Feynman Architecture Ushers in Chip Optical Interconnect Era, CPO Industry Faces Value Reassessment

Inter-chip connectivity, as the core infrastructure for global AI computing power, is experiencing a historic shift from “electric” to “optical.” On March 16, local time, NVIDIA announced the release of the Feynman chip, which introduces optical communication for the first time into inter-chip connectivity, reducing AI data center communication energy consumption by over 70%.

Analysts believe that with the establishment of overseas technological pathways and increased domestic industrial policies, A-shares are deeply integrated into the global computing power supply chain.

It is worth noting that while industry giants are advancing rapidly, domestic policies are also injecting strong momentum into the development of China’s optical communication technology. The Ministry of Industry and Information Technology and the State Administration for Market Regulation issued the “Electronic Information Manufacturing Industry 2025-2026 Stable Growth Action Plan,” which clearly states, “Focus on key areas in the photonics field for technological breakthroughs, increase R&D investment in high-speed optical chips and optoelectronic co-packaging, and promote the integration of optical architectures with existing electrical architecture ecosystems.” This policy precisely targets the core aspects of the CPO technology route, guiding domestic companies in technological breakthroughs.

At the local level, actions are also frequent. In March this year, the General Office of the Guangdong Provincial Government issued the “Guangdong Province Accelerating the Cultivation and Development of New Tracks to Lead Modern Industrial System Construction (2026–2035),” which proposes to “strengthen R&D of key materials, equipment, and processes for optical chips in fields such as optical communication, optical sensing, and optical computing.” It aims to demonstrate and apply optical chip products in scenarios such as next-generation information communication, data centers, and intelligent computing centers.

Deep integration into the supply chain

As early as two weeks before the GTC conference, NVIDIA had already sent a strong “optical signal.”

In early March, NVIDIA announced on its official website that it had reached strategic agreements with Lumentum and Coherent, investing $2 billion in each company. This $4 billion total strategic investment was interpreted by the market as NVIDIA’s “positioning battle” for core upstream materials in the CPO industry chain.

In fact, leading listed companies in A-shares have already deeply embedded into NVIDIA’s supply chain system. According to the research team at China Merchants Securities, Zhongji Xuchuang has a first-mover advantage in 800G high-speed optical modules and has already begun layout for 1.6T products. Wang Yihong, chief analyst of communications at Tianfeng Securities, states that New Ease is accelerating shipments of silicon photonics products, with a significant increase expected in the proportion of silicon photonics by 2026; the company has successfully launched series products based on silicon photonics, including 400G, 800G, and 1.6T. As NVIDIA’s Rubin platform begins mass production in the second half of the year and the technical path of the Feynman architecture becomes clear, product iterations from 800G to 1.6T and even 3.2T are accelerating.

At the same time, Chinese optical module manufacturers are also active at the OFC 2026 exhibition and seminar. On-site, 1.6T optical modules, silicon photonic integration, and CPO technology solutions are focal points. Industry insiders believe that optical interconnects are becoming a key variable in determining the upper limit of AI infrastructure performance. Multiple technological trajectories—such as product upgrades from 800G to 1.6T, architecture evolution from pluggable to CPO, and short-distance replacements from copper cables to optical—are advancing simultaneously, jointly building the technical foundation for this round of optical interconnect evolution. As AI cluster scales continue to expand into the hundreds of thousands of GPUs, the optical module industry is poised for structural growth. Chinese manufacturers, with their strong optical technology accumulation and rapid iteration capabilities, are expected to continue increasing their market share in high-end optical modules.

Industry chain value re-evaluation

With NVIDIA officially introducing optical communication into inter-chip connectivity at GTC 2026, the industrial pathway for CPO technology has become clear. Analysts generally believe that this technological breakthrough not only marks a new cycle of architecture upgrades in AI computing infrastructure but also signals a reshaping of the value distribution in the optical communication industry chain. From upstream optical chips, midstream optical modules, to downstream data center applications, the large-scale commercial use of CPO is opening new growth opportunities.

Xiao Qunxi, chief analyst of the machinery industry at Guotai Haitong Securities, states that Rubin is no longer just a single GPU product but an integrated AI supercomputing platform composed of CPUs, GPUs, interconnects, networks, and system components. NVIDIA is elevating the delivery unit of AI infrastructure from individual cards to entire cabinet systems. To achieve such high-density interconnects, Rubin may adopt a dual-layer network topology and realize “optical into copper” within the cabinet.

Xiao further notes that at the interconnect level, CPO and silicon photonics are becoming key directions for ultra-large-scale AI systems. In the future, data centers will gradually shift from traditional copper interconnects to higher bandwidth, lower-loss optical connection systems. Regarding cooling, air cooling is losing its suitability for ultra-high-power computing platforms, and liquid cooling will increasingly become a standard rather than an option.

Cheng Qiang, director of Debon Securities Research Institute and chief economist, states that from industry development trends, CPO is accelerating from the verification stage toward early commercialization. Global foundries are speeding up silicon photonics chip manufacturing capacity expansion, with Tower Semiconductor doubling its silicon photonics manufacturing capacity, and GlobalFoundries acquiring a Singaporean silicon photonics wafer foundry. UMC is collaborating with IMEC to achieve CPO-compatible silicon photonics processes. Domestically, manufacturers like Yandong Micro and Saiwei Electronics are also continuously advancing process development and wafer manufacturing layouts for silicon photonics chips.

Ouzi Xing, a telecommunications industry analyst at Dongwu Securities, believes that future optical interconnects will be driven by multiple network connection scenarios, and the overall market space is expected to maintain rapid growth. Different technological routes will not completely replace each other but will form differentiated positioning based on technical characteristics, cost structures, and application scenarios.

Ouzi further suggests that, considering market cycles and industry evolution stages, key focus should be on three main lines: first, core optical modules, where industry leaders will benefit from trends toward higher speeds and larger bandwidths, with both growth potential and certainty; second, secondary optical modules, which will benefit as downstream demand spreads from top-tier clients to other segments; and third, emerging optical interconnects, where new technologies like CPO and silicon photonics are in the early stages of industrialization, and early positioning now could enjoy incremental benefits from technological iteration.