AI Cluster Expansion Drives Explosive Growth in Optical Interconnect Demand

The quick and significant expansion of artificial intelligence clusters has notably moved the focus and attention from traditional compute hardware toward the intricate connections that enable thousands of accelerators to function as cohesive and unified systems. In its comprehensive research report published in April 2026, Goldman Sachs Global Technology Equity Research identified optical networking as the essential and critical enabler that unlocks higher effective computing power through the facilitation of low-latency, high-bandwidth data exchange. 

Analysts project that the total addressable market across both scale-up and scale-out configurations will experience substantial growth, expanding from approximately $15 billion associated with the GB300 NVL72 cycle in 2026 to an impressive $154 billion linked to the advanced Rubin Ultra NVL576 systems primarily anticipated in 2028. This remarkable growth represents a nearly ninefold increase, driven by the rising computing density per rack and the inherent physical limitations of copper interconnects at higher operational speeds. Optical networking is emerging as the defining and foundational infrastructure layer for the next generation of AI systems, with scale-up architectures and co-packaged optics poised to capture the majority of a projected $154 billion market by the year 2028, as the networking content per computing unit rises dramatically and significantly.

Goldman Sachs analysis shows scale-up networking, which connects larger numbers of GPUs within and across tightly coupled racks to form high-performance supernodes, is expected to account for 69 percent of the projected $154 billion total addressable market, or roughly $106 billion by 2028. This configuration prioritizes ultra-low latency and high bandwidth density inside the computing unit, moving beyond traditional scale-out approaches that rely primarily on switching between separate systems. The bank’s modeling assumes continued growth in rack-scale deployments, with NVIDIA-related systems progressing from GB300 NVL72 architectures toward denser Rubin Ultra NVL576 designs that require substantially more interconnect value. Aggregate networking dollar content per computing unit is projected to rise 29 times, from $315,000 in the GB300 NVL72 to $9.4 billion in the Rubin Ultra NVL576, reflecting both higher speeds and the shift toward optical solutions over shorter distances. 

Supporting data from company disclosures and shipment assumptions of approximately 48,000 racks for the earlier generation and 16,500 computing units for the later one underpin the ninefold overall TAM expansion. Industry observers note that scale-up demand aligns with hyperscaler requirements for coherent, high-throughput fabrics that minimize communication overhead during large-model training. Practical implications appear in the component mix, where copper cables, PCB midplanes, pluggable modules, and emerging co-packaged solutions all see rising content. Scale-up environments place particularly heavy emphasis on short-reach optics capable of sustaining multi-terabit aggregate bandwidth while controlling power and thermal loads. 

Goldman Sachs highlights that the addressable market for optical modules and engines expands thirteenfold when moving from pure scale-out to architectures that incorporate substantial scale-up elements. This structural shift creates multi-year visibility for suppliers positioned across the optical value chain, as operators prioritize interconnect performance to fully utilize advanced accelerators. Recent coverage of the report emphasizes that all major configuration types are expected to grow rather than one displacing another, supporting broad-based demand through the 2026-2028 period. Verification of these projections relies on the bank’s detailed product-cycle assumptions and publicly available server model data, ensuring the forecasts remain grounded in observable technology roadmaps.

Co-packaged optics technology, which places optical engines in close proximity to switch or accelerator silicon, is projected by Goldman Sachs to contribute approximately $91 billion, or 59 percent of the overall $154 billion TAM by 2028, assuming a 29 percent penetration rate in scale-out applications. By shortening electrical paths from centimeters to millimeters, CPO reduces power consumption and latency compared with traditional pluggable modules mounted at the front panel. Initial deployments focus on integration with switch ASICs, with subsequent expansion toward XPUs including GPUs and custom accelerators. 

The bank’s high-end estimates imply meaningful volume for scale-out CPO switches reaching tens of thousands of units by 2028 under accelerated adoption scenarios. Supporting analysis shows optical engines, external laser sources, fiber assemblies, and related components forming the bulk of this value. Industry commentary confirms that CPO addresses the density and power walls that emerge as lane speeds move beyond 200 Gbps and rack-level GPU counts increase. Adoption timelines remain subject to yield, interoperability, and field-replaceability considerations, yet the projected contribution underscores its strategic importance. Goldman Sachs modeling separates contributions from pluggable modules that continue to serve longer or more flexible links and CPO that targets the densest short-reach connections. 

Real-world progress includes supplier announcements of external laser source orders and roadmap alignments with major platform providers. Power-efficiency gains of CPO become increasingly relevant as data center operators confront rising energy costs and thermal constraints associated with multi-kilowatt racks. The technology’s role in unlocking higher effective utilization of expensive compute silicon reinforces its position within the broader optical networking expansion. Continued monitoring of penetration rates against the 29 percent scale-out assumption will determine whether the $91 billion figure materializes at the upper or lower end of the range.

Detailed Goldman Sachs calculations show the aggregate networking dollar content per computing unit climbing 29 times from $315,000 in GB300 NVL72 systems to $9.4 billion in Rubin Ultra NVL576 configurations. Breaking this down further reveals a 16-times increase in scale-out content and a 45-times increase in scale-up content on a per-unit basis. These multipliers arise from higher port counts, migration to 1.6T and eventually 3.2T speeds, greater optical penetration at shorter reaches, and the inclusion of advanced packaging elements such as midplanes and co-packaged engines. Equivalent 1.6T optical units per computing unit are expected to rise from hundreds in current generations to thousands in denser future systems. The bank’s assumptions incorporate realistic shipment volumes across full product cycles, translating unit-level content growth into the overall ninefold TAM expansion.

This content escalation has direct consequences for supplier revenue visibility and capacity planning. Components that previously represented modest fractions of system cost become material line items, elevating the strategic importance of optical and high-speed interconnect specialists. Analysis of related server models indicates that the transition is already underway with 800G and early 1.6T deployments, setting the stage for accelerated growth as Rubin-class systems ramp. Operators evaluating total cost of ownership increasingly factor interconnect performance and power into purchasing decisions, further supporting higher content. The projections remain tied to verifiable technology roadmaps rather than speculative demand scenarios, providing a solid foundation for assessing the multi-year opportunity.

As lane speeds and cluster sizes increase, copper interconnects encounter fundamental limits in reach, power, and signal integrity, prompting a progressive shift toward optical solutions even at shorter distances. Goldman Sachs documentation of mainstream AI server connection approaches shows copper remaining relevant for the most proximate links while optics expand into scale-up domains previously dominated by electrical signaling. Active copper cables and advanced PCB designs extend copper’s viability temporarily, yet the direction favors optics for bandwidth density and energy efficiency. Scale-out networks already rely heavily on pluggable optical modules, and scale-up architectures are incorporating optical engines and co-packaged designs at increasing rates. Industry data confirm that optical links are replacing copper at progressively shorter reaches as GPU counts move from thousands toward tens or hundreds of thousands in supernodes.

This transition carries practical engineering implications for data-center design, including structured cabling, liquid cooling integration, and power distribution upgrades that must accompany higher optical densities. Facility-level investments become part of the broader networking program rather than separate considerations. Supplier roadmaps reflect the dual requirement to support current pluggable volumes while preparing for denser optical integration. The net effect is a multi-year demand cycle spanning both traditional and emerging optical form factors. Verification through company disclosures and independent forecasts supports the conclusion that optical content will continue rising as a share of total interconnect spending.

The value market for pluggable optical modules in scale-out configurations is projected to expand tenfold on a per-computing-unit basis between second-half 2025 and second-half 2027 server models. Equivalent 1.6T unit counts are expected to climb sharply as speeds migrate and port densities increase. Silicon photonics adoption within the datacom market is forecast to rise from low single-digit percentages in early 2024 toward 45 percent by late 2028, offering potential bill-of-materials cost advantages approaching 32 percent and price advantages near 20 percent relative to traditional externally modulated laser solutions at 1.6T rates. These shifts support higher volumes while addressing cost and supply pressures associated with discrete laser approaches.

Pluggables and silicon photonics are expected to coexist with co-packaged optics rather than being fully displaced in the near term, serving different reach and flexibility requirements. Production capacity expansions announced by leading module and component suppliers align with this volume outlook. Cost and integration benefits of silicon photonics become particularly relevant as operators seek to control capital and operating expenditures amid rapid infrastructure buildouts. The combination of volume growth and technology migration underpins a substantial portion of the overall optical networking expansion through 2028.

Indium phosphide, the compound semiconductor essential for high-speed lasers used in optical transceivers and engines, faces tightening supply conditions that several executives describe as more acute than certain memory shortages. China accounts for roughly 70 percent of refined indium production, and export controls implemented on indium phosphide have slowed shipments and elevated prices. Substrate makers have implemented multiple price increases, with further hikes exceeding 10 percent under consideration. Leading device manufacturers report being sold out or undershipping demand by 25 to 30 percent or more despite capacity expansions. Long-term supply agreements and prepayments for capacity have become common as customers seek to secure material through 2028 and beyond.

These constraints affect the entire optical value chain, from substrates through epitaxial wafers and finished lasers to modules and co-packaged engines. Capacity additions require significant lead times for crystal growth, processing, and qualification, limiting the speed of response. Direct investments by major platform companies into key laser suppliers reflect recognition of the bottleneck’s strategic importance. Monitoring of export licensing, wafer pricing, and fab expansion timelines remains essential for assessing near-term supply availability against projected demand growth.

Companies specializing in lasers, optical engines, and networking systems have reported substantial order growth and backlog expansion tied to AI infrastructure. Lumentum has highlighted record shipments of high-speed lasers and capacity commitments, while Coherent has noted strong book-to-bill ratios extending into future years. Ciena has described optics as indispensable for next-generation AI architectures and projected meaningful expansion of its addressable market. These suppliers benefit from both current pluggable demand and longer-term co-packaged opportunities. Earnings growth and guidance increases in recent periods reflect the early stages of the content and volume expansion outlined by Goldman Sachs.

Differentiation arises from technology roadmaps, manufacturing scale, and customer qualification status. Suppliers with established positions in indium phosphide devices and high-volume module production hold advantages in the near term, while those advancing co-packaged and silicon photonics solutions prepare for subsequent phases. Ongoing capacity investments and strategic partnerships with platform providers further support visibility through the 2026-2028 horizon. The breadth of demand across scale-up and scale-out configurations provides a diversified opportunity set rather than dependence on a single architecture.

Speed transitions from 800G toward 1.6T in 2026 and subsequently toward 3.2T are expected to lengthen the overall migration cycle while elevating average selling prices and optical content. Goldman Sachs and related industry forecasts indicate substantial unit growth for higher-speed modules, with 1.6T and above categories showing particularly strong percentage increases. Silicon photonics and advanced laser approaches support these higher rates by improving integration density and power efficiency. The shift influences both pluggable and co-packaged form factors, requiring parallel advances in packaging, testing, and thermal management.

Operators standardize on multi-source qualified optics and laboratory validation of next-generation modules to manage risk during the transition. Facility upgrades for power, cooling, and cabling form an integral part of supporting higher-speed deployments. The combination of unit volume growth and speed-driven content increases contributes meaningfully to the overall TAM expansion. Continued tracking of shipment data against forecast models will confirm the pace of this structural change.

Major platform providers’ product cycles, including progressions from current NVIDIA generations toward denser Rubin-class systems, establish the concrete specifications that optical suppliers must meet. Connection solutions evolve across copper, PCB, pluggable optics, and co-packaged designs according to the specific scale-up and scale-out needs of each architecture. Goldman Sachs' mapping of mainstream AI server configurations illustrates the progressive incorporation of higher-speed optics and denser integration. These roadmaps create visibility for component demand while imposing rigorous qualification and volume requirements.

Alignment between platform roadmaps and supplier capacity plans is critical for timely delivery. Investments in external laser sources, optical engines, and related technologies by key suppliers reflect anticipation of these platform needs. The resulting demand is concentrated yet diversified across multiple hyperscalers and AI infrastructure builders, supporting a multi-year investment cycle. Practical examples include announced capacity expansions and long-term agreements that match expected platform ramps.

The projected market expansion creates differentiated opportunities at substrate, laser, module, and systems levels. Upstream material constraints elevate the strategic value of secure indium phosphide capacity, while midstream device and module producers benefit from volume and content growth. Downstream networking equipment providers capture value through integration and software-defined control of optical fabrics. Capital allocation toward capacity, technology development, and customer partnerships will determine relative positioning through 2028.

Risk factors include execution on yield improvements, geopolitical influences on material flows, and the precise timing of architecture transitions. Balanced assessment requires ongoing verification of shipment data, pricing trends, and backlog metrics against the original Goldman Sachs assumptions. The overall opportunity remains substantial for participants that successfully navigate these variables.

AI training and inference workloads generate continuous high-bandwidth traffic among large numbers of accelerators, placing stringent requirements on both latency and energy per bit. Optical solutions deliver superior performance on both metrics compared with copper at the speeds and distances required by modern clusters. Co-packaged approaches further improve efficiency by minimizing electrical path lengths. These technical advantages translate directly into higher cluster utilization and lower operating costs, supporting the economic case for higher optical content.

Data center operators increasingly evaluate interconnect choices through the lens of total system efficiency rather than component cost alone. The resulting preference for optical technologies across more layers of the network reinforces the multi-year demand trajectory. Engineering progress on laser efficiency, photonic integration, and packaging continues to widen the performance gap in favor of optics.

Optical networking forms one element within a larger set of AI infrastructure investments that also encompass power delivery, cooling, and compute silicon. The Goldman Sachs thesis positions networking as the next primary bottleneck after earlier constraints in accelerators and memory. Resolution of interconnect limitations enables more effective scaling of existing and future compute platforms. Industry-wide capital expenditure patterns already reflect rising allocations to networking and optical components.

The $154 billion projection by 2028 represents a concentrated opportunity within this broader spending cycle. Participants across the supply chain that align technology, capacity, and customer relationships stand to capture meaningful portions of the expansion. Continuous verification of underlying assumptions against real deployment data remains essential for accurate assessment.

1. How does the Goldman Sachs optical networking TAM projection compare with earlier AI infrastructure forecasts?

The bank’s estimate of growth from roughly $15 billion in 2026 to $154 billion by 2028 focuses specifically on scale-up and scale-out interconnect content rather than broader data-center spending. It incorporates detailed assumptions about NVIDIA product transitions and optical penetration rates that earlier general AI infrastructure analyses often treated at a higher level. The resulting ninefold expansion highlights the rising relative importance of networking within total system cost.

2. What role does co-packaged optics play relative to traditional pluggable modules? 

Co-packaged optics targets the densest, shortest-reach connections by integrating optical engines closer to silicon, delivering power and latency benefits that pluggables cannot match at the highest densities. Pluggable modules continue to serve longer reaches and applications requiring field replaceability. Goldman Sachs assumes both coexist, with CPO capturing a majority of incremental value under its 29 percent scale-out penetration scenario.

3. Why is indium phosphide supply particularly constrained? 

Indium phosphide is essential for the lasers that generate optical signals in high-speed modules and engines. Limited global substrate and device capacity, combined with export controls affecting a major producing region, have created shortages that lag demand by meaningful percentages despite expansion efforts. Qualification lead times and capital intensity further slow the response to rising AI-driven volumes.

4. Which companies stand to benefit most directly from the projected growth? 

Laser and optical engine specialists with established high-volume positions, including those expanding indium phosphide capacity and advancing co-packaged solutions, are positioned for the strongest near-term volume and content gains. Module manufacturers and networking systems providers that secure multi-year qualifications with hyperscalers also capture substantial portions of the expanding market.

5. How quickly are speeds expected to migrate beyond 800G? 

Industry roadmaps and Goldman Sachs analysis indicate 1.6T deployments ramping meaningfully in 2026, with further migration toward 3.2T in subsequent years. The pace depends on silicon readiness, optical component availability, and operator qualification cycles. Parallel progress on silicon photonics supports cost-effective higher-speed solutions.

6. What practical steps are data center operators taking to prepare for higher optical densities? 

Operators are upgrading power distribution, liquid cooling, structured cabling, and thermal management systems in parallel with optical deployments. Laboratory validation of next-generation modules and multi-source qualification processes help manage risk. Facility planning increasingly treats networking and supporting infrastructure as an integrated program.

7. Does the forecast assume continued dominance of any single platform provider? 

The modeling centers on progressive NVIDIA architectures because of their current market weight, yet the underlying drivers of higher interconnect content apply across multiple accelerator platforms and custom ASICs. Scale-up and scale-out requirements emerge wherever large clusters demand low-latency, high-bandwidth communication.

8. How should investors interpret supply-chain risks within the overall opportunity? 

Material constraints and capacity lead times introduce near-term volatility and potential delivery delays, yet they also support pricing power for qualified suppliers. Long-term agreements, capacity prepayments, and strategic investments by platform companies mitigate some risks while underscoring the structural nature of demand. Continuous monitoring of shipment, pricing, and backlog data remains necessary.

This content is for informational purposes only and does not constitute investment advice. Investments carry risk. Please do your own research (DYOR).