A year ago, only a few microLED companies were looking at the optical interconnect opportunity. Today, nearly all have some activity in this area. While AR glasses remain a hot area, cost and scale remain an issue. Automotive has some clear differentiated applications, but volume production is 4-5 years away. The optical communication need is driven by the boom in data centers as these expand rapidly to support AI. The need here is real and huge, but can the industry develop, test and scale solutions in a short time frame? This is the question a panel of experts took on at SID Display Week 2026.

Co-moderated by Sri Peruvemba (CEO, Marketer International, SID volunteer) and by Nag Patibandla (VP, Office of the CTO, Applied Materials), the panel brought together practitioners with direct commercial and research stakes in making this vision real:

• Nigel Alvarez, VP Product & Marketing, Avicena Tech
• Xi Wang, SVP and GM, Connectivity Business Group, Marvell
• John Kymissis, Professor, Columbia University; past president, SID; co-founder, Lumiode

Setting the Stage: Why Data Center Interconnect Is a Display Industry Problem

The question driving the session is one that would have seemed strange five years ago: why is a display technology conference hosting a panel on data center networking? The answer, as the panel made clear, is that the AI infrastructure buildout has created an interconnect problem that laser-based optical solutions cannot solve at the required scale, and that microLED technology — with its roots in display manufacturing — may be the most viable path forward.

Xi Wang of Marvell opened with a historical framing. Thirty years ago, the optical communications challenge was transcontinental and transoceanic — thousands of kilometers, requiring high-performance Continuous Wave (CW), Distributed Feedback (DFB) lasers with Dense Wavelength Division Multiplexing (DWDM). Data storage was local, and carriers handled long-distance movement. About fifteen years ago, the data center era began: office-size server clusters grew to building-scale facilities, the interconnect problem compressed to hundreds of meters, and the industry adapted laser technology from telecom — cheaper, higher volume, but still laser-based — to link hundreds of thousands of servers.

AI has now pushed the problem inside the rack and between co-packaged chips. The distance is sub-30 meters in many cases, and in the most demanding applications — chip-to-chip, chip-to-High Bandwidth Memory (HBM) — the relevant dimension is millimeters to centimeters. The performance requirements are entirely different from what any telecom-derived architecture was designed to deliver, and the power and density constraints are severe.

The core tension: copper, the incumbent short-distance interconnect technology, is running out of road. Driving a 200 Gbps Serializer/Deserializer (SerDes) over copper is already pushing thermal and signal integrity limits, and the industry is hitting a wall trying to reach 400 Gbps. Every increase in per-lane speed on copper requires faster SerDes electronics that consume 3–6 picojoules per bit (pJ/b). And roughly half of a laser-based optical link’s total power budget goes to the drive electronics, not the light source. In data centers, as Kymissis noted, power is paid for twice: once to supply it and again to cool the heat. Power Usage Effectiveness (PUE) is a metric that measures how efficiently a data center uses energy — specifically, how much of the total power actually reaches the IT equipment vs. being lost to overhead. Using copper at these high speeds makes the true system cost – the PUE –  quite substantial.

Why MicroLED? The Technical and Supply Chain Case

Kymissis provided the foundational physics comparison between LEDs and lasers for this application. The two devices are manufactured similarly, and both produce light via band-gap recombination, but LEDs have no lasing threshold. Semiconductor lasers require a current drive above threshold — and in VCSELs that threshold represents a meaningful minimum power floor regardless of how much optical power the application actually needs. For a link that only needs to traverse centimeters or meters, this minimum intensity represents wasted energy. LEDs can be operated at much lower current densities, require simpler driver circuitry, and are more temperature-tolerant — all of which translates to lower system power and simpler cooling.

Critically, the drive simplicity and the lack of a threshold enable dense integration. Because you do not need precision threshold management per element, you can pack hundreds or thousands of individual microLED emitters into a 2D array and drive them in parallel. This is the architectural shift that makes microLED interesting for data center interconnect — not raw per-channel speed, but massive spatial parallelism.

Alvarez reinforced the temperature tolerance argument. GaN-based microLEDs are inherently more robust in the thermal environment of a data center than VCSELs, and this translates directly to higher reliability — which, as he noted, is the number one qualification criterion for any technology that hyperscalers will bet infrastructure on.

The supply chain argument is equally compelling. The entire global laser market ships approximately 100–200 million units per year — a figure Kymissis noted can be matched on a single wafer of microLEDs. If copper replacement requires billions of light sources annually, as Wang framed it, the laser supply chain cannot meet that demand without massive and risky capital investment.

MicroLED production on GaN wafer infrastructure scales in a way that laser production does not. Alvarez added that the laser market is currently sold out — Coherent and II-VI earnings confirm the supply constraint is real today — making microLED’s potential supply chain advantage immediately relevant, not hypothetically future.

The Architecture: Parallel and Wide, Not Fast and Narrow

Understanding why microLED is architecturally interesting for data center interconnect requires understanding the fundamental tradeoff between fast-and-narrow and slow-and-wide interconnect approaches.

Today’s dominant architecture inside switches and any processing units (XPUs) runs at 1–2 Gbps internally, then serializes that data to drive out over a single high-speed lane. This is the gearbox — the SerDes — and it is where most of the power goes. The laser approach requires it: you have a single emitter per channel, so you have to pack all the data onto one fast lane.

MicroLED arrays flip the architecture. Instead of one emitter at 260 Gbps (Wang’s reference point for where laser drive electronics currently operate), you have 1,000 or more emitters each running at a modest few gigabits per second. The data never needs to be serialized, because the parallel array provides the bandwidth density natively. This eliminates the SerDes entirely in the ideal architecture — the data moves from the processor at its native processing speed, in parallel, directly to a parallel optical output. No gearbox, no serialization overhead, no power burning to create a high-speed optical signal that immediately gets converted back to low-speed electrical.

Wang summarized the vision: terabit-per-fiber aggregate bandwidth, with cost approaching $0.01–0.02 per Gbps to be competitive with copper, achieved through massive 2D arrays rather than high per-lane speed. The bandwidth density target is 10 to 100 times that of copper, and the 2D array architecture enables going aerial — volumetric rather than planar — which is where the paradigm shift really happens.

Alvarez put specific numbers to the density target: from roughly 1–2 Tbps per millimeter in today’s 1D copper architectures, the target for microLED is over 25 Tbps/mm in 1D, with 2D array configurations going substantially beyond that.

Key Specifications: What Does “In the Game” Require?

Peruvemba pushed the panel for concrete performance targets. The numbers that emerged, drawing on an Applied Materials internal workshop:

• Modulation frequency: ≥5 GHz per LED (Avicena Tech’s current product operates at 3.5 GHz; the roadmap target is ~5 GHz)
• Data rate per LED: ≥10 Gbps at NRZ encoding (NRZ is a binary line encoding scheme where the signal holds its voltage level for the entire bit period without returning to zero between bits)
• Array size: minimum ~1,000 LEDs per bundle
• EQE: ≥10% for 5–10 µm device size LEDs
• Aggregate bandwidth: targeting ≥1 Tbps per link (Avicena Tech has demonstrated this with a 400-LED array)
• Bandwidth density: >25 Tbps/mm in 1D array
• Cost target: $0.01–0.02 per Gbps (comparable to copper)
• Device dimensions: single-digit micron (5–10 µm) pixel pitch to achieve the required array density

On wavelength, Avicena Tech currently operates at approximately 430 nm (blue visible). Patibandla noted that silicon photodetectors respond well across visible wavelengths, which is a practical advantage — no need to develop new detector materials. The industry has not converged on a single wavelength, and Alvarez acknowledged that different epi recipes will produce different wavelengths with different efficiency and power tradeoffs. This is an area of active optimization rather than settled practice.

The question of drive electronics was addressed by Wang in the context of the AR display community’s question about “refresh rates”: this is not a display refresh problem but a modulation bandwidth problem. Today’s laser drivers operate at 260 GHz for a single-lane link. For microLED, the per-element drive frequency is far lower — 2, 4, 8 GHz — but the challenge shifts to building scalable, high-density driver ICs and transimpedance amplifiers (TIAs) that can address a dense 2D array. As Wang framed it, this is the classic fast-and-narrow versus slow-and-wide tradeoff in electronics design.

Why Not OLED?

The question was raised from the floor and addressed directly by Kymissis. OLEDs have been explored, but the fundamental physics works against them for this application. Switching speed is the primary constraint: OLED devices have been operated at many megahertz, but reaching double-digit gigahertz — essential for data rate targets — is, as he stated flatly, “unfortunately unrealistic.” The recombination rate in organic materials cannot match that of GaN-based inorganic LEDs, and the tricks available in LED design to push switching speed are not available in the organic system. Temperature tolerance is a secondary issue: OLED lifetime is sensitive to operating temperature, and the thermal environment of a data center works against OLED longevity in a way it does not for GaN.

Where MicroLED Stands Today: The Avicena Tech Benchmark

Alvarez was asked directly about Avicena Tech’s demonstrated performance — the “Avicena Tech spec” that the industry references. The company has demonstrated individual LEDs operating at 16 Gbps, and its evaluation kit (the first such kit commercially available for customer qualification) delivers 1 Tbps aggregate using an array of 400 microLEDs, each running at 3.5 Gbps. A commercial product is targeted for end of 2026.

These are genuine industry firsts, and Alvarez was clear that reaching them was not straightforward: taking individual LEDs, speeding them up to multi-gigahertz operation, integrating them into fiber-coupled bundles, and achieving the reliability required for data center qualification is a non-trivial systems engineering problem. The evaluation kit exists precisely to give hyperscalers and XPU vendors the hardware needed to begin their own qualification processes before committing to the architecture.

Kymissis added a useful data point on the maturity level at the box-to-box and board-to-board interconnect layer: Corning’s high-density cable bundle, demonstrated at Touch Taiwan, represents a viable solution at that level, with competitors also approaching market readiness. The harder — and more architecturally important — problem is the short-reach, tightly integrated chip-to-chip and chip-to-memory layer, where Kymissis argued the solution will likely need to be done in-situ at the fab level rather than through hybridization of off-the-shelf components. This is the intersection where semiconductor packaging and display manufacturing expertise converge, and where the opportunity for Applied Materials and similar infrastructure players is largest.

The Path to Market: Two Incumbents, One New Ecosystem

Wang articulated the competitive landscape with precision: microLED is not challenging one incumbent but two simultaneously. On one side is copper, which is cheap, well-understood, and still being actively improved (Peruvemba noted he had recently spoken with an engineer in Silicon Valley working specifically on pushing copper further). On the other is the mature laser-based optical ecosystem — thirty years of development in fiber types, connectors, qualification standards, testing methodologies, and supply chain — none of which directly translates to microLED-based array interconnect.

The first commercial deployment will be as an active optical cable (AOC), replacing a direct-attach copper (DAC) cable or active electrical cable in an existing SerDes architecture via a gearbox. Alvarez acknowledged this is a transitional, “kludgy” use of the technology — the full power efficiency benefit requires re-architecting the chip to eliminate the SerDes — but it is the correct first step. It is the use case that operators can validate without infrastructure commitment, building the reliability and scalability track record that enables deeper architectural integration in subsequent generations.

Wang’s concern about ecosystem readiness was candid: we can manufacture billions of microLEDs, but can we manufacture billions of the fiber arrays and connectors specific to microLED-based links? The fiber-optic ecosystem took twenty years to mature for laser-based communication. Building the equivalent for microLED — including new fiber array form factors, connector standards, and assembly processes — is not a device-level problem but a supply chain and standards problem, and it has barely started.

Wang proposed an intriguing cross-industry knowledge transfer: the display industry’s investment in pixel-level yield, camera-based testing, and array-level quality control could be the basis for microLED datacom qualification, rather than adapting point-to-point telecom test methods that were never designed for array-based incoherent sources.

The Fiber Coupling Problem: Patibandla’s Shower Head

Patibandla closed with what he called a “homework question” for the audience — an unsolved problem that he framed with a memorable analogy. Connecting a single laser to a fiber is like connecting a garden hose to a tap. Connecting 1,000 microLEDs to a fiber bundle is like connecting 1,000 small pipes to each hole in a shower head.

The coupling challenge is real: how do you efficiently couple light from a dense 2D microLED array into a corresponding dense fiber or waveguide array, with low loss and high positional precision at scale?

An audience member — Mike Lee of Vuzix — offered a direct response: don’t use a fiber bundle. Instead, align each individual microLED pixel with a single waveguide channel, creating a one-to-one pixel-to-photonic-channel mapping. This eliminates any sharing of optical paths between emitters, which Patibandla confirmed is essential — the moment you introduce sharing, you reintroduce the multiplexing (Certes-type) problem that the parallel architecture is designed to eliminate. The exchange illustrated how much fundamental systems engineering remains to be worked out, even as the device performance advances.

Who Leads? Hyperscalers, and the Question Is When Not Who

Patibandla asked the panel — somewhat self-interestedly, as he acknowledged — who in the value chain would pull this ecosystem. The consensus was clear: hyperscalers.

Wang’s evidence: at OFC 2026, held in the same building just two months prior, every hyperscaler present named IO density as their central interconnect concern, and nearly all specifically mentioned microLED as a candidate for the next technology step. Hyperscalers control their supply chains, move faster than enterprise or telecom ecosystems, and have both the capital and the urgency to pull new technologies to market. Wang’s framing: it is not a question of who will lead, but when the transition happens.

Alvarez described the challenge as a full-stack problem that runs from AI workload and model architecture down through infrastructure, chips, and power — the “five-layer cake” approach. No single company drives it; each hyperscaler has its own architecture, and XPU vendors have distinct requirements. Marvell’s approach, as Wang explained, is technology-agnostic: with several billion dollars invested across the space (including acquisitions and microLED-specific company investments), the strategy is to identify the best technology for each specific application segment rather than betting exclusively on one approach. AI’s transformation of the total addressable market has changed the ROI calculus: the market is now large enough to justify specialized technology for specialized use cases, which was not true in the traditional telecom era.

An audience question raised the possibility of a Multi-Source Agreement (MSA) or formal industry consortium. Alvarez confirmed that Avicena Tech is actively investigating this — talking to peers and competitors about alliance formation to align on common components and interfaces. Applied Materials, as Patibandla made clear, is already convening workshops and is positioned as a neutral ecosystem enabler that benefits from broad adoption regardless of which specific microLED architecture prevails.

Bottom Line

The Display Week 2026 microLED datacom panel made clear that the question is no longer whether microLED belongs in the data center interconnect conversation. It does. Avicena Tech has demonstrated 1 Tbps aggregate from a 400-element array; hyperscalers are asking for it at OFC; Marvell is investing across the technology space; and Applied Materials is convening the ecosystem.

What remains unresolved is the harder part: building the standards, test methodologies, fiber array ecosystem, driver IC infrastructure, and ultimately the chip-level architectural integration that realizes the full power efficiency and density advantage. The first commercial milestone — an active optical cable replacing a DAC — is targeted for end of 2026 and represents the industry’s first volume proof point. What comes after depends on how quickly the ecosystem can be assembled, and whether the display industry’s manufacturing expertise transfers effectively to a datacom context.

The shower head problem is still unsolved. But at least now the right people are in the room asking about it.