Avicena Tech
sees the light.

The hum in the rack.

Walk into any AI training cluster in 2026 and you'll hear it - the constant whine of fans pulling heat off copper. Most of that heat isn't from doing math. It's from moving data between chips that need to talk faster than electrons over thin metal can carry. Avicena Tech builds the thing meant to replace those wires: thousands of microscopic LEDs, smaller than red blood cells, blinking data into fiber bundles at a trillion bits per second.

The company calls the platform LightBundle. The pitch fits on a napkin - same kind of microLED engineers have spent a decade trying to cram into watches and AR glasses, now repurposed as a chip-to-chip interconnect. It's the kind of idea that sounds obvious in hindsight and impossible before someone ships it.

Copper is running out of room.

For thirty years, the inside of a computer has been a small miracle of conducting metal. Then GPUs got greedy. A modern training cluster wants every accelerator to whisper to every other accelerator, constantly, in step. Copper carries the data fine over a few centimeters - it just costs you watts, real estate, and signal integrity at every millimeter past that. Stretch it across a rack and the bill comes due.

Conventional silicon photonics promised to fix this, and to its credit, it has fixed some of it - the long stuff, between racks. But silicon lasers are temperamental, hot, expensive, and they don't really enjoy being miniaturized to the densities AI chips now demand. They're the wrong tool for short-reach, ultra-dense links. Which, inconveniently, is exactly where the next generation of AI bottlenecks lives.

Avicena saw this gap and walked through it.

Two photonics lifers and a strange idea.

Bardia Pezeshki spent years inside the optical networking industry - including a stint building optical components used inside Google's data centers - watching engineers reach for silicon photonics every time the question came up. In 2019, he and co-founder Rob Kalman made a different bet: that the right material wasn't silicon at all. It was gallium nitride. The same junk that lights up the LED in your phone's flashlight.

GaN microLEDs are small, cheap to fabricate in dense arrays, run cool, and tolerate temperature swings that would cook a silicon laser. They are also, charmingly, off-the-shelf-ish - the display industry has been pouring money into this exact material for years for entirely different reasons. Avicena's wager was that you could co-opt that supply chain, point the LEDs at fiber instead of human eyeballs, and end up with a denser, cooler, far cheaper interconnect than any laser-based competitor.

Optical engineers used to laugh at this. They mostly aren't laughing now.

LightBundle, in plain English.

Picture a chip with a 256-pixel microLED array glued to one edge. A matching array of CMOS photodetectors sits on the other end. Between them, a multi-core fiber bundle - cheap, passive, off-the-shelf. The transceiver ASIC sits in 16nm finFET CMOS, which means it scales the way everything else in a data center scales: along with TSMC's roadmap.

If you can read this caption, your eyes are doing what an Avicena microLED does, roughly a trillion times slower.

If watts were honey, AI data centers would be a bear problem. Avicena is mostly selling smaller bears.

People who write the checks.

Tiger Global led the Series B in May 2025. SK hynix joined - which matters more than the dollar amount, because SK hynix is the memory company that will need this interconnect for the disaggregated memory architectures it has been quietly drawing on whiteboards for years. The earlier rounds pulled in Lam Capital, Micron Ventures, Cerberus and Prosperity7. The cap table reads like an industrial conspiracy: the people who make the chips, the people who make the wafer tools, the people who make the memory.

Then there's TSMC. In April 2025, Avicena announced TSMC would build the CMOS photodetector arrays. TSMC does not normally announce things. When it does, it's because someone inside the building decided this was worth a signal to the rest of the industry. IEEE Spectrum called it "an unorthodox bet" - which, in semiconductor reporting, is roughly as close to a rave review as one gets.

Make the wire boring.

If you ask Bardia Pezeshki what success looks like, the answer is not the kind of thing that lands well on a venture deck. It's: nobody talks about the interconnect anymore. The wire just works. AI engineers stop thinking about bandwidth tax the same way you stopped thinking about clock speed somewhere around 2010.

That's the entire game. Photonics has historically failed not because it didn't work, but because it worked expensively. Avicena's microLEDs are betting that "expensively" is a problem you can fix with the right material and the right manufacturing partner. Cheap fiber. Cheap LEDs. CMOS at TSMC. None of the components want to be exotic.

The disaggregation argument.

The reason SK hynix is on the cap table isn't politeness. It's that the next big architectural shift in AI infrastructure is disaggregated memory - separating the GPU from its HBM stacks by literal meters, then pooling memory across many GPUs. That shift only works if you can move data across those meters as if it were happening on-chip. Copper can't. Silicon photonics gets you part of the way. MicroLEDs, at the densities and power numbers Avicena is now demonstrating, get you all the way.

If that future arrives - and a lot of very serious people are now building hardware that assumes it will - the wire stops being a copper trace at all. It becomes a bundle of light, terminated by something that started its life trying to be a screen.

Walk back into the data center.

It still hums. Of course it does. But picture this: the fans pull less heat. The cables are thinner. The racks stretch farther apart, the floor space rearranges itself around different physical constraints. Somewhere on a board you'd need a microscope to find, a 256-pixel microLED array is blinking faster than human nerves can register, lighting a multi-core fiber that runs to the next rack over.

Avicena didn't build the AI. It built the part underneath the AI that everyone forgets about until it stops scaling. Which is, on balance, exactly the kind of company that gets boring and rich at the same time. If they're right about the wire, you'll never hear of them again. That's the point.