Enabling network scalability through advanced photonic integration

AI is reshaping optical networks from the rack to the data centre and beyond. Advances in photonic integration, vertical integration and heterogeneous PICs are enabling the bandwidth, power efficiency and scalability needed for the next generation of AI infrastructure.

By Fady Masoud, Optical Pluggables Solutions; Ajay Mistry, Leader, Silicon Photonics Design; Teresa Monteiro, Optical Components Solutions; Nokia

The rapid growth of AI, cloud computing, hyperscale data centers and global networks is creating unprecedented demand for bandwidth. Networks must now scale compute inside racks, across data-center fabrics and over metro, long-haul and subsea links. These three models scale-up, scale-out and scale-across are reshaping how photonic integrated circuits (PICs) are designed for higher bandwidth, lower power, smaller footprints and faster deployment (Figure 1).

Figure 1: Network scaling models

Scale-up increases compute capacity within a rack by connecting accelerators such as GPUs and TPUs, memory and switching resources with ultra-high bandwidth and very-low latency. These architectures require dense optical links that can deliver terabit-class throughput while minimizing power and latency.

Scale-out expands compute capacity horizontally across racks and nearby data centers. It places heavy demands on pluggable optics, coherent interconnects and switching fabrics as AI and cloud workloads distribute traffic across larger pools of machines.

Scale-across extends GPU or TPU clusters beyond the physical walls of a single building, in effect fusing two or three data centres and their accelerated compute resources into one giant synchronous super-cluster. Scale-across is most applicable to hyperscalers training large language models, but the scale involved, connecting 100,000s of GPUs in one facility to 100,000s in a second or third site 100s of kilometres or more away.

Although each model has different reach, latency and capacity requirements, they all share the same basic need: higher bandwidth, lower power per bit, smaller form factors, lower cost, and faster deployment. Coherent optics, used for DCI, metro, long-haul and subsea transport, prioritizes reach, spectral efficiency, optical signal quality and tight integration with DSP-based signal processing. Conversely, intensity modulation direct-detect (IMDD) optics are used for short-reach intra-data-center connectivity, prioritizes low power, rack density, low latency, cost and high-volume manufacturability. This is why photonic integration and vertical integration are becoming central to the AI networking era.

Figure 2: Vertical integration

The value of vertical integration
Deep vertical integration is a critical asset for advanced optical networking. Vertical integration refers to designing and controlling multiple layers of the technology stack, from photonic devices and DSPs to packaging, subsystem assembly and final optical system design. This includes in-house development of lasers, modulators, waveguides, photodetectors, optical amplifiers, RF drivers and Transimpedance Amplifiers (TIA), RF packaging, DSPs, controls, firmware, coherent engines, pluggable optics, optical transport, IP switching and routing platforms. Vertically integrated companies gain several important advantages.

First, vertical integration accelerates innovation. As baud rates increase and system architectures become more complex, performance optimization increasingly depends on co-design between optics, electronics, packaging and firmware. When all these domains are developed internally, engineering teams can optimize the entire system holistically rather than treating components independently.

Second, vertical integration improves quality and reliability. High-speed coherent systems require extremely precise alignment between optical and electrical subsystems. Integrated development and testing reduce interface mismatches and manufacturing inconsistencies, resulting in higher yields and improved field reliability.

Third, vertical integration enhances supply chain security. Demand for optical components is increasing dramatically in the AI Supercycle, placing stress on global manufacturing ecosystems. Companies with in-house manufacturing, packaging and testing capabilities are better positioned to maintain product availability and protect intellectual property.

Finally, vertical integration enables faster product introduction cycles. AI and cloud operators expect rapid deployment of new technologies across both coherent transport and intra-data-center connectivity.

Owning multiple stages of the value chain allows vertically integrated companies to respond faster to technology changes while optimizing performance, power and manufacturability (Figure 2).

At the foundation of vertical integration in high-performance optics is the photonic integrated circuit (PIC).

Figure 3: Monolithic InP coherent PIC

Building photonic integrated circuits
PICs generate and manipulate light to perform transmission, modulation, amplification, switching and detection functions on a compact microchip. They can integrate optical functions such as distributed-feedback (DFB) lasers or laser arrays, tunable lasers, Mach-Zehnder modulators (MZMs), electro-absorption modulators (EAMs), photodiode arrays, and multiplexers or demultiplexers into a single chip. This enables high-speed transmission with low power and simpler integration. Today, the two dominant PIC platforms are indium phosphide (InP) and silicon photonics (SiPh), while heterogeneous integration is increasingly used to combine the strengths of multiple material systems.

Silicon photonics
Silicon photonics uses mature semiconductor manufacturing to build optical circuits on silicon wafers. Its main advantage is scale: it can take advantage of established silicon foundries, packaging infrastructure and CMOS process control. Silicon photonics is well suited for passive optical functions such as waveguides, multiplexers, polarization management, filters and routing structures. It also supports advanced electronics integration and packaging, making it attractive for high-volume data-center applications. Modern SiPh process design kits (PDKs) include these functions, along with germanium-on-silicon photodetectors, edge couplers, tunable optical filters, optical attenuators, monitoring photodiodes and polarization-management functions. Integrating these building blocks on one PIC reduces packaging complexity and optical loss while improving transceiver performance, yield and cost.

Heterogeneous integration extends silicon photonics beyond the limits of native silicon. For example, thin film lithium niobate (TFLN) can be bonded to silicon to create very high-bandwidth modulators with lower drive voltages. This supports higher baud rates and better modulation efficiency than conventional silicon photonics. One example is the use of TFLN-based Mach-Zehnder modulators (MZMs) on a silicon photonics platform for high-performance coherent pluggable optics. These systems combine high-speed modulators, SiPh PDK components and high-speed photodetectors, enabling integrated transmit and receive functionality on a common platform.

Looking ahead, heterogeneous integration can combine InP and TFLN devices on a common silicon photonics platform. Because silicon cannot natively generate light, laser sources must be integrated through flip-chip assembly or wafer bonding.¹ In this architecture, InP devices enables active functions such as nano-integrable tunable laser assemblies (nITLAs), distributed-feedback (DFB) lasers and semiconductor optical amplifiers (SOAs).

By combining the strengths of each material system (InP for light generation and optical gain, TFLN for high-speed modulation and silicon for large-scale photonic integration), future PICs can deliver higher functionality and performance on a scalable platform. This approach also supports wafer-scale production and provides a path to future 1.6T, 3.2T and higher-capacity coherent optical modules.

Figure 4: Scalability and sustainability of power and capacity of InP PIC optical engines

Monolithic indium phosphide PICs
While silicon photonics relies on heterogeneous integration for adding functionality or improving performance, indium phosphide itself offers a capable and established monolithic integration platform. InP can generate, modulate, amplify, and detect light directly on the same substrate. This enables tunable lasers, semiconductor optical amplifiers, modulators, detectors and monitoring functions on one chip, reducing optical interfaces, insertion loss, footprint, packaging complexity and power consumption.

InP also allows optical gain to be placed where it delivers the greatest system benefit. Integrated SOAs can improve optical signal-to-noise ratio, increase launch power and enhance receiver sensitivity, enabling high-performance coherent transmission over long distances (Figure 3). By integrating lasers, amplifiers, modulators and detectors on one semiconductor platform, InP PICs support stable optical performance, simpler packaging and long operating lifetimes by minimizing mechanical bonds that often become failure points.

Monolithic InP solutions are especially attractive for coherent pluggables and long-haul transport systems. Recent demonstrations have shown 800 G coherent pluggables operating over uncompensated subsea fiber links spanning thousands of kilometers. These trials demonstrated the feasibility of using compact pluggable optics in applications traditionally dominated by large optical transport platforms.

InP is also compelling for intra-datacenter connectivity. Its monolithic integration can reduce modulation voltages and support power-efficient linear pluggable optics, helping address rack power density, cooling constraints and sustainability goals. Power-efficient intra-datacenter InP transmitter PICs were recently reported.^2,3 When packaged in the linear pluggable optics (LPO) form factor, these InP PICs can be used to interconnect GPUs through switches and network interface cards with about 8 W of power dissipation per 1.6 Terabits/sec.

The sustainability of optical engines for power, space and throughput is an increasingly critical factor and requires a wide range of technical innovations to meet the application demands. Figure 4 illustrates an example for long-reach optical engines realized with monolithic InP PICs over multiple decades and the sustained long-term reduction in power per bit enabled by successive generations of optical engines.

Challenges and future directions
Despite major progress, the next generation of photonic integration will require innovation across devices, materials, electronics, packaging and system architecture. As baud rates increase, electrical interconnect loss, signal integrity and thermal density become more difficult to manage. Future systems will need shorter RF paths, lower-loss materials, tighter DSP-to-optics integration and advanced thermal designs suitable for compact pluggable modules.

Manufacturing scalability is equally important. High-volume optical platforms must maintain cost efficiency, reliability and stable performance while supporting increasingly complex assemblies.

This will require better packaging materials, improved degradation models, scalable characterization methods and simulation tools that model optical, thermal and electrical behavior together.

Photonic integration has already transformed optical networking by enabling bandwidth scaling from 10 Gbps to multi-terabit systems.

The AI era now requires deeper co-design across photonics, electronics, materials, packaging and systems. The most scalable solutions will combine monolithic integration, heterogeneous integration, advanced packaging and vertical integration to deliver higher capacity, lower power and faster deployment.