Optics Move Closer to Chips: NPO and CPO Reshape AI Compute Interconnect
TREE NEWS reports: In a new research report, Zhongtai Securities analysts Wang Fang and Sun Yuewen argue that the explosive growth of AI inference demand is pushing data center interconnect to a critical inflection point. Citing Google data, they note that monthly token processing across Google’s products surged from 9.7 trillion in May 2024 to over 3,200 trillion by May 2026—a more than 300-fold increase in two years. This surge is forcing a shift from copper cabling to near-package optics (NPO) and co-packaged optics (CPO), as high-speed electrical signals become increasingly difficult to transmit over even short distances.
What Happened
The report outlines a clear technological progression: from copper interconnects and pluggable optical modules, to NPO and CPO. The core driver is that as token volumes explode, GPUs need to collaborate more frequently and rapidly, demanding low-latency, high-bandwidth full-mesh networks. Capital expenditure data from hyperscalers confirms this trend: Google raised its 2026 capex guidance to $195–205 billion, Amazon to ~$220 billion, and Meta to $130 billion.
The interconnect architecture is bifurcating into two main tracks: Scale-up (tightly coupled, low-latency compute domains) and Scale-out (connecting servers and racks across a network). Both require optical interconnects but with different power, fault, and maintenance constraints. Copper’s limits are also shrinking: as SerDes speeds advance from 25G NRZ to 200G PAM4, effective copper reach drops from ~4 meters to ~1 meter, making optical solutions increasingly necessary.
Market Impact Analysis
The shift from pluggable optics to NPO and CPO represents a significant reallocation of value across the AI supply chain. Key beneficiaries include high-power CW lasers (for external light sources), fiber array units (FAUs), high-density connectors, and advanced packaging. For example, Cisco estimates CPO can reduce whole-system power by 25–30% compared to pluggable architectures in a 51.2T switch. Broadcom’s data shows CPO optics power at 7.1kW versus 16.2kW for DSP pluggable in a 576-GPU interconnect.
NVIDIA is the most prominent CPO adopter, with its Quantum-X800 and Spectrum-X switches already in production. TSMC’s COUPE platform is emerging as the key manufacturing foundation, with NVIDIA, Broadcom, and Ayar Labs all moving toward it. However, CPO faces five major hurdles: joint yield, thermal management, external light sources, testing/diagnostics, and repair. The yield issue is critical—any defect in the co-packaged assembly can lead to expensive scrap.
Key Takeaways for Investors
- Value migration: Traditional pluggable optical module makers may face disruption, while suppliers of CW lasers, FAUs, high-density connectors, and advanced packaging stand to gain.
- Focus on execution: The report emphasizes that despite the clear logic, actual earnings will depend on customer design wins, volume deliveries, yield rates, and reliability data.
- Technology divergence: Silicon photonics and VCSEL will coexist, with VCSEL suited for short-reach, high-parallelism links and silicon photonics for longer distances and WDM.
- Watch for milestones: Key indicators include Alibaba’s 3.2T NPO pilot, Tencent’s dual validation of 3.2T silicon photonics and VCSEL, and Huawei’s Hi-ONE platform.
Investors should monitor the pace of NPO/CPO adoption in hyperscale data centers, as this will determine the timing and magnitude of the value chain shift.
