by Neophotonics in 2011), then ran
the optical-transceiver startup Kaiam from 2009 to 2019. When he founded
Avicena in 2019, he wasn’t inheriting Apple’s failed display project. He
was looking at the same wafer fabs Apple had been pouring billions into and
asking: what else can a GaN microLED be?

The answer, eight years and 39 patents later, is “the cheapest dense optical
I/O ever built.” The display industry was the wrong adjacent. The data
communication industry was the right one. The pattern is older than
photonics: companies and academics often get the physics right years before
they get the application right, and the eventual buyer is rarely the buyer
the inventors first had in mind. In 2014 Apple bought LuxVue thinking about
your wrist. In 2025 SK hynix bought a piece of Avicena thinking about your
GPU. Same atoms, different building.

The microLED isn’t dead. It just changed careers.

Method note. Patent counts come from the full USPTO bulk grant feed
(9.3M US utility patent grants), filtered to the chip-to-chip optical
interconnect niche by joining the Avicena assignee variants and the broader
microLED-data-communication keyword set, then deduping by document ID.
Literature counts come from OpenAlex (357M scientific works). Funding,
partnerships and product specifications come from Optics.org, SPIE,
BusinessWire, Microled-Info, NVIDIA developer blog, IEEE Spectrum,
and Avicena’s own press releases between 2022 and ECOC 2025. Apple
microLED program history and cancellation timing come from TechCrunch,
Display Daily, Yole Group, and AppleInsider. Numbers in this post
were re-verified against the source databases on 6 May 2026; the niche
contains 25 distinct US grants in total, 21 of them assigned to one of four
spelling variants of Avicena Tech Corp.