The high power silicon photonics (SiPh) chip market is moving from optical-component experimentation toward the architecture of AI-scale connectivity. The supplied report values the market at USD 5.473 billion in 2024 and projects USD 10.36 billion by 2032, but current company programs show why the technology is strategically important: Intel reports more than 8 million photonic integrated circuits shipped and more than 32 million on-chip lasers, Marvell is pushing 1.6T optical DSPs into mass volume with hyperscaler customers, and Coherent is planning a 1.6T/3.2T and 12.8T-class optical roadmap. GlobalFoundries is working toward 200G per wavelength and a path to 400G per wavelength, while Lightmatter is organizing an OCP effort for optical connectivity across AI clusters. The market is therefore becoming a systems question involving lasers, modulators, packaging, DSPs and co-packaged optics.

Intel

Intel has one of the clearest commercial records in silicon photonics, combining photonic integrated circuits, optical transceivers and integrated lasers. Its significance is amplified by the need to connect increasingly large AI clusters without allowing electrical I/O power and distance to become bottlenecks.

Intel reports more than 8 million photonic integrated circuits shipped and more than 32 million on-chip lasers. Its optical compute interconnect work includes a first-generation co-packaged optical I/O platform at 4Tbps bidirectional bandwidth, positioning silicon photonics as a path toward tighter integration between compute and optical connectivity.

Competitive significance: Intel demonstrates that SiPh can be a high-volume manufacturing platform, not only a laboratory technology.

Intel’s shipment history is important because it demonstrates that photonic integration can be industrialized. The challenge now is extending that manufacturing experience into higher optical bandwidth and tighter package integration.

Marvell

Marvell is important because it connects silicon photonics to the optical DSP and interconnect electronics that hyperscale data centers need at 1.6T and beyond. The company’s advantage is the ability to co-design electrical and optical interfaces around the same system requirements.

In March 2026 Marvell highlighted its 1.6T optical DSP portfolio and described its 3nm Ara platform entering mass volume with hyperscalers. The company has also received recognition for its 1.6T silicon-photonics light-engine work. These developments show that the competitive target is no longer simply a faster transceiver; it is an optical system that fits the power and thermal envelope of an AI cluster.

Competitive significance: Marvell is competing where photonics, DSP and packaging converge at the data-center link level.

Marvell’s 1.6T roadmap shows how photonics and DSP must evolve together. Higher optical lane rates are only useful if the electrical interface, FEC and thermal budget can support them in the same platform.

Coherent

Coherent has a broad photonics portfolio spanning silicon photonics, InP lasers, VCSELs and optical modules. Its position is important because future AI networks will likely use more than one photonic material system, depending on reach, power and integration requirements.

In March 2026 Coherent outlined a roadmap covering 1.6T and 3.2T optical connectivity and a 12.8T-class XPO roadmap. The company also described a 6.4T socketed co-packaged-optics approach using 400G per lane. Those specifications show the scale of the bandwidth problem the industry is trying to solve without linearly increasing electrical I/O power.

Competitive significance: Coherent competes across the full photonic stack and can shift architectures as data-center requirements evolve.

Coherent’s multi-technology portfolio is strategically valuable because different links may favor different photonic materials and lasers. Flexibility across SiPh, InP and VCSEL technologies reduces architecture lock-in.

GlobalFoundries

GlobalFoundries is strategically relevant because its silicon-photonics platform brings photonic device manufacturing into a 300mm foundry environment. The foundry model can allow multiple optical-system companies to build products without owning a dedicated photonics fabrication plant.

The company’s platform targets 200G per wavelength with a path toward 400G per wavelength. It also describes low-loss CWDM filters below 1dB insertion loss, multi-wavelength bidirectional DWDM structures and copper-pad pitches below 45 micrometers. A $300 million preliminary funding arrangement from the U.S. Commerce Department also supported expansion of its U.S. semiconductor manufacturing ecosystem.

Competitive significance: GF is competing on manufacturing scale and ecosystem access, turning SiPh into a foundry platform rather than a vertically integrated product only.

GlobalFoundries demonstrates the role of a foundry model in scaling photonics. A mature 300mm manufacturing process can make it easier for fabless companies to iterate optical designs without building their own wafer lines.

Lightmatter

Lightmatter is an AI-connectivity company using photonics to attack the bandwidth and power limits inside large AI systems. Its role is more architectural than conventional transceiver supply: the company is interested in optical interconnect as infrastructure for scaling clusters.

In August 2026 Lightmatter announced participation in an Open Compute Project initiative with a 19-company coalition focused on infrastructure that can scale AI clusters from 72 to more than 1,024 nodes. The program is significant because photonics is being considered at system-architecture level, not merely as a faster pluggable transceiver.

Competitive significance: Lightmatter shows how silicon photonics can become a platform-level interconnect technology for AI cluster design.

Lightmatter’s cluster-scale focus changes the unit of competition from one transceiver to the network fabric. Optical connectivity can become part of the topology decision for large AI systems.

Cisco

Cisco is relevant because it combines silicon-photonics technology with networking systems, giving it visibility from the optical component to the switch and network architecture. Its optical strategy spans 400G through 1.6T-class pluggable connectivity.

The competitive advantage is vertical understanding: optical modules have to match switch ASIC bandwidth, thermal limits, signal integrity and network architecture. Cisco’s SiPh-based optics are therefore designed in the context of the complete network rather than as an isolated optical component.

Competitive significance: Cisco demonstrates the pull from system vendors for photonics that is tightly matched to switching and network-scale requirements.

Cisco’s position matters because networking silicon and optical modules have to operate as one system. A photonic component that performs well on a test bench but raises switch power can still be unattractive commercially.

Ayar Labs

Ayar Labs is developing optical I/O technologies that move data optically between compute devices, targeting a future in which electrical interconnect becomes a major limitation for AI accelerator scaling. Its approach is relevant to silicon photonics because it uses integrated photonics to shorten the electrical path.

The company’s technology combines optical engines with compute-oriented interfaces, aiming to improve bandwidth density and energy efficiency. Its relevance is therefore closer to chip-to-chip and package-level I/O than to conventional telecom optics.

Competitive significance: Ayar Labs represents the optical-I/O route in which photonics is integrated more tightly with compute packages.

Ayar Labs’ optical-I/O approach targets the electrical distance between compute components. The strategic goal is to move data without allowing interconnect energy to rise as quickly as compute density.

Celestial AI

Celestial AI is another AI-interconnect company applying photonics to the memory and compute fabric. Its platform focuses on optical connections that can move data between processors and memory resources with lower distance and power penalties.

The technology is relevant because AI accelerators increasingly depend on moving data between compute and memory, and the electrical network can become the limiting factor before the compute engines themselves are fully utilized.

Competitive significance: Celestial AI shows that silicon photonics is expanding from network links toward memory-centric AI architectures.

Celestial AI’s memory-centric approach expands the photonics discussion beyond Ethernet. If optical links can connect memory and compute more efficiently, the architecture of AI systems can change as well.

Broadcom

Broadcom participates across optical interconnects, network switching and high-speed connectivity, giving it a broad position in the silicon-photonics ecosystem. Its relevance is strongest where switch bandwidth growth forces a corresponding increase in optical port capacity.

The company’s optical portfolio spans high-speed transceivers and interconnect components that sit close to the network ASIC. This system adjacency is important because optical performance, power consumption and switch throughput have to be optimized together.

Competitive significance: Broadcom illustrates the system-level pull for photonics created by higher-bandwidth Ethernet and AI networking.

Broadcom’s combination of switching and optics gives it visibility into the bandwidth growth of data-center networks. This system perspective can help align optical ports with the switching roadmap.

Soitec

Soitec is relevant to silicon photonics through engineered substrates and photonic-SOI materials. The substrate is not the final optical engine, but it can determine how waveguides, optical isolation and device integration are implemented.

The company’s photonic-SOI platform serves optical and photonic components where a controlled silicon layer and buried oxide structure support waveguide and device fabrication. This makes substrate engineering part of the photonics manufacturing chain.

Competitive significance: Soitec represents the upstream substrate layer of the SiPh ecosystem, where material architecture influences device integration and manufacturing yield.

How the Competitive Landscape Is Changing

Competition in silicon photonics is becoming increasingly layered. Intel and GlobalFoundries demonstrate manufacturing-oriented SiPh platforms; Marvell and Broadcom tie photonics to high-speed networking silicon; Coherent supplies multiple photonic material systems; and Lightmatter, Ayar Labs and Celestial AI are pushing photonics deeper into AI system architecture.

The practical bottleneck is no longer the ability to make an optical link. The challenge is delivering bandwidth density at acceptable power, thermal load and package cost. A 1.6T link changes the requirements for DSPs, lasers, modulators, packaging and cooling simultaneously.

This is why co-design is becoming the defining competitive model. Suppliers that can coordinate photonics with switch ASICs, electrical SerDes, packaging and system architecture have a stronger path to hyperscale adoption than suppliers offering a photonic component in isolation.

Key Technology Trends Shaping the Top Players

1.6T is becoming a system-level target

Marvell and Coherent are already discussing 1.6T-class optical solutions, with roadmaps extending to 3.2T and beyond.

At these rates, power and package density become as important as optical modulation speed.

Photonic I/O is moving closer to the compute package

Intel, Ayar Labs and Celestial AI are pushing optical interfaces toward co-packaged or chip-adjacent architectures.

Shorter electrical paths can reduce I/O energy and enable scaling beyond conventional board-level connectivity.

Foundry manufacturing is becoming critical

GlobalFoundries’ 300mm SiPh platform shows that photonics can be manufactured using semiconductor-style foundry infrastructure.

This model can lower the barrier for system companies that need photonics but do not own fabs.

AI clusters are creating new interconnect topologies

Lightmatter’s OCP initiative for clusters above 1,024 nodes shows photonics being considered at the architecture level.

The future competitive question is how optical fabrics can reduce bottlenecks between accelerators, memory and switches, not simply how fast one transceiver can run.

Packaging and thermal design may determine which photonics architectures scale

Optical bandwidth can rise quickly, but the surrounding package still has to remove heat from DSPs and lasers while maintaining optical alignment and electrical integrity. At 1.6T and above, packaging is therefore part of the system specification.

The companies moving toward co-packaged optics and optical I/O are effectively competing on this integration problem. Manufacturing scale alone is not enough if a photonic engine cannot be assembled, cooled and serviced economically.