Goldman Sachs projects the market for AI-related optical networking will surge by 900% from approximately $15 billion today to $154 billion by 2028, identifying a new primary investment thesis in AI infrastructure. The forecast, spurred by strong earnings from component maker Zhongji Innolight, argues that the architecture of AI data centers is creating massive demand for high-speed optical connections inside server racks, not just between them.
"The core logic is the evolution of data center architecture from horizontal to vertical, which brings higher bandwidth and more connection demand," Allen Chang, Goldman's Head of APAC Technology, wrote in the report. This shift means optical interconnects are no longer just a supporting part of AI expansion but are becoming an independent, quantifiable investment theme, according to the bank.
The main driver for this 10-fold market expansion is the need for "scale up" networking—interconnecting GPUs within a single server rack or super-node. Goldman estimates this internal networking will constitute about $106 billion, or 69%, of the total addressable market. For instance, the networking value for a single Nvidia GB300 NVL72 unit is $315,000, a figure projected to jump 29-fold to $9.4 million for the Rubin Ultra NVL576 system expected in 2027-2028.
For investors, the report reframes the AI hardware opportunity, suggesting a significant growth runway for component suppliers beyond just GPU manufacturers. The analysis points to a rally in optical communication stocks like Lumentum, Coherent, and Zhongji Innolight, whose stock rose 4% to a new record high after reporting a 262% jump in quarterly net profit.
CPO and Pluggables to Coexist
A key technology enabling this growth is Co-packaged Optics (CPO), which moves the optical engine next to the chip for lower latency and power consumption. Goldman projects CPO will contribute $91 billion to the market by 2028. However, its adoption is not a zero-sum game against traditional pluggable optical modules.
CPO's main drawback is its high maintenance cost; a failure could require replacing the entire switch ASIC. Therefore, Goldman expects CPO to be used in performance-critical scenarios, while pluggable modules will remain in use where operational flexibility is key. Even with CPO adoption reaching an estimated 29% in "scale out" applications by 2028, the absolute market for pluggable modules is still forecast to grow 10 times as the total number of GPU clusters expands.
Supply Bottlenecks and Silicon Photonics
The industry's most immediate challenge is a supply bottleneck for Indium Phosphide (InP), the core material for high-speed lasers. Demand is being pressured by the growth of AI servers, the upgrade to 1.6T and 3.2T speeds, and new laser demand from CPO. While suppliers like Lumentum and Coherent are expanding capacity, Goldman expects supply to remain tight until 2027, with a potential balance only in late 2028, barring further geopolitical trade restrictions on InP.
Concurrently, Silicon Photonics (SiPh) is rapidly gaining market share from traditional laser technology. A SiPh-based 1.6T module has a 32% lower bill-of-materials cost than the alternative, boosting supplier gross margins. The report sees SiPh penetration growing from 6% in early 2024 to 46% by the end of 2028.
The OCS Wildcard
Optical Circuit Switching (OCS) is an emerging technology whose primary advantage is "cross-generation compatibility," allowing a single switch to handle 800G, 1.6T, and 3.2T signals without replacement. This value has been validated by Google's adoption in its TPU v7 supercomputers and a reported $400 million order backlog for Lumentum's OCS systems. The ultimate market penetration for OCS, however, remains uncertain and will depend on how much of a premium customers will pay for its future-proofing capabilities.
This article is for informational purposes only and does not constitute investment advice.