Optoelectronic Semiconductors

Optoelectronic semiconductors generate, detect, or modulate light. They are built primarily on III-V compound semiconductor substrates — GaAs, InP, GaN — whose direct bandgap enables efficient photon emission and absorption that silicon's indirect bandgap cannot match. The supply chain consequence is that optoelectronics sits on a different substrate ecosystem from digital logic and memory: GaAs and InP wafers come from a small number of specialized growers (Wafer Technology, AXT, Freiberger, Sumitomo), epi layers are deposited in dedicated MOCVD reactors, and the resulting die must be precision-aligned and hermetically packaged — a process incompatible with standard semiconductor assembly automation.

Silicon photonics is the technology attempting to bring optical functions onto conventional silicon wafers and CMOS process infrastructure. It does not replace III-V optoelectronics — silicon cannot generate light efficiently — but it integrates passive optical functions (waveguides, modulators, splitters, photodetectors) with CMOS electronics, reducing the packaging complexity of optical transceivers. Co-packaged optics (CPO), which integrates silicon photonic transceiver dies directly with switch ASICs in the same package, is the dominant architecture trend for AI cluster datacenter interconnect at 800G and 1.6T speeds.

Optoelectronic Device Categories — Products & Process

Deployment & Supply Chain Risk

Co-Packaged Optics — The AI Cluster Interconnect Inflection

The AI training cluster has a networking problem that copper cannot solve at scale. At 800G and above, copper electrical interconnects between switch ASICs and server NICs face fundamental signal integrity limits beyond approximately 3 meters. Optical interconnect has always solved reach, but pluggable optical transceivers add power consumption, latency, and physical port density constraints that become acute inside a dense GPU cluster. Co-packaged optics solves the problem by integrating silicon photonic transceiver dies directly in the same package as the switch ASIC — removing the pluggable transceiver entirely and placing the optical-electrical interface at the package edge rather than at the faceplate.

Intel has been the most aggressive CPO proponent, building its Optical I/O program around its silicon photonics process. Broadcom, Marvell, and Ayar Labs are the other significant CPO participants. The critical supply chain dependency that CPO does not eliminate is the laser source: silicon cannot generate light, so every CPO module requires an InP or GaAs laser die bonded to the silicon photonic waveguide. Lumentum and II-VI/Coherent are the primary laser sources for CPO. The transition from pluggable transceivers to CPO is therefore a transition that reduces VCSEL and pluggable transceiver demand while increasing demand for precision InP laser die bonded at the package level — a shift in where the III-V compound semiconductor content sits in the supply chain, not an elimination of it.

InGaAs APD — The LiDAR Scale-Up Constraint

Automotive and AV LiDAR operating at 1550nm uses InGaAs avalanche photodiodes (APDs) as detectors because 1550nm is eye-safe at the power levels required for long-range sensing, and InGaAs has peak sensitivity in the 1550nm telecom window. The supply chain for InGaAs APDs is a direct derivative of the telecom InP supply chain — the same InP substrate growers, the same MOCVD epitaxy infrastructure, and many of the same device manufacturers (Lumentum, II-VI/Coherent) who built the telecom InP supply chain are now the primary InGaAs APD suppliers for LiDAR.

The problem is volume. Telecom InP APD production is measured in millions of units per year for submarine cables, DWDM amplifiers, and coherent receivers. Automotive LiDAR at AV fleet scale would require orders of magnitude more InGaAs APD units. The existing InP supply chain is not sized for this transition, and expanding it requires the same long-lead-time MOCVD reactor procurement, substrate supply development, and device qualification cycles that constrain every compound semiconductor scale-up. This is why InGaAs APD scarcity is consistently identified as the binding constraint on 1550nm LiDAR deployment at automotive volume — not the laser, not the signal processing, but the detector.

Substrate Dependencies — The III-V Foundation

Supply Chain Bottlenecks

Related Coverage

Power & Analog Hub | Solar PV Semiconductors | LiDAR Sensors — Sensing & Connectivity | CMOS Image Sensors | GaAs & InP Wafer Supply Chain | RF & Networking (GaN, GaAs RF) | Semiconductor Bottleneck Atlas

Cross-Network — ElectronsX Demand Side

Automotive LED demand scales with every EV and ICE vehicle headlight, DRL, and tail light design-in requiring AEC-Q102 qualified LED die. LiDAR VCSEL and InGaAs APD supply constraints directly gate AV sensor suite deployment timelines — the detector supply chain is the binding constraint on 1550nm LiDAR at AV fleet scale. Co-packaged optics for AI cluster interconnect is the datacenter infrastructure counterpart to the power and cooling demand covered on ElectronsX.

EX: ADAS/AV Compute Architecture | EX: EV Semiconductor Dependencies | EX: Humanoid Robots