NVIDIA splashes out on "light": Secures Lumentum and Coherent's scarce indium phosphide component capacity

The “Light Progress” trend is becoming increasingly clear, with industry waves fueling new hotspots and opportunities. NVIDIA announced investments of $2 billion each in photonics giants Lumentum and Coherent, along with long-term procurement commitments worth billions of dollars and future capacity rights. This marks another historic move in NVIDIA’s layout in the optical interconnect field, following the launch of the world’s first mass-produced CPO switch in March last year.

This article explains the industry development progress and future blueprint.

1. What happened? NVIDIA’s massive investments

In March 2026, ahead of the NVIDIA GTC conference, NVIDIA announced investments of $2 billion each in photonics giants Lumentum and Coherent, along with long-term procurement commitments worth billions and future capacity rights. This is another historic move in NVIDIA’s optical interconnect domain, following the launch of the world’s first mass-produced CPO switch last March. Undoubtedly, CPO technology will become the core foundation of the next-generation “Gigawatt AI Factory.” Key components such as light engines, external laser sources, and fiber connection units will see a reassessment of value, with suppliers holding supply chain advantages likely to benefit first.

Both deals are non-exclusive agreements but include “billions of dollars” in procurement commitments and priority access to future advanced laser components and optical network products.

NVIDIA explicitly states in both announcements: “Optical interconnects and advanced packaging integration are the foundation of next-generation AI infrastructure because they can provide ultra-high bandwidth and high energy efficiency connections for AI factories.”

Lumentum and Coherent continue to rise, and optical communication companies listed in Hong Kong’s AH shares also surged collectively. Core CPO industry chain targets such as Yuanjie Technology, Shijia Photonics, Tianfutong, and Taichen Optical have attracted high market attention. Behind this reaction is the capital market’s rapid digestion of NVIDIA’s strategic intent—competition in AI infrastructure is shifting from mere GPU computing power to a deep game of high-speed optical interconnect capabilities.

To understand NVIDIA’s strategic layout, we need to step back to a broader industry perspective.

Over the past decade, GPU computing power growth mainly depended on chip process evolution (from 7nm to 2nm). However, by 2026, the performance gains from simply increasing transistor density are sharply diminishing due to power constraints (Power Wall). NVIDIA’s current core strategy has shifted from “selling chips” to “selling factories.” In the visionary “Gigawatt AI Factory” concept presented at GTC 2026, thousands of GPUs need to be interconnected via ultra-high-speed, ultra-low-power networks. At this point, interconnect bandwidth has overtaken raw computing power (TFLOPS) as the primary metric of AI system performance.

CPO technology, by directly packaging light engines next to compute chips, eliminates lengthy copper wiring, reducing interconnect power consumption by 30%-50% and latency by about 20%. This is the fundamental reason NVIDIA is locking in laser components—core parts of CPO—from Lumentum and Coherent.

2. Why is this important? Is CPO a must-have for “Gigawatt AI factories”?

To understand why NVIDIA is heavily investing in optical interconnects, we must revisit the key factor driving chip interconnect bandwidth—the evolution of SerDes (High-Speed Serializers/Deserializers).

NVIDIA’s NVLink SerDes speeds have advanced from 56Gbps in the Ampere architecture to 224Gbps in the Blackwell architecture, supporting intra-chip interconnect bandwidth from 600GB/s to 1800GB/s. Rubin era will push this further to 448G PAM4 and even 896G PAM6. However, behind these speed increases are two major physical bottlenecks:

First, signal attenuation: signals above 224G have a Nyquist frequency up to 56GHz; at 448G, this jumps to 112GHz. In this frequency band, traditional PCB traces can have insertion losses of 20-50dB, causing signals to decay to irrecoverable levels after just a few inches of transmission.

Second, power wall: as SerDes speeds reach 448G, DSP power consumption will exceed 50%, and SerDes in exchange chips will account for over 40% of total power, with thermal density reaching 50W/cm². Traditional air cooling has approached physical limits. Shortening the electrical distance between optical transceivers and exchange chips, and reducing or eliminating high-power DSPs, becomes the core logic of optical interconnect evolution.

NPO (Near-Package Optics), as a transitional solution, deploys optical engines on switch boards via LGA connectors, shortening the physical distance to within 150mm, reducing power consumption by over 50%.

CPO (Co-Packaged Optics) is the ultimate solution, integrating optical engines and exchange chips on the same IC substrate or silicon interposer, compressing electrical connection distances to within 50mm, and reducing overall system energy consumption by over 65%. According to NVIDIA’s technical data, CPO can cut port power from 30W to 9W, with signal integrity improved by 64 times.

NVIDIA’s Rubin Ultra, the next-generation AI data center GPU flagship, offers a total bandwidth of 1.5PB/s across 144 GPUs. This breakthrough bandwidth challenges unprecedented scale-up network architectures.

Rubin Ultra’s cabinet adopts a stacked dual-layer architecture with four canisters:

① First layer: Non-blocking switching within canisters via orthogonal backplanes, requiring ultra-low-loss CCL materials at M9 grade.

② Second layer: Inter-canister interconnect via 72 NVSwitches and 648 NPO optical engines, with a GPU-to-optical engine ratio of 1:4.5.

This means a single Rubin Ultra cabinet needs nearly 650 3.2T optical engines. When such cabinets are deployed at large scale in global AI data centers, the demand for optical interconnect components will be astronomical.

NVIDIA’s optical strategy comprises three steps. First, launching switches.

In 2025, NVIDIA will release the world’s first mass-produced CPO switch—Quantum X800-Q3450, with a total switching bandwidth of 115.2T. This device uses four Quantum-X800 ASICs in a multi-plane switching configuration, each surrounded by six detachable optical sub-assemblies, each containing three optical engines.

NVIDIA plans to launch the Spectrum-X Photonics series CPO switches in late 2026, covering Ethernet ecosystems and forming a complete technology layout from InfiniBand to Ethernet.

Second, securing core supplier capacity

The $2 billion investments in Lumentum and Coherent are an extension of this strategy.

Lumentum is one of the top global suppliers of optimized CPO laser modules, capable of producing 1311nm laser beams with integrated temperature management for network performance. Since September last year, Lumentum has initiated capacity expansion for CPO components. Based on the X800-Q3450 technical breakdown, this device uses 18 ELS modules as laser sources, each containing 8 CW DFB laser chips, requiring relatively high-power laser sources (~350mW per CW-DFB chip). Key manufacturers capable of mass-producing continuous-wave lasers include Lumentum, Coherent, Broadcom, and Furukawa Electric.

Coherent’s business spans industrial and data center laser fields, recently launching laser emitters optimized for CPO systems, along with fiber cables and related equipment. Coherent also has a 20-year partnership with NVIDIA.

Although the $4 billion investment appears huge, its deeper value lies in the included “billions of dollars” in procurement commitments and “future capacity rights.” This means NVIDIA not only provides capital support for capacity expansion but also locks in future key component supplies through long-term procurement commitments. In a tight capacity cycle, this “funds + orders” approach ensures NVIDIA’s priority position in the CPO supply chain.

Third, building a complete industry chain ecosystem

NVIDIA leverages its large profits to develop a robust ecosystem supporting advanced AI systems. Previously, NVIDIA invested directly in data center company CoreWeave, as well as AI model developers OpenAI and Anthropic.

This investment logic is clear: lock core suppliers’ capacity through capital, bind downstream customer demand via strategic investments, and ultimately form a complete “chip-network-application” closed loop. As CPO connects compute chips and AI applications, it naturally becomes a strategic hub in this ecosystem layout.

3. What to watch next? The “smile curve” of the CPO industry chain

Based on disassembly of NVIDIA’s CPO switch products, the core segments and value distribution of the CPO industry chain are as follows:

① Lumentum: technology leadership + capacity expansion

Lumentum’s quarterly revenue is $665.5 million, with product lines including traditional pluggable transceivers, optical switches for AI cluster network traffic optimization, and industrial lasers for metal welding and other industrial scenarios. Since September last year, Lumentum has started capacity expansion for CPO components. More critically, Lumentum is the sole supplier of laser components for NVIDIA’s Quantum X3450 CPO switch.

② Coherent: long-term partnership + broad product portfolio

Coherent has a 20-year cooperation with NVIDIA. Besides laser emitters, Coherent also sells fiber cables and participates in the spectral synthesis instrument market, providing engineering support for network equipment vendors to simulate different optical network traffic types.

③ Opportunities in the domestic industry chain

• Fiber connection units (FAU): Tianfutong may become the main supplier of FAU for X800-Q3450 CPO switches, having invested in an advanced packaging factory in Suzhou, China.

• Fiber switch boxes: Taichen Optical is a leading enterprise collaborating with Corning to serve NVIDIA, Broadcom, and other clients.

• MT plug-in modules: Fokushima (a subsidiary of Shijia Photonics) has strong mold design and manufacturing capabilities.

• Optical chips: Yuanjie Technology is a leading domestic IDM optical chip manufacturer, with CW 70mW laser chips already mass-delivered and CW 100mW laser products validated by customers.

④ Challenges facing CPO

Despite NVIDIA’s strong focus on CPO, industry consensus indicates that full-scale optical transmission still faces practical limitations:

First, maintenance costs: integrating optical engines with switch chips into a single package means that when faults occur, the entire device often needs replacement, making data center operators cautious about adoption.

Second, standardization is incomplete: current optical transmission standards are not fully unified, and pluggable optical connectors and modular repair mechanisms are still under development.

In summary, at the technical level, CPO is the necessary path to address the bottlenecks of power consumption and signal integrity caused by SerDes speed upgrades; at the industry level, locking core supplier capacity is the foundation for building “Gigawatt AI factories”; at the competitive level, optical interconnect capability is becoming a new moat after GPU computing power. For investors, understanding this strategic shift means moving beyond a single GPU track and re-evaluating AI infrastructure investment value from a full-stack perspective of “optics, boards, and copper.” As Rubin Ultra cabinets and CPO switches scale up, core components like light engines, laser sources, and fiber connection units will enter a golden window for value reassessment.