Author: Godot
The two hottest sectors in AI are storage and optics. I previously wrote about the storage framework (“Understanding the Profit Pools and Industry Landscape of AI Storage Layers”); this time, let’s cover optics.
Silicon photonics is used for communication between computing chips, replacing traditional copper wires, as shown clearly in the diagram below.
LPO (Linear Pluggable Optics), CPO (Co-Packaged Optics), OCS (Optical Circuit Switching), and Optical I/O (Optical Input/Output) are different technical approaches to achieving silicon photonics.
Typically, chips communicate using copper wires. Silicon photonic chips integrate light-generating lasers, light-modulating modulators, and light-detecting detectors directly onto the silicon chip, using photons for communication.
So why replace copper? And why use silicon photonics instead of something else?
First, copper wires almost reach their physical limit when transmitting signals above 1.6T, causing signal degradation. A material change must be considered. This is the most critical issue and absolutely necessary. The term for this is the bandwidth wall.
Second, copper is a physical material, and as GPU clusters grow larger, there simply isn’t enough space to accommodate all the copper wiring. This is precisely why replacing copper becomes necessary. Light, on the other hand, allows optical interfaces to be directly soldered next to switch chips, eliminating the need for extensive cabling. This is known in technical terms as the “scale wall.”
Again, copper is too power-hungry; in multi-megawatt facilities, silicon photonics can save tens of thousands of kilowatt-hours per day—electricity that would otherwise be consumed by copper-based communication. By switching to optical, that saved power can be redirected to GPUs for actual computation. This is known as the power wall.
More interestingly, silicon photonics can leverage the mature CMOS manufacturing processes already in use for 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 relies 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 milestone is March 2025, when NVIDIA unveiled the Quantum-X and Spectrum-X photonic switches at GTC, and Jensen Huang announced that starting with the next-generation Rubin, "optical interconnects are no longer optional—they are standard."
One week later, NVIDIA announced a combined investment of $4 billion in Coherent and Lumentum to secure critical supply chains.
The silicon-based photonic effect paper was published in the 1980s, and Intel and IBM manufactured silicon-based optical modulators between 2004 and 2014.
In the previous decade, hyperscale cloud providers such as AWS, Google, and Meta adopted silicon photonics, but it was only a part of optical fiber communication at that time.
Current industry landscape
1) The bottom layer: Foundry
Manufacturing photonic chips. TSMC $TSM leads with its COUPE process, Tower Semiconductor $TSEM specializes in silicon photonics foundry services, with silicon photonics revenue growing 70% year-over-year in 2025. GlobalFoundries $GFS, through its acquisition of Singapore’s AMF, has become the world’s largest dedicated silicon photonics foundry.
2) Layer 2: Core Component Suppliers
Fewer than five companies worldwide provide lasers, modulators, and other components, primarily indium phosphide (InP) lasers capable of manufacturing high-speed EML lasers.
Lumentum $LITE is the only manufacturer capable of mass-producing 200G/lane EML lasers, which are core components of 1.6T optical modules. NVIDIA has secured its production capacity by signing orders through 2027 and beyond.
3) Third layer: Modules and system manufacturers
Assemble components into finished products. Coherent holds a 25% global market share in optical transceivers. Chinese manufacturers InnoLight, Eoptolink, and Accelink are formidable in terms of manufacturing scale and cost competitiveness.
4) Top level: System integrator
NVIDIA, Cisco, Broadcom, and Marvell are all at this layer.
Overall,
NVIDIA$NVDA
Dominant position, determine which interconnection standards AI data centers adopt, then secure the supply chain through strategic investments.
Broadcom$AVGO
The absolute leader in network switching chips, with nearly 80% market share in Ethernet switches. The Tomahawk 6-Davisson is the world's first 102.4 Tbps CPO switch.
Marvell$MRVL
Broadcom's strongest competitor, dominating the PAM4 optical DSP market with a 60-70% share. Recently acquired Celestial AI to enter chip-to-chip optical interconnects.
Lumentum$LITE
The leading supplier of EML lasers. The only manufacturer globally capable of mass-producing 200G/lane EMLs, with NVIDIA having secured orders through 2027 and beyond.
Coherent$COHR
An integrated player across the entire industrial chain, with presence in materials, lasers, and modules. Revenue of $5.8 billion in FY2025, making it the market leader in optical transceivers.
TSMC$TSM
Process standard setter. The 65nm silicon photonics process is already in mass production, and the COUPE platform is the most advanced 3D heterogeneous integration solution currently available, deeply integrated with NVIDIA's CPO roadmap.
Tower Semiconductor$TSEM
The purest beneficiary of silicon photonics foundry. Silicon photonics revenue grew 70% year-over-year in 2025, with $650 million being invested to triple capacity. It has the strongest market cap elasticity among all targets.
Lightmatter / Ayar Labs - Not yet public · IPO candidate
Lightmatter is valued at $4.4 billion and specializes in 3D photonic interconnects; Ayar Labs has received investments from AMD, Intel, and NVIDIA simultaneously and focuses on optical I/O chiplets. Both are potential major IPO candidates.
The surge in silicon photonics is changing the valuation logic.
For example, Wall Street previously valued Tower Semiconductor as a conventional analog foundry, with a price-to-sales ratio of about 2 to 3 times.
But as the silicon photonics business grows from 5% of total revenue to 30%-40%, the market begins to revalue it as a scarce asset in AI infrastructure, with the price-to-sales ratio expected to rise to 6-10x.
Lumentum and Coherent were previously telecommunications component suppliers, now redefined as essential providers of AI-interconnected components. BofA analyst Vivek Arya raised Marvell’s price target to $200, valuing Marvell as an AI infrastructure platform rather than a communications chipmaker.
Evercore ISI’s assessment of Cisco is similar: as silicon photonics products penetrate hyperscale data centers, Cisco’s AI core revenue could surge from $3 billion to $12–15 billion over the next 3–4 years.
The moat of the silicon photonics industry
The silicon photonics industry exhibits a clear winner-takes-all characteristic, as each process has undergone long-term development prior to the AI boom.
Only 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 critical bottleneck in the entire supply chain.
TSMC's COUPE process. The process barrier for 3D heterogeneous integration leaves competitors at least two generations behind, requiring years of yield experience to catch up.
The foundry PDK ecosystem. Once a customer completes a design at a particular foundry, the switching cost is extremely high, requiring 12 to 18 months for redesign and recertification.
Thermal management and packaging. CPO requires managing the coupling of electrical, thermal, and optical domains within a few millimeters of space—impossible without years of system integration experience.
The supplier certification process for giants like AWS and Google typically takes 12 to 24 months. Once completed, customer loyalty is extremely high.
Risks and Cool-headed Reflection
The growth of the entire industry chain is highly dependent on the capital expenditures of the five hyperscale cloud providers: Microsoft, Google, Meta, Amazon, and Oracle.
The technical pathways are interchangeable—LPO (Linear Pluggable Optics), CPO (Co-Packaged Optics), OCS (Optical Circuit Switching), and Optical I/O. If one pathway is displaced by another, the capital previously invested may face depreciation.
Research institutions such as LightCounting judge that large-scale CPO deployment will not arrive until after 2028, with transitional solutions like LPO dominating prior to that.
Therefore, betting on the industry winning is safer than betting on a single company winning.