The Point Where Copper Runs Out of Physics
The electrical interconnects that carry data between the graphics processing units, switching hardware, memory, and storage in a modern AI data centre are copper conductors whose signal integrity, power consumption, and maximum practical length create the physical constraints that limit the density and performance of AI computing infrastructure. At the data rates that current and next-generation AI accelerators operate at, copper interconnects within a server rack can carry signals over distances measured in metres before the electrical resistance, capacitance, and skin effect losses degrade signal quality to the point where reliable data transmission requires power-intensive signal conditioning hardware. Between racks and between rows of racks, copper's signal degradation at high data rates requires either active copper cables with equalisation electronics, which add cost and power consumption, or optical fibres with electrical-to-optical transceivers at each end, which add the cost of the transceiver modules but achieve the low-loss, bandwidth-dense optical transmission that copper cannot match at distance. The migration of AI computing to increasingly dense GPU clusters whose interconnect bandwidth requirements scale with the number of GPUs being coordinated has pushed the physical limits of copper interconnect to the point where optical interconnect is becoming economically competitive not just between racks but within racks and potentially within the server itself.
Silicon photonics is the technology platform that makes optical interconnect practical at the manufacturing scale that data centre volumes require. A photonic integrated circuit fabricated on a silicon wafer using processes derived from CMOS semiconductor manufacturing integrates the optical waveguides, modulators, photodetectors, and multiplexing components required for high-speed optical data transmission in a chip whose manufacturing cost and yield can approach those of silicon electronic integrated circuits at volume. The ability to fabricate silicon photonic components using existing semiconductor foundry infrastructure, without the specialised compound semiconductor manufacturing that III-V photonics requires, creates the cost trajectory that conventional photonics cannot match and that has driven the silicon photonics market's 45 percent compound annual growth rate from $95 million in 2023 toward the $863 million projection for 2029 that reflects AI data centre interconnect demand alone.
NVIDIA's $2 Billion Lumentum Investment
NVIDIA's announcement in March 2026 of a $2 billion investment in Lumentum, combining equity investment with multi-year laser component purchase commitments to scale silicon photonics for AI interconnects, is the most commercially significant single corporate investment in optical interconnect technology and the clearest signal that the dominant AI accelerator manufacturer has assessed silicon photonics as a structural necessity for the next generation of AI data centre architecture rather than an optional performance enhancement. NVIDIA's GPU-to-GPU interconnect architecture, whose NVLink fabric connects multiple GPUs within a server and whose InfiniBand and Ethernet connections span between servers and racks, faces the copper distance and bandwidth limitations whose resolution through optical interconnect at increasingly short distances is the technology shift that the NVIDIA-Lumentum investment is funding. The laser components that Lumentum manufactures are the light sources whose performance determines the power efficiency and output wavelength stability of silicon photonic transceivers, making Lumentum's pump laser and edge-emitting laser supply critical to the silicon photonic transceiver manufacturing whose scale NVIDIA's AI infrastructure demand requires.
Marvell Technology's release in March 2026 of its 1.6 terabit optical DSP platform for AI data centre interconnects, demonstrated at the OFC 2026 conference in Los Angeles alongside demonstrations from Cisco, Broadcom, and Coherent of next-generation co-packaged optics, represents the commercial product portfolio that the silicon photonics investment wave is producing in real deployable hardware. Co-packaged optics, which integrates the optical transceiver directly alongside the switching or computing silicon in the same package rather than in a pluggable module connected by copper traces on the circuit board, reduces the copper path length between the electrical and optical domains to millimetres, improving energy efficiency and signal integrity at the highest data rates where pluggable transceiver approaches encounter their physical limits. The industry-wide pivot toward co-packaged optics demonstrated at OFC 2026, with every major data networking vendor committing to CPO roadmaps, signals that silicon photonics has crossed the inflection from emerging technology to standard infrastructure component.
The Material Platform Competition
Silicon on insulator is the dominant silicon photonics material platform for data centre transceivers, but the material landscape includes silicon nitride, indium phosphide, thin-film lithium niobate, and barium titanate platforms whose different optical properties create performance advantages in specific applications that silicon-on-insulator cannot achieve. Thin-film lithium niobate modulators, whose electro-optic bandwidth and low insertion loss exceeds that of silicon's plasma dispersion modulators, are being integrated into hybrid silicon photonic chips whose lithium niobate modulator sections are bonded to silicon waveguide routing using wafer bonding techniques. Ayar Labs' optical I-O chiplets, which implement the optical transceiver function in a chiplet that is co-integrated with computing or switching chips using industry-standard chiplet interconnect interfaces, represent the most aggressive integration of silicon photonics with electronic computing that any commercial product currently demonstrates, achieving the photonic interconnect bandwidth density that future AI accelerator generations require from an optical interface whose electrical interface is compatible with the standard chiplet ecosystem.
Top 10 Companies in Silicon Photonics and Data Centre Optical Interconnect Globally
- Marvell Technology: US semiconductor company with 1.6T optical DSP platform and AI data centre interconnect solutions released in March 2026; its acquisition of Polariton Technologies for plasmonics-based modulators and its optical circuit switching collaboration with Lumentum create the silicon photonics chip company whose AI data centre networking products serve the hyperscale interconnect market at terabit rates.
- Lumentum: US photonics company with $2 billion NVIDIA investment for laser component scale-up and silicon photonics transceiver supply; its high-power pump lasers and edge-emitting lasers for silicon photonic transceiver integration and its coherent optics components create the laser supply chain that AI data centre optical interconnect scaling depends on.
- Coherent Corp.: US optical components company with silicon photonic transceiver modules launched at OFC 2026 for AI data centre applications; its co-packaged optics development and its indium phosphide laser integration into silicon photonic platforms create the optical component supply for the 800G and 1.6T interconnect standards that AI cluster networking is deploying.
- Intel: US semiconductor company with silicon photonics transceiver products for data centre optical interconnect; its Intel Optical I/O and its EMIB packaging technology for integrating optical components with silicon chips create the semiconductor company's silicon photonics commercial position whose foundry manufacturing scale differentiates it from smaller photonics specialists.
- Broadcom: US semiconductor company expanding co-packaged optics capability for AI switching infrastructure; its Tomahawk and Jericho switch silicon photonics integration and its 51.2T switch chip with integrated optical I/O create the network switching company whose co-packaged optics roadmap is the reference for hyperscale data centre spine fabric optical integration.
- STMicroelectronics: Swiss-French semiconductor company with the PIC100 data centre photonics chip developed with AWS for 800G and 1.6T optical interconnects; its established silicon foundry manufacturing and its AWS partnership create the semiconductor company's silicon photonics data centre product whose foundry manufacturing scale supports the cost trajectory that data centre volume requires.
- Ayar Labs: US silicon photonics company with optical I/O chiplets for co-integration with computing and switching chips; its TeraPHY optical I/O chiplet and its chiplet interconnect standard compatibility create the optical interconnect chiplet that disaggregated computing architectures use to achieve optical bandwidth density at the chip-to-chip level.
- Cisco: US networking company with 1.6T optical transceivers and 800G linear pluggables demonstrated at OFC 2026 built on silicon photonics from its Acacia and Luxtera acquisitions; its data centre networking market position and its optical transceiver production scale create the system integrator's silicon photonics commercial position in the hyperscale networking market.
- OpenLight: US silicon photonics company with $50 million Series A financing for standards-based 1.6T and 3.2T reference designs; its open photonic reference designs that enable optical transceiver manufacturers to implement next-generation products without developing the photonic platform from scratch create the silicon photonics IP licensing and reference design business.
- Lightmatter: US photonic computing company with Passage photonic interconnect and Envise photonic AI accelerator; its silicon photonics platform for both optical interconnect and photonic computing creates the company that is applying silicon photonics beyond transceiver applications to the photonic computing architecture whose optical matrix multiplication could accelerate AI inference beyond what electronic hardware can achieve.