◉
Silicon photonics
Waveguides, modulators, detectors and multiplexers fabricated with semiconductor-compatible processes. It is the integration platform underlying many optical engines.
Dense WDMHigh bandwidth density
Interactive architecture
Photonics enters at every distance scale.
The best interconnect depends on reach. Copper remains extremely effective across millimeters, but its power and signal-integrity penalties become harder to manage as bandwidth and distance rise.
Optical I/O at the package edge
Photonic chiplets or optical interposers shorten the high-speed electrical path and move conversion closer to GPUs, TPUs and custom accelerators. The goal is to increase bandwidth density while reducing the energy cost of chip-to-chip communication.
Optical I/OPhotonic chipletsOptical interposers
Commercial maturityEarly deployment
Potential architectural impactVery high
GPU / TPU / XPU compute package
Optical I/O or co-packaged optical engine
Fiber links across board, rack or cluster
Reconfigurable optical fabric
Core technologies
Three building blocks are reshaping the network.
⬡
Co-packaged optics
Optical engines move from the faceplate toward the switch ASIC or accelerator package, shortening power-hungry electrical traces and increasing front-panel bandwidth density.
CPOShort electrical reach
⇄
Optical circuit switching
OCS establishes direct light paths through a physical-layer fabric. Software can reconfigure topology around traffic, upgrades or failures without packet-by-packet optical processing.
OCSProgrammable topology
| Technology | Best fit | Main advantage | Primary challenge | Readiness |
|---|---|---|---|---|
| Pluggable optical transceivers | Rack-to-rack and data-center links | Mature, modular and serviceable | Faceplate density and retiming power | Mature |
| Co-packaged optics | High-radix switches | Shorter electrical path, higher density | Thermals, fiber attach and serviceability | Early adoption |
| Optical I/O chiplets | Accelerator-to-accelerator scale-up | Bandwidth density and reach beyond copper | Packaging, standards and total cost | Pre-volume |
| MEMS / LCOS OCS | Spine and cluster reconfiguration | Low steady-state switching power | Reconfiguration, control and topology design | Deployed selectively |
| Integrated silicon-photonic OCS | Fast rack or scale-up fabrics | Potentially faster, compact switching | Loss, radix, packaging and yield | Emerging |
Industry adoption
Deployment is real—but uneven.
Pluggable optics are ubiquitous. Optical circuit switching has proven itself in hyperscale networks. CPO and in-package optical I/O are moving from demonstrations toward early commercial systems, with economics and maintainability still decisive.
Indicative maturity
Illustrative editorial scoring, not an industry standard or market forecast.
What drives adoption?
System implications
The payoff is bigger than a faster cable.
01
More useful compute
Lower communication overhead can keep expensive accelerators productive instead of waiting on data movement.
02
Larger scale-up domains
Optical reach can expand tightly coupled accelerator fabrics beyond traditional copper boundaries.
03
Adaptive topology
OCS can create workload-specific light paths for collective communication, mixture-of-experts routing and bursty traffic.
04
Resilient growth
Programmable optical layers can support incremental upgrades, traffic engineering and physical-layer rerouting.
Open questions
What still has to be solved?
Packaging and thermals
Optical engines must coexist with very hot switch and accelerator silicon without compromising laser stability, coupling efficiency or yield.
Serviceability
Replacing a faceplate module is simple. Repairing optics integrated next to a high-value ASIC changes spares, maintenance and failure-domain economics.
Control software
Dynamic optical fabrics require telemetry, scheduling and orchestration that understand AI collective communication and topology changes.
Standards and interoperability
Chiplets, lasers, fibers, packaging and control planes need common interfaces to avoid proprietary islands and supply-chain risk.
Loss and scale
Integrated optical switches must reduce insertion loss while increasing port count and maintaining manufacturable fiber coupling.
Total cost of ownership
Energy savings matter, but adoption depends on complete lifecycle economics: capital cost, reliability, repair, utilization and deployment speed.
Sources and further reading
Primary references.
- Torrijos-Morán, L. & Pérez-López, D., “Industry insight: photonics to scale AI data centers,” npj Nanophotonics 3, Article 8 (2026). DOI source.
- Google Research, “Jupiter Evolving: Transforming Google’s Datacenter Network via Optical Circuit Switches and Software-Defined Networking.” Research paper.
- Google Cloud, “Speed, scale, reliability: 25 years of data center networking” (2024), reporting Jupiter scale up to 13 Pb/s. Official overview.
- NVIDIA, “Spectrum-X Photonics Co-Packaged Optics Networking Switches” (2025). Press release.
- Broadcom, “Third-Generation Co-Packaged Optics Technology with 200G/lane Capability” (2025). Press release.
- Optica Publishing Group, “High-Density Optical I/O for ML/AI Applications” (OFC 2024). Abstract.
Vendor performance figures are identified as company claims and should be evaluated against system configuration, workload and deployment conditions. Editorial maturity scores on this page are explanatory, not investment recommendations.