Nvidia spends heavily for "light": Secures scarce indium phosphide device production capacity from Lumentum and Coherent

```

The trend of “optics advancing” is becoming increasingly evident, and industrial waves are fostering new hotspots. NVIDIA announced that it will invest $2 billion each in photonics technology giants Lumentum and Coherent, as well as securing long-term procurement commitments worth several billion dollars and future production rights. This is another historic move in optical interconnect by NVIDIA after launching the world’s first mass-produced CPO switch in March of last year.

This article explains the current state and future blueprint of the industry in one read.

1. What happened? NVIDIA’s massive investment

Before NVIDIA’s GTC conference in March 2026, NVIDIA announced it would invest $2 billion each in photonics giants Lumentum and Coherent, securing long-term procurement commitments worth several billion dollars and future production rights. This is another historic layout in optical interconnect by NVIDIA after launching the world’s first mass-produced CPO switch in March last year. Undoubtedly, CPO technology is set to become the cornerstone of the next generation of “gigawatt-level AI factories.” Key segments such as optical engines, external laser sources, and fiber connection units are set for revaluation, and companies with a supply chain positioning advantage are expected to benefit first.

Both deals are non-exclusive agreements but include multi-billion dollar purchase commitments and prioritized access to future capacity for advanced laser components and optical networking products.

NVIDIA made it clear in both statements: “Optical interconnects and advanced packaging integration are the foundation of next-generation AI infrastructure, as they can deliver ultra-high bandwidth, highly energy-efficient connectivity for AI factories.” 

Lumentum and Coherent stocks continue to rise, and optics communications companies in the A-H market also surged collectively. Core CPO industry chain companies like Yuanjie Technology, Shijia Photonics, TFC Communication, and Taichen Photonics are all under high market attention. This response reflects capital markets’ rapid digestion of NVIDIA’s strategic intentions—the competition for AI infrastructure is shifting from a pure GPU compute power race to a deep game of high-speed optical interconnect capabilities. 

To understand NVIDIA’s strategic plan, one needs to step back for a more macro industry perspective.

In the past decade, GPU compute power mainly relied on chip process evolution (from 7nm to 2nm). However, by 2026, the performance gains from increasing transistor density alone are rapidly diminishing due to power constraints (the Power Wall). NVIDIA’s core strategy has shifted from “selling chips” to “selling factories.” In the forward-looking “Gigawatt AI Factory” vision unveiled at the 2026 GTC conference, thousands of GPUs need to be connected as a whole through extremely high-speed and ultra-low-power networks. At this point, interconnect bandwidth has replaced compute power (TFLOPS) as the primary measure of AI system performance. 

CPO technology eliminates the long copper driving lines by directly packaging the optical engine next to the compute chip, cutting interconnect power by 30%–50% and reducing latency by about 20%. This is the core reason why NVIDIA is targeting Lumentum and Coherent: to control the key CPO components—lasers.

2. Why is this important? Is CPO a must-have for the “gigawatt-level AI factory”?

To understand why NVIDIA is betting big on optical interconnect, we must return to the key determinant of compute chip interconnect bandwidth—SerDes (high-speed serial deserializer) rate advancement.

NVIDIA’s NVLink SerDes rate has evolved from 56Gbps in Ampere architecture to 224Gbps in Blackwell, supporting a leap in single-chip interconnect bandwidth from 600GB/s to 1800GB/s. In the Rubin era, it will advance further to 448G PAM4 and even 896G PAM6. However, this rate increase brings two worsening physical bottlenecks:

First, signal attenuation: At 224G, Nyquist frequency reaches 56GHz, and upgrading to 448G would push frequency to 112GHz. At these frequencies, traditional PCB trace insertion loss can reach 20–50dB, meaning the signal becomes irrecoverable after only a few inches. 

Second, power wall: As SerDes rate approaches 448G, DSP power consumption accounts for over 50%, with SerDes taking up more than 40% of switch chip power. Heat flux density can hit 50W/cm². Traditional air cooling reaches physical limits.

Shortening the electrical distance between electro-optical conversion and the switch chip, and reducing or even eliminating the use of high-power DSPs, becomes the central logic of optical interconnect technology evolution.

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

CPO (Co-Packaged Optics) is the ultimate solution, co-packaging the optical engine and switch IC on the same substrate/interposer, compressing electrical distance to under 50mm, lowering total system power by more than 65%. According to NVIDIA’s technical data, CPO can lower per-port power from 30W to 9W and improve signal integrity 64-fold.

NVIDIA Rubin Ultra, as the flagship GPU for next-generation AI data centers, achieves total rack bandwidth of 1.5PB/s with 144 GPUs working together. This breakthrough bandwidth poses unprecedented challenges for scale-up network architecture.

The Rubin Ultra rack adopts a two-level, four-canister stacked design:

① Layer 1: Internal canister uses orthogonal backplanes for non-blocking switching, utilizing M9-grade ultralow-loss CCL materials.

② Layer 2: Cross-canister interconnect is handled by 72 NVSwitch chips with 648 3.2T NPO optical engines. The GPU-to-optical-engine ratio is as high as 1:4.5.

This means a single Rubin Ultra rack requires almost 650 3.2T optical engines. When these racks are widely deployed in global AI data centers, the need for optical interconnect components will be astronomical. 

NVIDIA’s photonics strategy breaks down in three steps. Step one: launch the switch.

In 2025, NVIDIA will launch the world’s first mass-produced CPO switch—the Quantum X800-Q3450, with total switching bandwidth as high as 115.2T. This device uses four Quantum-X800 ASICs in a multi-plane switching configuration, each surrounded by six removable optical submodules, and each submodule integrates three optical engines. 

NVIDIA plans to release the Spectrum-X Photonics CPO switch family covering Ethernet in the second half of 2026, forming a comprehensive layout from InfiniBand to Ethernet.

Step Two: Secure core supplier capacity

This $2 billion investment each into Lumentum and Coherent is the extension of the second step of this strategy.

Lumentum is one of the world’s top CPO-optimized laser module players; its transmitter produces 1311nm beams and uses thermal management to optimize network performance. Lumentum started expanding CPO component production capacity in September last year. X800-Q3450’s technical teardown shows it uses 18 ELS modules as laser sources, each with 8 continuous-wave (CW) DFB laser chips; the CPO system needs higher-powered laser sources (each CW-DFB chip outputs about 350mW). Key companies capable of mass-producing such CW laser units include Lumentum, Coherent, Broadcom, Furukawa Electric, etc.

Coherent’s business also spans both industrial and data center lasers; it recently launched a laser emitter optimized for CPO systems and provides related fiber optic cables. Coherent and NVIDIA have a 20-year partnership.

On the surface, a direct investment of $4 billion is huge, but the deeper value lies in “multi-billion dollar purchase commitments” and “future production rights.” This means NVIDIA not only funds the suppliers’ capacity expansion, but also locks in a portion of key components deliveries in coming years through long-term commitments. In generally tight supply cycles, this strategy of "funding + orders" secures NVIDIA’s leading position in the CPO supply chain.

Step Three: Build an entire industry ecosystem

NVIDIA is leveraging its strong profits to create a powerful ecosystem that supports advanced AI systems. Previously, NVIDIA directly invested in data center company CoreWeave and AI model developers OpenAI and Anthropic.

The investment logic is clear: use capital ties to lock in supplier capacity, use strategic investment to tie down client demand, and ultimately form a closed loop of “chip – network – application.” As the core network link connecting compute chips to AI apps, CPO naturally becomes the strategic high ground in this layout. 

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

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

① Lumentum: first-mover technology + capacity expansion

Lumentum’s last quarter revenue was $665.5 million. Its product line covers traditional pluggable transceivers, optical switches for optimized AI cluster traffic, and industrial lasers for metal welding, etc. Lumentum started expanding CPO component production in September last year. More critically, Lumentum is the sole supplier of laser components for NVIDIA’s Quantum X3450 CPO switch. 

② Coherent: 20-year partnership + broad product line

Coherent has a 20-year history with NVIDIA. Besides laser emitters, Coherent sells optical fiber cables and makes spectrum synthesis analyzers and testing tools for network suppliers to engineer and simulate various optical traffic types. 

③ Opportunities for the domestic industry

- Fiber array units (FAU): TFC Communication may become the primary FAU supplier for the X800-Q3450 CPO switch and has invested in an advanced packaging plant in Suzhou, China.

- Optical distribution boxes: Taichen Photonics is a leading industry player, serving NVIDIA, Broadcom, etc., in conjunction with Corning.

- MT ferrules: Fooke Xima (subsidiary of Shijia Photonics) has strong mold design and manufacturing capability.

- Optical chips: Yuanjie Technology is a domestic leader in optical chip IDM. Its CW 70mW laser chips have been mass delivered; 100mW chips have passed customer verification.

④ Real-world challenges facing CPO

Though NVIDIA is betting big on CPO, the industry widely believes that full-scale optical transmission faces practical constraints:

First, maintenance cost. CPO integrates the optical engine with the switch chip; if a fault occurs, the entire device often needs replacement, making data center operators cautious.

Second, standards are incomplete. Optical transmission standards are not fully unified, and pluggable connectors and modular repair mechanisms are still developing.

In summary, on the technical level, CPO is the only path to solve the power and signal integrity bottlenecks from SerDes rate increases. On the industrial level, locking in core supplier capacity underpins the gigawatt AI factory supply chain. On the competitive level, optical interconnect is becoming the new moat after GPU compute strength. For investors, understanding this strategic leap means moving beyond the single GPU track and reassessing AI infrastructure investment value from a holistic “optical-board-copper-power” perspective. As Rubin Ultra racks roll out and CPO switches scale up, key segments—optical engines, lasers, fiber connection units—face a golden window for value revaluation.

Risk Warning and DisclaimerThe market entails risk and investments must be made prudently. This article does not constitute personal investment advice nor does it take into account the special investment objectives, financial situation, or needs of any individual user. Users should consider whether any opinions, views, or conclusions in this article suit their specific situation. Investments are made at your own risk. ```