Author: Godot

The two hottest sectors in AI are storage and optics. I previously wrote about the framework of storage ( Understanding the Profit Pool and Industry Landscape of AI Storage Layers in One Article ), and this article will focus on optics.

Silicon photonics, also known as silicon photonics, is used for communication between computing chips, replacing the original copper wires. The diagram below illustrates this clearly.

The terms LPO (Linear Pluggable Optics), CPO (Co-packaged Optics), OCS (Optical Path Switching), and Optical I/O (Optical Input/Output) that people may encounter and find confusing are different technical approaches to achieving silicon photonics.

Traditionally, chips use copper wires for communication. Silicon photonics chips integrate a laser that generates light, a modulator that modulates light, and a detector that receives light directly onto a silicon chip, enabling photonic communication.

So, why replace copper? And why use silicon photonics instead of something else?

First, copper wires reach near their physical limit when transmitting signals above 1.6T, at which point the signal becomes incomplete. Changing the material must be considered. This is the most critical issue, and the only option. The technical term is "bandwidth wall."

Secondly, copper is a tangible physical component, and with the increasing size of GPU clusters, there simply isn't enough space for copper cabling. This is another reason why copper had to be replaced. Optical is different; the optical interface can be soldered directly next to the switch chip, saving a significant amount of wiring. The technical term is "scale wall."

Secondly, copper consumes too much electricity. Silicon photonics, in megawatt-scale facilities, can save tens of thousands of kilowatt-hours of electricity per day, all of which is consumed in copper wire communication. Switching to optical fiber allows that electricity to be redirected to GPUs for actual computation. The technical term is the power wall.

What's even more interesting is that silicon photonics can leverage the mature CMOS manufacturing process of existing semiconductors, eliminating the need to build a new factory from scratch, and enabling low-cost mass production.

Of course, silicon photonics also has a drawback: silicon itself cannot emit light efficiently and must rely on indium phosphide (InP) materials. This has become the most critical bottleneck in the entire industry chain.

Evolution of silicon photonics technology

The most important watershed moment was in March 2025, when NVIDIA released the Quantum-X and Spectrum-X photonic switches at the GTC conference. Jensen Huang announced that starting with the next generation Rubin, "optical interconnects will not be an option, but a standard feature."

A week later, NVIDIA announced a combined $4 billion investment in Coherent and Lumentum to secure key supply chains.

Papers on silicon-based photoelectric effects were published in the 1980s, and Intel and IBM manufactured silicon-based optical modulators between 2004 and 2014.

In the last decade, large cloud service providers such as AWS, Google, and Meta used silicon photonics, but at that time it was only a part of fiber optic communication.

Current industrial landscape

1) Bottom layer: Wafer foundry

Manufacturing photonic chips. TSMC (TSM) leads with its coupe process, while Tower Semiconductor (TSEM) specializes in silicon photonics foundry services, with silicon photonics revenue projected to grow by 70% year-on-year in 2025. GlobalFoundries (GFS) , through its acquisition of Singapore's AMF, has become the world's largest dedicated silicon photonics foundry.

2) Second tier: Core component suppliers

They provide lasers, modulators, etc., mainly indium phosphide (InP) lasers. There are fewer than five companies worldwide that can manufacture high-speed EML lasers.

Lumentum $LITE is the only manufacturer capable of mass-producing 200G/lane EML lasers, a core component of 1.6T optical modules. NVIDIA has already placed orders with Lumentum until 2027 to secure its production capacity.

3) Third layer: Module and system manufacturers

They assemble the components into products. Coherent holds a 25% market share in the global optical transceiver market. Chinese companies like InnoLight, Eoptolink, and Accelink are formidable competitors in terms of manufacturing scale and cost competitiveness.

4) Top level: System integrator

NVIDIA, Cisco, Broadcom, and Marvell are all on this floor.

In summary,

NVIDIA $NVDA

A dominant position determines which interconnect standards AI data centers adopt, and then secures the supply chain through strategic investments.

Broadcom $AVGO

The undisputed leader in network switching chips, holding nearly 80% of the Ethernet switch market share. The Tomahawk 6-Davisson is the world's first 102.4 Tbps CPO switch.

Marvell $MRVL

Broadcom's strongest challenger, dominating the PAM4 optical DSP market with a 60-70% share. Its recent acquisition of Celestial AI allows it to penetrate the chip-to-chip optical interconnect market.

Lumentum $LITE

The most critical supplier of EML lasers. The only manufacturer in the world capable of mass-producing 200G/lane EML, NVIDIA has already locked in orders until after 2027.

Coherent $COHR

A full-industry chain integrator, with a presence in materials, lasers, and modules. With FY2025 revenue of $5.8 billion, it is the number one manufacturer in the optical transceiver market share.

TSMC $TSM

NVIDIA is a standard setter in process technology. Its 65nm silicon photonics process is already in mass production, and the COUPE platform is currently the most advanced 3D heterogeneous integration solution. NVIDIA's CPO roadmap is deeply tied to it.

Tower Semiconductor $TSEM

The purest beneficiary of silicon photonics foundry services. Silicon photonics revenue is projected to grow by 70% year-on-year in 2025, and the company is investing $650 million to triple its capacity. Its market capitalization elasticity is the strongest among all stocks.

Lightmatter / Ayar Labs: Unlisted/IPO candidate

Lightmatter, valued at $4.4 billion, focuses on 3D photonic interconnects; Ayar Labs, which received investment from AMD, Intel, and NVIDIA, focuses on optical I/O chips. Both are potential heavyweight IPO candidates.

The Silicon Photonics Boom Shifts Valuation Logic

For example, Wall Street used to value Tower Semiconductor based on the valuation of a typical contract manufacturer, with a price-to-sales ratio of about 2-3 times.

However, as the silicon photonics business grows from 5% to 30%-40% of total revenue, the market begins to revalue it as a scarce asset in AI infrastructure, and the price-to-sales ratio is expected to rise to 6-10 times.

Lumentum and Coherent, formerly telecommunications component suppliers, are now being redefined as essential component providers for AI connectivity. BofA analyst Vivek Arya raised his price target for Marvell to $200, valuing Marvell as an AI infrastructure platform rather than a communications chip manufacturer.

Evercore ISI has a similar assessment of Cisco: as silicon photonics products penetrate deeper into hyperscale data centers, Cisco's core AI revenue could surge from $3 billion to $12-15 billion in the next 3-4 years.

The moat of the silicon photonics industry

The silicon photonics industry exhibits a clear winner-takes-all characteristic because each process actually underwent a long period of development before the AI boom.

For InP lasers, fewer than five companies worldwide can mass-produce high-end EML lasers, with a capacity expansion cycle of 3-5 years. This is the most bottleneck link in the entire industry chain.

TSMC's Coupe process. A technological barrier in 3D heterogeneous integration, leaving pursuers at least two generations behind and requiring years of experience in improving yield rates.

The OEM PDK ecosystem. Once a customer has done design work with a particular OEM, switching is extremely costly, as redesigning and recertification can take 12-18 months.

Thermal management and packaging. A CPO needs to manage the coupling of three physical domains—electric, thermal, and optical—in a space of just a few millimeters. This requires several years of system integration experience.

Supplier certification processes for giants like AWS and Google typically take 12-24 months. Once certified, customer loyalty is extremely high.

Risk and Cold Thinking

The growth of the entire industry chain is highly dependent on the capital expenditure of the five hyperscale cloud vendors: Microsoft, Google, Meta, Amazon, and Oracle.

There are alternatives among the technical routes, such as LPO (Linear Pluggable Optics), CPO (Co-packaged Optics), OCS (Optical Path Switching), and Optical I/O (Optical Input/Output). If one route is overturned by another, the capital invested in the previous route may face depreciation and impairment.

Research institutions such as LightCounting predict that large-scale CPO deployments will not arrive until after 2028, with LPO and other transitional solutions being used more for testing before then.

Therefore, it is safer to bet on an industry to win than to bet on a single company to win.

Google and Blackstone form a joint venture AI cloud company, with Blackstone committing to invest $5 billion.

Popular Articles