Bardia Pezeshki

The Light Speed Bet

The first thing to know about Bardia Pezeshki: he has spent over two decades making infrastructure nobody else thought was ready yet. At Santur, he shipped tunable lasers into a market that was technically skeptical and economically broken. He got to 80% market share anyway. At Kaiam, he built multi-wavelength transceivers for 40G and 100G datacenter links years before the hyperscalers needed them at scale. And now, at Avicena Tech, he's making the argument that microLEDs - the same technology showing up in next-generation displays - can solve AI's most underappreciated problem: how much energy it takes to move data between chips.

Pezeshki studied physics at Harvey Mudd College, one of the few undergraduate institutions that operates at graduate rigor from day one. He followed that with a PhD in electrical engineering at Stanford, finishing in 1991. His early career took him through IBM's T.J. Watson Research Center and SDL Inc., where the technology of record for long-haul communications was still being worked out. He understood lasers, waveguides, and the physics of getting light to behave like a reliable wire - not as a theoretical matter, but as a product engineering problem. That distinction would define everything he built after.

Santur: The Company That Moved the Internet

In 2000, with co-founder Ed Vail, Pezeshki started Santur Corporation in the wreckage of the dot-com bust. The idea was elegant: rather than building a single tunable laser from complex optics, Santur used an array of DFB lasers - each fixed to a different wavelength - and a MEMS mirror to select between them. It was a hardware hack that produced a product which was cheaper, more stable, and faster to manufacture than competing approaches.

The results were not subtle. By 2007, Santur was the highest-volume supplier of widely tunable lasers on earth. Over 500,000 units shipped. More than $500M in revenue generated over the company's lifetime. For years, most of the world's long-distance internet traffic ran through fiber carrying light generated by Santur's chips. The scale of that is still strange to contemplate: a company of a few hundred engineers built the physical layer of the global internet.

Kaiam: The Alignment Problem

After Santur, Pezeshki founded Kaiam Corporation in 2009 - this time attacking a different optical engineering problem. Optical transceivers require precise physical alignment between lasers, waveguides, lenses, and fibers. The alignment step was expensive, slow, and hard to automate. Kaiam's insight was to use a planar silicon photonics platform to make alignment a lithographic step rather than a mechanical one. The cost came down. The yield went up.

Kaiam shipped the first practical multi-wavelength transceivers at 40Gbps and 100Gbps - products that ended up sole-sourced at major cloud companies and deployed across hyperscale datacenters. The technology worked. The business ran into trouble: a combination of cash-flow constraints and patent litigation forced Kaiam to wind down in early 2019. Pezeshki moved on the same year.

Avicena: The LED Bet on AI

Pezeshki founded Avicena Tech in 2019 with a specific thesis: as AI workloads exploded, the energy cost of moving data between chips would become the binding constraint on GPU cluster performance. Copper interconnects bleed power as a function of distance and bandwidth. Existing optical solutions - laser-based silicon photonics, VCSEL arrays - were cheaper than coherent but still expensive enough and power-hungry enough to leave headroom for something radically different.

The something different was microLEDs. Where a laser requires a resonant cavity and precise temperature control, a microLED is a simpler, stabler device. It switches fast enough for multi-gigabit modulation, consumes far less power per photon, and can be manufactured in arrays of hundreds of elements on a single chip. Avicena's LightBundle technology packages 256 microLED elements with CMOS photodetectors, micro-lens arrays, and multi-core fiber bundles into a transceiver architecture that runs at 4Gbps per lane - delivering over 1 Tbps per fiber bundle.

The 200-femtojoule-per-bit figure demonstrated at ECOC 2025 in Copenhagen is not a theoretical projection. It is a measured result on real hardware. For context: 200 femtojoules is 0.2 picojoules. Silicon photonics transceivers typically run at 1-5 picojoules per bit. Traditional copper SerDes links run at 10-30 picojoules per bit. Avicena's number is in a different order of magnitude.

The Investors Who Read the Physics

Avicena's May 2025 Series B closed at $65M, oversubscribed, with a remarkably strategic cap table. Tiger Global led the round - an unusual move for a semiconductor hardware company whose products don't run software. SK hynix joined as an investor, connecting Avicena to one of the two largest DRAM manufacturers in the world, a company with deep incentives to solve memory disaggregation and die-to-die bandwidth. Lam Research - which makes the deposition and etch equipment that builds semiconductor devices - also invested, as did Hitachi Ventures, Maverick Silicon, Prosperity7, and Cerberus Capital, which had backed earlier rounds.

The TSMC partnership, announced separately in April 2025, confirmed that Avicena's silicon photodetector arrays are being produced at the world's most advanced foundry. That matters because TSMC doesn't take partnerships with companies whose manufacturing roadmap lacks credibility.

The CTO Pivot

In September 2025, Pezeshki stepped back from the CEO role at Avicena, handing the operational reins to Marco Chisari, a former SVP at Samsung Foundries with decades of semiconductor manufacturing experience. Pezeshki moved into the Chief Technology Officer role - which reads less as a demotion than as a return to the work he's been doing anyway. The company was reaching product launch and needed a CEO who had scaled a semiconductor manufacturing operation before. Pezeshki, by his own consistent positioning, is the person who finds the physics first.

The LightBundle evaluation kit launched in March 2026 - the first device of its kind, allowing hardware engineers to integrate microLED optical interconnects into their own systems. The world's first. That the evaluation kit exists at all is its own kind of benchmark.