The Zhitong Finance App learned that Bernstein pointed out in the latest research report released on November 10 that as Nvidia (NVDA.US)'s next-generation AI computing platform Rubin Ultra accelerates the adoption of co-packaged optics (CPO) technology, the demand for optical communication equipment in artificial intelligence data centers is expected to increase dramatically. The bank expects the Nvidia ecosystem's optical communications market to reach US$46.8 billion by 2028, an increase of about 91% over US$24.5 billion in 2027. Among them, Rubin Ultra will be an important driving force for market expansion.
Bernstein analyst Daniel Zhu said that as AI computing clusters continue to expand, traditional copper interconnections gradually face transmission distance and bandwidth limitations, and the application of optical communication technology within data centers is expected to increase significantly. The CPO technology adopted by Nvidia on the Rubin Ultra platform will create new growth opportunities for the optical communications industry chain.
According to the technical architecture previously announced by Nvidia, Rubin Ultra will use a computing cluster composed of 8 cabinets, each with 72 GPUs, for a total of 576 GPUs. Bernstein estimates that the total optical communication bandwidth of this cluster will reach about 15.9 Pb/s, an increase of about 7.7 times over the Vera Rubin platform's configuration of about 2.1 Pb/s on the same scale. This change means that as high-speed interconnections between GPUs gradually shift from copper cables to optical connections, the value of optical components required for a single AI computing cluster will increase significantly.
Optical communication equipment manufacturer Coherent (COHR.US) previously disclosed that for chips with a data transmission capacity of 100 Tbps, the value of optical components may reach 15,000 US dollars. Based on this calculation, Bernstein estimates that the value of the optical components of the Rubin Ultra's single 8-cabinet cluster is expected to reach about US$5.35 million, compared to the previous Vera Rubin platform's optical component value of about US$1 million.
Among them, the CPO optical components of a single cluster of Rubin Ultra are worth about $1.04 million, pluggable optical components to expand GPU interconnections are worth about $3.32 million, and the remaining $1 million comes from other optical connectivity requirements.
According to Bernstein's semiconductor research team, shipments of Rubin Ultra related R300 chips will reach about 1 million units in 2027, and further increase to about 3.9 million in 2028. Based on the above shipments and the value of optical components in a single computing cluster, the bank anticipates that Rubin Ultra will create an optical communication market opportunity of approximately US$9.4 billion in 2027 and further expand to approximately US$36.1 billion in 2028.
If Nvidia's other AI computing platforms and future Feynman architectures are included in the calculation, Bernstein predicts that the optical communication market for Nvidia's entire ecosystem will grow from US$19.5 billion in 2026 to US$24.5 billion in 2027, and reach US$46.8 billion in 2028.
In addition to optical transceiver equipment, fiber suppliers are also expected to benefit. Bernstein pointed out that Corning (GLW.US) previously disclosed that under an architecture that uses copper cables for internal GPU expansion and interconnection, each GPU corresponds to the requirement of about 16 optical fibers. However, according to the bank's estimates of the Vera Rubin architecture, the actual demand may reach about 36 fibers per GPU.
As Rubin Ultra further uses optical interconnects, each GPU may require an additional 240 fibers, bringing the total requirement to about 276, which is nearly 8 times that of the Vera Rubin architecture.
Although a sharp increase in the number of optical fibers does not mean that related revenue will increase in the same proportion, Bernstein believes that the value of Corning's optical communication products corresponding to each GPU will increase by about 50% is a reasonable estimate. If optical interconnections are also fully used inside cabinets in the future, the room for long-term growth may expand further.
However, there is still uncertainty about whether the optical communications industry chain can meet rapidly growing demand.
Bernstein pointed out that indium phosphide substrates may become a major bottleneck limiting the supply of high-end optical communication products. Indium phosphide is an important material for manufacturing high-performance optical communication lasers. As demand for high-speed optical interconnection in AI data centers surges, it may be difficult for related production capacity expansion to keep up with market growth.
According to the bank's estimates, the production capacity of the indium phosphide substrate industry is expected to increase by about 60% in 2027 and further increase by about 67% in 2028. Although major supplier AXT (AXTI.US) is actively expanding production, companies such as Sumitomo are also planning to increase production capacity. Considering that additional substrate production capacity usually takes 9 to 12 months to be converted into end customer supply, the tight supply situation in the industry is likely to continue until 2028.
Bernstein believes that unless overall demand in the AI industry chain slows down significantly, the prices of key materials for optical communication are unlikely to drop sharply.
Judging from the competitive landscape of the industry, Nvidia is clearly more active than most large cloud computing companies in terms of the CPO technology adoption schedule. However, other tech giants are also promoting related technology. According to the research report, AWS, a subsidiary of Amazon (AMZN.US), allegedly plans to use CPO technology in Trainium 4 chips in 2027, and Google (GOOGL.US) is also reported to be purchasing a large number of new optical interconnect modules for its TPU computing platform.
This means that the future growth of the optical communications market may not only come from Nvidia, but also gradually expand to other large-scale AI computing platforms.
At the same time, competition among domestic manufacturers is also receiving increasing attention from investors. Bernstein pointed out that companies such as Yuanjie Technology are making progress in the field of high-end optical communication lasers and are gradually becoming competitors that international manufacturers cannot ignore.
However, the bank believes that Lumentum (LITE.US) and Coherent currently still have technical advantages in the field of high-end lasers. Although domestic manufacturers may expand their share in some low-end product markets, demand for high-end optical communication products continues to grow, which is expected to support the business performance of these two companies.
In terms of specific investment proposals, Bernstein maintains Coherent's “outperforming the market” rating and a target price of $350, believing that it is expected to benefit from increased demand for AI optical communication and may gain more market share in the process of industry consolidation.
For Lumentum, the bank also maintained a “outperforming market” rating and raised the target price from $635 to $700, which mainly reflects the latest stock valuation involved in the MRVL.US acquisition exchange.
Furthermore, Bernstein maintains Ciena (CIEN.US)'s “outperforming the market” rating and a target price of $440, and believes that its cross-data center interconnection business is expected to benefit from increased demand for electricity and land resources from AI computing clusters. For Corning, the bank maintains a “same as the market” rating and a target price of $120, believing that the market already reflects the future growth potential of the optical communications business to a large extent.
Overall, Bernstein continues to be optimistic about the long-term prospects of the AI optical communication industry chain. The bank believes that Rubin Ultra's adoption of CPO technology will significantly increase demand for optical components in AI computing clusters, and that tight supply of key materials may further strengthen the industry's supply and demand pattern. However, the progress of technology deployment, supply chain expansion, and changes in industry competition will still be important factors in determining the future growth performance of related enterprises.