Photonics Stocks: Silicon Photonics and the AI Bandwidth Race

This is not financial advice. Do your own research before making any investment decisions.

By Daniel Reyes, S4Tips Markets Desk

The bottleneck in AI compute is not the GPU anymore. It is the wire between the GPUs. As AI clusters scale from hundreds to tens of thousands of accelerators, the copper interconnects that once worked fine for data centers are hitting a hard wall: they burn too much power, generate too much heat, and cannot move data fast enough at the distances involved. Silicon photonics, a technology that uses light rather than electrical signals to transmit data, is now the central engineering bet for solving that problem. The companies building the chips, transceivers, and optical packaging that enable this shift represent a distinct investment category within semiconductor stocks, and one that is getting serious attention from infrastructure buyers.

This article maps the technology, identifies the key players, and explains why the AI bandwidth thesis makes photonics one of the more interesting structural stories in the current build-out of AI infrastructure stocks.

What Silicon Photonics Actually Is (The Citable Definition)

Photonics is the branch of physics and engineering concerned with the generation, transmission, and detection of light (photons) as a carrier of information. In data communications, photonic systems replace copper-wire electrical signals with optical signals traveling through waveguides or fiber, achieving higher bandwidth at lower power consumption over distances where copper degrades.

Silicon photonics is a specific manufacturing approach that integrates optical components, including lasers, modulators, waveguides, and photodetectors, directly onto a silicon chip using standard CMOS fabrication processes. This matters commercially because it allows optical components to be manufactured at semiconductor scale and cost, rather than as expensive bespoke assemblies.

Co-packaged optics (CPO) takes the concept a step further: instead of a separate pluggable optical transceiver module at the port of a network switch or GPU board, the optical engine is packaged directly alongside the switch ASIC or processor on the same substrate. This eliminates the copper trace distance between the chip and the optical interface, reducing latency and cutting power consumption substantially compared to conventional pluggable form factors.

Why AI Clusters Created a New Demand Curve for Optical Interconnects

Traditional enterprise data centers moved data at rates where pluggable optical transceivers were more than adequate. AI training clusters operate differently. A large-scale GPU cluster running distributed training across thousands of accelerators must continuously exchange gradient updates, activations, and parameter states between nodes. The bandwidth required scales roughly with the number of GPUs, and the power budget for moving that data is a genuine engineering constraint, not a theoretical one.

Copper-based electrical links work at short distances and low speeds, but as per-port speeds climb past 400 Gbps toward 800 Gbps and 1.6 Tbps, copper’s physics become punishing. Signal loss increases with both distance and frequency, which means more repeaters, more power amplification, and more heat, all of which cost money and rack space. Optical fiber does not have that problem: a photon traveling through glass does not degrade the way an electron does through copper at high frequencies.

This is not a speculative shift. The hyperscalers building the current generation of AI data center stocks have publicly committed to optical networking architectures. The question for investors is which part of the optical stack captures the most durable value.

The Three Technology Tiers: Where Value Accumulates

The photonics supply chain has three distinct tiers, and their competitive positions are quite different.

The transceiver market is where most of the near-term volume flows, because pluggable form factors work with existing switch and server hardware. The CPO market is smaller today but is where the industry’s long-term power efficiency argument lives. If CPO deployment accelerates alongside next-generation switch ASICs, the companies already designing CPO solutions in partnership with major ASIC vendors will have a structural head start.

The Key Players in Photonics Stocks

Several companies across the optical supply chain are positioned to benefit from the AI bandwidth build-out. Their situations vary considerably in terms of scale, customer concentration, and technology differentiation.

Coherent Corp. (COHR)

Coherent is one of the largest vertically integrated optical component and transceiver companies in the world. It manufactures the indium phosphide (InP) laser chips that sit at the heart of most high-speed transceivers, along with the finished transceiver modules themselves. Vertical integration in lasers is a real advantage because InP fabrication requires specialized foundry capacity that silicon fabs cannot simply replicate. Its 800G transceiver volumes have grown substantially as hyperscaler AI spending has accelerated, and it is shipping into data center builds at scale. The company publishes quarterly earnings with data-center-specific revenue disclosures, which makes its AI exposure more directly traceable than some peers.

II-VI / Now Coherent Corp. Note on Naming

Coherent Corp. was formed from the merger of II-VI Incorporated and legacy Coherent in 2022. References to “II-VI” in older analyst coverage refer to the same entity.

Lumentum Holdings (LITE)

Lumentum is a laser and photonics components specialist with meaningful exposure to both telecom and data-center markets. It supplies pump lasers, edge-emitting lasers, and photonic integrated circuits (PICs), and has been gaining ground in the silicon photonics transceiver supply chain. Lumentum’s data-center revenue has grown as a share of total revenue as AI-driven demand has displaced the slower telecom upgrade cycles that traditionally defined its business. It also has exposure to the 3D sensing (lidar, face recognition) market, which is unrelated to AI interconnects but adds complexity to revenue modeling.

Marvell Technology (MRVL)

Marvell deserves a place in this category even though it is more widely known as a custom ASIC and networking silicon company. Marvell has made silicon photonics a central plank of its data-center interconnect strategy, developing optical DSPs (digital signal processors) that interface between electrical and optical domains at very high speeds. Its Teralynx switch platform and its PAM4 DSP portfolio are both designed around the optical networking architectures that AI clusters require. Marvell’s relationship with major cloud customers for custom AI silicon gives it visibility into where next-generation optical interconnect standards will land before those decisions become public. See our broader coverage of semiconductor stocks for context on where Marvell sits in the chip ecosystem.

Broadcom (AVGO)

Broadcom is the dominant provider of switch ASICs for data-center networking, with its Tomahawk and Trident series installed in a very large share of the world’s hyperscale switch fabrics. Broadcom is now actively developing co-packaged optics solutions that integrate optical engines directly with its next-generation switch silicon. Its scale and its position at the center of switch ASIC design means it can define the physical interface standards that transceiver vendors must conform to. For investors, Broadcom’s CPO program is both an opportunity (significant incremental revenue if CPO adoption accelerates) and a risk factor for standalone transceiver vendors who could be partially bypassed.

Nvidia (NVDA)

Nvidia is relevant here because its NVLink and NVSwitch interconnect architectures directly determine the bandwidth requirements between GPU clusters, and because Nvidia has been increasingly open about the role of optical technologies in its future networking roadmap. Nvidia’s acquisition of Mellanox brought it deep networking expertise, and its Quantum InfiniBand and Spectrum Ethernet switch platforms are deployed in many of the same AI clusters that are driving optical transceiver demand. Nvidia is not a photonics manufacturer per se, but its product decisions set the bandwidth specifications that the entire supply chain must meet.

Ciena Corporation (CIEN)

Ciena has historically served the long-haul and metro optical networking market for carriers, but the line between telecom optical transport and data-center interconnect (DCI) is blurring as hyperscalers build private wide-area networks connecting their campuses. Ciena’s WaveLogic coherent optical platform is deployed in DCI applications, and the company has been positioning itself for the AI-driven surge in between-data-center bandwidth. Its exposure is different from transceiver pure-plays: Ciena sells complete optical transport systems, not just modules, which means larger average deal sizes and longer sales cycles.

Applied Optoelectronics (AAOI)

Applied Optoelectronics is a smaller, more concentrated pure-play on data-center transceivers. It manufactures its own lasers and assembles finished transceiver modules, with a customer base historically concentrated among a small number of hyperscalers. That concentration is both its strength (deep customer relationships, co-development visibility) and its primary risk. Revenue can swing sharply when a single major customer shifts its sourcing mix or changes a network upgrade cycle. It is the highest-risk, highest-leverage name in this peer group for investors who want direct photonics exposure without the diversification of larger conglomerates.

The Co-Packaged Optics Bet: Long-Dated but Potentially Decisive

CPO is the technology most discussed at the engineering level but least reflected in current revenue figures for any of the companies above. The technical argument for CPO is clear: by eliminating the copper path between a switch ASIC and its optical interfaces, CPO can reduce the power consumed by I/O by a substantial fraction compared to pluggable transceivers, and it can enable higher bandwidth density per rack unit.

The practical challenge is that CPO requires the optical vendor and the ASIC vendor to co-design their solutions, which means customer lock-in runs in both directions. A hyperscaler that commits to a specific CPO platform effectively commits to the ASIC-photonics pairing for that generation of hardware. The companies building those partnerships now, whether at the component level (lasers, PICs) or the module level (optical engine design), are positioning for a product cycle that may not fully ramp until the second half of this decade.

The Ethernet Alliance and major switch ASIC vendors have published CPO interoperability specifications, which are a prerequisite for broad ecosystem adoption. The fact that those standards exist and are actively being tested by multiple vendors is a signal that CPO has moved from research to pre-commercial development, even if volume production timelines remain uncertain.

Power Efficiency as the Real Investment Thesis

The bandwidth argument for photonics is real, but the power argument may be the more durable investment thesis. AI data centers are running into power density limits: the amount of electricity a building’s electrical infrastructure and cooling systems can deliver per square meter of floor space. A rack of GPU accelerators running at full utilization draws power at rates that require substantial cooling investment, and the network switches and transceivers serving that rack add to the load.

If CPO and silicon photonics can genuinely reduce the power consumed by data-center networking by a meaningful fraction relative to copper or conventional pluggable transceivers, that efficiency gain has real dollar value for hyperscalers operating at scale. It reduces power bills, reduces cooling costs, and potentially allows more compute density in the same physical footprint. That is a purchasing argument that data-center operators understand clearly, and it is separate from the performance argument about bandwidth and latency.

The photonics industry has published credible technical papers (including work from groups at MIT and Intel’s photonics research teams) supporting the power-per-bit advantage of optical interconnects at high speeds. According to IEEE, silicon photonics interconnects operating at 800 Gbps and above show substantially lower energy per bit than equivalent electrical links at comparable distances. The Ethernet Alliance has documented CPO interoperability work that underpins the ecosystem standardization required for broad deployment.

Risk Factors Worth Naming

The photonics opportunity is real, but investors should understand the specific risks rather than treating this as a simple AI-upside trade.

Hyperscaler concentration is the most immediate risk. A small number of customers, primarily the major US cloud providers plus a handful of large Chinese internet companies, account for a disproportionate share of data-center transceiver demand. If any of those customers slows its networking upgrade cadence, shifts sourcing between vendors, or builds more vertical integration into optical components (as some have attempted), revenue at the pure-play optical vendors can move sharply.

Technology transition risk runs in both directions. The move from 400G to 800G to 1.6T creates upgrade revenue, but it also requires vendors to successfully qualify new products with demanding customers on compressed timelines. A company that misses a speed-generation transition can lose shelf position that is difficult to recover. CPO, specifically, represents a technology transition with no guarantee of the adoption timeline that current bullish sentiment implies.

Inventory correction cycles have historically been severe for optical vendors. The transceiver market went through a significant inventory digestion period in 2022-2023 as hyperscaler capex slowed temporarily after aggressive pandemic-era build-outs. AI-driven demand has changed the underlying growth trajectory, but it has not eliminated the possibility of future inventory corrections if build plans are revised.

FAQ: Photonics Stocks and the AI Interconnect Story

What are photonics stocks?

Photonics stocks are shares of companies that design, manufacture, or sell products based on light-based data transmission. In the AI infrastructure context, the relevant segment includes makers of optical transceivers, silicon photonics integrated circuits, co-packaged optical engines, and the laser and modulator components that these systems require. Companies like Coherent Corp., Lumentum, and Applied Optoelectronics are the most direct pure-play exposures; Broadcom and Marvell are larger-cap names with significant photonics programs embedded in broader businesses.

Why does AI create demand for optical interconnects specifically?

AI training and inference clusters require very high bandwidth between GPU accelerators, often at distances where copper electrical signaling becomes too power-hungry or too signal-degraded to operate reliably. Optical interconnects, which move data as light through fiber or silicon waveguides, can transmit data at higher speeds over longer distances with lower power consumption per bit. As GPU cluster sizes increase, the aggregate bandwidth required grows accordingly, and photonics is the dominant technology being deployed to meet that demand in the networking layer.

What is the difference between an optical transceiver and co-packaged optics?

An optical transceiver is a pluggable module that slots into a port on a network switch or server, converting electrical signals to optical and back. It is hot-swappable and works with existing hardware. Co-packaged optics places the optical engine on the same substrate as the switch ASIC itself, eliminating the short copper trace between chip and port. CPO is more power-efficient and can support higher bandwidth density, but it requires the switch silicon and the optical engine to be designed together, which means it is not backward-compatible with installed hardware and requires closer supply-chain integration between vendors.

Are photonics stocks only about AI data centers?

No, though AI is the dominant growth driver right now. Photonics companies also serve the long-haul telecom market (undersea cables, carrier backbone networks), metro optical transport, cable television networks, and in some cases consumer applications like 3D sensing in smartphones and lidar for autonomous vehicles. For most of the pure-play photonics vendors, data-center revenue has grown to dominate their mix, but understanding their telecom and non-data-center exposure matters for modeling revenue cyclicality. Telecom capital spending tends to move on different cycles from data-center spending.

How do silicon photonics companies fit within the broader semiconductor sector?

Silicon photonics companies occupy a specific niche within the semiconductor and photonics supply chain, sitting between traditional chip fabs and optical component specialists. They use CMOS fabrication processes (often at foundries like TSMC or GlobalFoundries, or in their own specialized fabs) to manufacture integrated optical circuits, but their end-markets and customers overlap with both the networking semiconductor world and the fiber-optic component industry. From an investor classification standpoint, most photonics companies appear in communications semiconductor or electronic components categories, and their revenue trends correlate with data-center capex cycles more than with consumer electronics or mobile chip demand.

What should I look at in a photonics company’s earnings to gauge AI exposure?

The most direct indicators are: the share of revenue classified as “datacom” versus “telecom” (higher datacom exposure means more AI data-center sensitivity); revenue from 400G and 800G transceiver products specifically (speed transitions reflect AI cluster build timelines); any disclosed CPO program revenue or design wins; and customer concentration disclosures, which reveal hyperscaler dependency. Companies with significant revenue from two or three named hyperscaler customers have high upside in an AI capex boom and high downside if those customers slow or shift sourcing. Watch gross margin trends as well: optical vendors with proprietary laser manufacturing tend to defend margins better through competitive cycles than those sourcing components externally.

Daniel Reyes

Daniel Reyes is a markets writer for S4Tips covering the AI infrastructure and semiconductor supply chain. He focuses on the companies that build and power the AI compute stack. His articles are for information only and are not financial advice.