Global Optical Chip Expansion Driven by AI Data Center Demand

It has to be said that the demand for optical chips is extremely high.

In recent days, the global optical chip industry chain has seen a series of concentrated actions involving capacity expansion, long-term agreements, investments, and supply chain commitments: Coherent is expanding its 6-inch InP compound semiconductor production line in Sherman, Texas; Nokia is increasing its advanced testing and packaging capacity for photonic chips in Allentown, Pennsylvania; Japan’s JX Advanced Metals plans to invest up to 120 billion yen to increase InP substrate capacity by 7 to 10 times; IQE has entered into a multi-year agreement with Tower Semiconductor for InP epitaxial wafer supply; and China’s Sorsun Photonics, a subsidiary of Dongshan Precision, has announced an expansion project in Changzhou for optical chips and high-speed optical modules, with a total investment of $1.2 billion.

A production race centered on the optical interconnection capabilities of AI data centers has already begun.

First, let's look at the U.S. expansion efforts.

On June 16, Coherent announced that it has signed a letter of intent to receive up to $50 million in direct funding from the U.S. Department of Commerce under the CHIPS and Science Act to expand its world-leading 6-inch indium phosphide (InP) semiconductor manufacturing facility in Sherman, Texas. The following day, Coherent held a groundbreaking ceremony for the expansion at its Sherman, Texas facility. Coherent emphasized that this site hosts the world’s first and currently largest 6-inch InP manufacturing platform. Upon completion of the expansion, the facility’s manufacturing space will double, and wafer production capacity will increase fourfold.

Notably, NVIDIA’s founder and CEO, Jensen Huang, attended the ceremony in person alongside Coherent’s new CEO, Jim Anderson. NVIDIA had previously announced a strategic investment of $2 billion in Coherent to secure future capacity for its most advanced lasers, photonic engines, and photonic modules. Huang spoke on-site: “AI runs on compute, but scaling is constrained by connectivity—and the Sherman facility is where these ‘neural connections’ are built.”

Image source: techpowerup

Nvidia has integrated "light" into its AI infrastructure supply chain through capital investments. As early as March this year, Nvidia announced investments of $2 billion each in Coherent and Lumentum, along with multi-year procurement commitments and future capacity/access rights for advanced lasers, photonic networking products, R&D, and expansion of U.S. manufacturing capabilities.

Lumentum is also an essential part of the U.S. photonic chip expansion landscape. In March, Lumentum announced plans to build a new advanced laser manufacturing facility in Greensboro, North Carolina. The facility, covering approximately 240,000 square feet, will focus on producing indium phosphide (InP) photonic devices for large AI data centers worldwide. In May, AIXTRON announced it had received multiple G10-AsP MOCVD system orders from Lumentum. Over the past year, Lumentum’s stock price has risen by 769%.

On June 16, Nokia also announced plans to expand its advanced photonic chip testing and packaging capabilities in Allentown, Pennsylvania, USA, by further packaging photonic chips into optical modules for use in AI and communications infrastructure. Nokia stated that this facility is one of the few in the U.S. with such capabilities; after expansion, production capacity could increase up to ten times the current level, with commercial availability expected by the end of Q3 2026.

Nokia contributes photonic chip packaging, testing, and modular capabilities; Coherent brings InP photonic device front-end manufacturing capabilities; NVIDIA’s prior investments in Coherent and Lumentum effectively secure funding, orders, and capacity from core suppliers of lasers and optical networking. The United States is integrating optical interconnects for AI data centers into its domestic semiconductor manufacturing ecosystem.

Japan specializes in upstream material sectors, which have long been its strength in the semiconductor industry.

On June 16, JX Advanced Metals, one of the world’s two leading InP substrate manufacturers, announced plans to invest up to 120 billion yen over the next four years to expand its InP substrate production capacity. Combined with previously announced investments, the company’s total investment in InP capacity expansion will reach approximately 150 billion yen. These investments will increase the company’s production capacity to seven to ten times its current level.

JX Advanced Metals has been producing indium phosphide substrates since the 1980s. In fiscal year 2025, the company invested 25 billion yen to increase capacity for this material. According to India Strait Research, the global indium phosphide wafer market is projected to reach $507.21 million by 2034, nearly triple its size in 2025. Currently, JX Advanced Metals and its competitor Sumitomo Electric each hold approximately 40% of the market share.

In Europe, several key actions have also been taken.

When discussing optical communications, the market often pits "silicon photonics" against "InP," as if the adoption of silicon photonics would replace InP. Added to this is the previous intellectual property (IP) lawsuit between IQE and Tower Semiconductor, which further fuels this perception. However, the real industry trajectory is far more complex, as evidenced by the actions of IQE and Tower.

On June 15, IQE entered into a multi-year supply agreement with Tower Semiconductor to provide InP epitaxial wafers, supporting Tower’s silicon photonics platform in scaling up production for 200 Gb/channel pluggable transceivers, next-generation 400 Gb/channel modulators, and optical circuit switching. Under the agreement, Tower is required to commit to minimum purchase volumes in the first year, with IQE corresponding with supply commitments, followed by ongoing minimum purchase obligations. This highlights a key trend: next-generation silicon photonics platforms are not entirely abandoning III-V materials, but rather integrating high-performance InP components into mature silicon photonics platforms. Silicon photonics enables large-scale integration, CMOS process compatibility, and platform-based manufacturing, while InP continues to deliver critical functions such as high-performance light sources, modulation, and optoelectronic conversion.

Under another agreement, Tower will also grant IQE a broad, worldwide, royalty-free license for porous silicon patents. Previously, the two companies were involved in an intellectual property dispute, which Tower will resolve through a settlement, dismissing all pending litigation.

In its Q1 2026 financial report released on May 13, Tower stated that it is executing an aggressive global multi-fab silicon photonics capacity expansion plan, aiming to increase monthly silicon photonics wafer output capacity to more than five times the level at the end of 2025 by the end of 2026. Tower also announced that it has secured long-term silicon photonics supply contracts worth up to $1.3 billion with several key major customers, and has already received $290 million in advance payments from customers in Q1 2026. As equipment across multiple facilities comes online, Tower’s cumulative global investment in silicon photonics-related processes, equipment, and packaging will reach approximately $920 million.

In March 2026, ST announced it is considering modular expansion at its Crolles facility in France, aiming to quadruple 300mm silicon photonics capacity by 2027, with further expansion plans scheduled for 2028. The project is also supported by the European Sovereign Supply Chain Initiative. ST’s PIC100 silicon photonics platform, based on its 300mm wafer line, has entered full-volume production for leading global cloud providers, primarily serving core chips for 800G and 1.6T optical transceivers.

On June 2, Swedish chipmaker Sivers Semiconductors, specializing in high-power multi-wavelength laser arrays, entered into a deep strategic partnership with U.S.-based pure-play foundry giant GlobalFoundries to develop next-generation optical interconnect solutions for AI data center infrastructure. Specifically, Sivers’ advanced laser arrays will be directly integrated into GlobalFoundries’ silicon photonics platform.

Domestically, the development of optical chips is experiencing rapid growth.

According to industry statistics from Securities Times · Data Treasure, as of the first quarter of 2026, the total construction-in-progress scale of China’s seven core optical module listed companies rose to RMB 3.898 billion, more than six times higher than the same period four years ago (2022). China Post Securities noted in its research report that overseas giants account for 95% of the global indium phosphide market, with an overall supply-demand gap in the indium phosphide industry nearing 70%, and high demand is expected to persist until 2028.

On the evening of June 16, Dongshan Precision announced its approval for its wholly-owned subsidiary, Thorstone Photonics, and its subsidiaries to expand their operations in Changzhou with a project focused on optical chips and high-speed optical modules. The total investment amounts to $1.2 billion, funded entirely through the company’s own resources. Thorstone is a vertically integrated enterprise with capabilities in optical chip design, manufacturing, packaging, optical module assembly, and testing. Following its acquisition of Thorstone, Dongshan Precision has entered the core segment of AI optical communications, transitioning from traditional electronics manufacturing and consumer electronics supply chains.

From a financial contribution perspective, the consolidation of Thorlabs has already significantly outpaced its revenue contribution to Dongshan Precision’s profits. In 2025 and the first quarter of 2026, Thorlabs’ revenue contribution accounted for 3.58% and 16.02%, respectively, while its profit contribution reached 22.69% and 52.92%, respectively. This demonstrates that the optical communications business is not only growing rapidly but also exhibits strong profit elasticity—explaining why Dongshan Precision is willing to invest another $1.2 billion to further bet on this segment.

On June 3, Sanan Optoelectronics responded on its interactive platform, stating that its indium phosphide (InP) epitaxial growth, chip manufacturing, and packaging and testing technologies are leading domestically. The company has already achieved mass production capabilities for 6-inch InP photonic chips, with a current photonic technology output of 2,750 wafers per month, and its core epitaxial process capacity expanded to nearly 6,000 wafers per month. Regarding products, Sanan Optoelectronics noted in its 2025 annual report that it provides laser and detector chips for optical modules, including CW lasers, VCSELs, EMLs, and PDs. Photonic chips for 400G and 800G optical modules are already in bulk shipment, while samples of photonic chips for 1.6T optical modules have been delivered to customers for validation.

On the materials side, in April this year, Yunnan Germanium officially launched the "High-Quality Indium Phosphide Wafer Production Project." The project plans to expand a production line with an annual capacity of 300,000 wafers (equivalent to 4-inch wafers, including 6,000 six-inch wafers). Building upon the existing capacity of 150,000 wafers per year, the total annual capacity will reach 450,000 wafers upon completion, with a construction period of 18 months. Work is currently progressing as scheduled, including industry validation and equipment installation, with capacity being gradually ramped up in line with construction progress.

China's domestic optical chip industry chain is transitioning from "module assembly" to completing the full value chain: materials → epitaxy → chips → packaging and testing → modules.

It is widely known that in the field of optical chips, CPO is considered the industry's "holy grail." However, the actual deployment of CPO has consistently been delayed. As a result, the industry has a major concern: if CPO (co-packaged optics) continues to be delayed or underperforms, will optical module companies lose their growth potential?

Morgan Stanley's latest optical report provides a clear rebuttal. Morgan Stanley points out that investors are overly focused on the timing of CPO adoption, while overlooking the underlying constant—the demand for bandwidth growth.

Regardless of whether the market ultimately scales through pluggable optics, NPO, CPO, OBO, or a hybrid architecture, the demand for higher bandwidth should continue to drive increased adoption of optical engines, lasers, and related components per GPU or rack. Morgan Stanley’s view is that the specific architectural evolution is merely a matter of path, but the overall surge in optical content is certain.

What are CPO, NPO, and pluggable?

Traditional pluggable: Optical modules are inserted into the front panel of the switch, similar to a USB drive, and connected to the internal switching chip (ASIC) via copper cables.

NPO (Near-Package Optics): Move the optical engine inside the switch, adjacent to the switching chip, to reduce copper trace distance.

CPO (Co-Packaged Optics): Integrates optical chips directly onto the same substrate as switching chips (or GPUs), eliminating long-distance copper traces and minimizing power consumption and latency.

Currently, CPO indeed faces critical pain points such as extremely complex packaging, low yield rates, and the risk that if one component fails, the entire motherboard may be rendered unusable (due to poor repairability and serviceability). As a result, widespread adoption of CPO is likely to slow down. However, even if the market temporarily avoids CPO and continues using traditional pluggable optical modules or adopts a “copper/CPO hybrid approach,” the number of optical engines and lasers required per AI server and per GPU will still increase significantly.

The controversy surrounding CPO is not just about packaging location, but also about light source pathways. At its core, CPO aims to position the optical engine as close as possible to the switching or computing chip to shorten the transmission distance of high-speed electrical signals, thereby reducing power consumption and alleviating bandwidth bottlenecks. However, the industry currently has no single definitive solution for the light source.

The three most prominent approaches currently are: SiPh + CW Laser (silicon photonics + continuous wave laser), VCSEL (vertical-cavity surface-emitting laser), and MicroLED (micro-light-emitting diode). Differences in maturity, cost, distance, and power consumption among these approaches indicate that CPO is unlikely to adopt a single form; instead, multiple solutions will coexist across different distance tiers within AI data centers.

The SiPh + CW Laser solution, meaning "Silicon Photonics Chip + Continuous Wave Laser," has the highest technology readiness level, enabling effective transmission distances exceeding 1 kilometer, making it more suitable for connections in data centers with high demands on bandwidth, distance, and reliability. However, challenges remain in system-level power consumption, coupling, packaging, and cost.

The advantages of VCSELs lie in their high energy efficiency, low cost, strong arrayability, and high technological maturity; however, their effective range is typically limited to within a hundred meters, making them more suitable for short-distance interconnects within or between racks. Therefore, the role of VCSELs is not to replace SiPh + CW lasers, but rather to serve as a complementary solution in short-reach, low-cost, high-density optical interconnect scenarios.

MicroLED is a promising future-oriented solution with potential for low latency, low cost, and high energy efficiency, but it has a shorter effective range and the lowest level of technological maturity. It has emerged in recent years as a highly watched "dark horse" pathway in the field of optical interconnects. Startups such as Ayar Labs are actively exploring the integration of MicroLED—originally developed for display applications—into chiplet-level high-density near-field optical interconnects. This approach primarily utilizes arrays of extremely small (micrometer-scale) LEDs as light sources, directly integrated onto the edges or substrates of computing chips such as GPUs and HBM, transmitting data by electrically driving the MicroLEDs to emit rapid light pulses.

Thus, it is clear that the future of CPO is unlikely to be dominated by a single light source technology; instead, multiple solutions—such as SiPh, VCSEL, and MicroLED—will coexist in a layered architecture, tailored to varying distances, bandwidth densities, and cost constraints within AI data centers. This further underscores that expanding optical chip production is not simply a bet on one specific CPO technology, but rather an investment in the increased value across the entire ecosystem of light sources, optical engines, packaging and testing, and materials as AI clusters transition from electrical to optical interconnects.

In this global surge of photonic chip expansion fueled by AI computing power, no region is willing to fall behind: the United States is reshaping its domestic manufacturing chain through policy and major capital investments; Japan is fiercely defending its upstream material advantages; Europe is actively promoting the engineering implementation of silicon photonics and compound semiconductor hetero-integration; while China demonstrates strong industrial resilience through its astonishing speed in deploying production lines, the massive scale of ongoing projects, and its growing capability to extend upstream into materials and vertically integrated chips.

On the surface, this is a capacity race among manufacturers from the U.S., Japan, Europe, and China; in essence, it is a collective bet by the global semiconductor industry on “more light” following the shift of AI data centers from expanding computing power to expanding bandwidth.

The arms race in the photon era has entered a heated phase.

This article is from the WeChat public account "Semiconductor Industry Watch" (ID: icbank), authored by Du Qin (DQ).