Nvidia Invests $4 Billion in CPO: The Next Stop for AI Factory Networking

In early March 2026, Nvidia injected $2 billion each into photonics manufacturers Lumentum and Coherent, securing future production capacity and technology rights for advanced optical components with a total investment of $4 billion. This massive funding, which includes billions in procurement commitments, not only sent the stocks of these two optical giants soaring but also sent a clear signal to the market: Co-Packaged Optics (CPO), by integrating optical modules with chips, provides faster and more energy-efficient data transmission solutions for AI data centers.

From "Electrical Computing, Optical Transmission" to "Optical and Electrical Co-Packaging"
As AI models evolve towards clusters of tens of thousands to hundreds of thousands of GPUs, the interconnect technologies inside traditional data centers are hitting physical bottlenecks. The industry has long reached a consensus on "electrical computing, optical transmission" – using electricity for computation and light for transmission. However, the widely used pluggable optical modules currently face pain points in high-compute scenarios, including high power consumption, significant signal loss, and limited bandwidth.

The core breakthrough of CPO technology lies in co-packaging the network switch chip and optical module within the same socket, achieving tight integration. This "extreme integration" strategy promises to fundamentally solve the aforementioned issues of traditional solutions. Some industry insiders believe that looking at the technological roadmap for interconnect development, CPO might be an irreplaceable solution, while NPO (Near-Packaged Optics) could be a transitional phase.

Copper cable solutions face increasingly severe challenges in transmission density and energy efficiency. Data indicates that to achieve high transmission rates above 1.6Tbps, traditional copper cables consume over 10 pJ/bit, significantly increasing overall system power consumption. In contrast, Micro LED CPO solutions can reduce energy consumption per bit to just 5% of copper cable solutions, around 1-2 pJ/bit, making them a promising alternative for optical interconnects thanks to their energy efficiency.

The Technology Roadmap Behind the Strategic Investment
Nvidia's latest investment directly targets the CPO technology roadmap it is vigorously pursuing. It is reported that in March 2025, Nvidia launched two series of data center switches utilizing CPO technology, with both Lumentum and Coherent being core suppliers of the necessary laser components for this technology.

Nvidia is partnering with Lumentum to advance the world's most cutting-edge silicon photonics technology, aiming to build the next generation of gigawatt-scale AI factories. Nvidia will also collaborate with Coherent to develop next-generation silicon photonics technologies tailored for AI infrastructure.

The mainstream CPO architecture is expected to adopt an "external laser source + internal silicon photonic modulation" structure. Due to the fundamental principle of stimulated emission, laser sources have high power density and require operation in relatively low ambient temperatures, ideally below 60°C. The External Laser Source (ELS) approach places high-performance laser sources externally, reducing thermal management and integration complexity on the chip side, effectively solving this tricky heat dissipation problem.

The Rubin Platform and the CPO Scaling Deployment Timeline
Looking at Nvidia's next-generation product roadmap, the Rubin platform will offer multiple CPO switch options. The Rubin Ultra, expected to be launched in 2027, will likely adopt the Kyber rack architecture.

Nvidia remains a pioneer in Scale-out CPO. Optical engine shipments are projected to reach 600,000 to 1 million units in 2026, with switch shipments around 23,000 units, where Spectrum is a key product; looking ahead to 2027, the next-generation Quantum switches are expected to begin volume shipments.

In the Scale-up CPO domain, Nvidia may introduce a CPO version for the Rubin Ultra NVL144, with Quantum serving as the NVLink GPU interconnect solution to support GPU-to-GPU communication. Optical I/O solutions for next-generation GPUs are also under development.

Regarding production capacity, TSMC aims to achieve a monthly production capacity of 10,000 Photonic Integrated Circuit (PIC) wafers by the first quarter of 2027 to meet strong CPO demand, with Nvidia anticipated to be a primary user.

Industry Resonance: The Tipping Point for CPO Commercialization Nears
Nvidia's investment not only benefits existing optical module leaders but also points the way for future technological evolution. With the deployment of Rubin Ultra racks and the volume ramp-up of CPO switches, demand for core components like M9-grade copper-clad laminates, NPO/CPO optical engines, external laser sources, and fiber connection units is poised for explosive growth.

Companies like Microsoft and Meta have announced plans to spend nearly $700 billion on AI data center construction, which will extensively utilize CPO technology. Broadcom has already introduced its first 102.4T CPO switch, the Davisson platform, with shipments expected to begin in the second half of 2026. The next decade is expected to witness explosive growth for CPO. By 2030, the global CPO market size could exceed $40 billion.

Nvidia's $4 billion bet this time represents a final confirmation of the CPO technology roadmap and a crucial step in the evolution of AI factory network architecture. From pluggable to co-packaged, from electrical to optical interconnects, the "next stop" for AI infrastructure is accelerating towards reality.