The Zhitong Finance App learned that Orient Securities released a research report saying that OCS will become a common optical interconnection infrastructure in AI computing power clusters. The number of customers and expansion of application scenarios is expected to open up a new round of growth space, and the future market is expected to exceed expectations. The OCS industry has changed dramatically, and I am optimistic that the entry of major companies such as Nvidia and OpenAI will bring about a significant increase in industry sentiment.
Orient Securities's main views are as follows:
From fixed connections to dynamic reconstruction, resource pooling and fault recovery drive OCS penetration rate improvement
AI cluster networks are evolving from simply pursuing connection bandwidth to taking into account computing power utilization, flexible scheduling, and reliability, and the need for elastic and reconfigurable clusters is expected to become a core factor driving the increase in OCS penetration rate. Since traditional fixed physical topology networks are difficult to flexibly adjust connection relationships according to computational tasks, resource allocation, and fault conditions, it is easy to cause fragmentation of computing power resources and idle local resources. However, through dynamic configuration of physical optical circuits, OCS realizes network topology reconstruction and flexible networking of computing power resources while retaining packet switching capabilities, which is expected to greatly improve the overall utilization rate of the cluster and optimize TCO. By giving the AI computing network the ability to reconfigure as needed, OCS is expected to push it from specific scenarios to a wider range of application scenarios and platforms.
OCS moves from Google TPU to GPuscale-up, and market space expands at an accelerated pace
In the past, OCS was mainly used in Google TPU connectivity and Jupiter data center networks, and market demand was relatively concentrated. As manufacturers such as Nvidia promote the evolution of GPuscale-Up from a single cabinet to multiple cabinets, the bandwidth and scale of optical connectivity across cabinets continue to increase, which is expected to create a new incremental market for OCS. According to Cignal AI's forecast, the global OCS market is expected to exceed 8 billion US dollars in 2030, and the compound growth rate will exceed 40% from 2025 to 2030. Among them, GPusScale Up is expected to contribute significantly from 2028. OCS is expanding from a Google-led application system to a wider GPU computing network. The number of customers and expansion of application scenarios is expected to open up a new round of growth space and become a common optical interconnection infrastructure in AI computing power clusters.
Multiple technology routes evolve in parallel, and the whole machine and core optical devices are expected to benefit first
Currently, OCS technology routes are showing diversified development trends, and technologies such as MEMS, LCD, piezoelectric, and optical waveguide each have advantages. Among them, MEMS occupies an important position in current large-scale applications with large port expansion capabilities and industrialization foundation; optical waveguides have development potential in small to medium port and high density scenarios with high integration, potential large-scale manufacturing, and rapid reconstruction capabilities, but there is still a need to continuously improve insertion loss, port expansion, and manufacturing yield. With the release of OCS demand, industrial chain links such as complete machine systems, MEMS micromirrors, collimator arrays, photonic chips, and FAU are expected to benefit. The bank believes that enterprises with core process barriers, system integration capabilities, and large-scale delivery capabilities are expected to be the first to achieve performance.
Risk Alerts
Demand fell short of expectations due to falling OCS port ratios, delays in the introduction of GpuScale-Up to OCS, technology routes and competitive landscape risks, and AI capital expenditure falling short of expectations.