Beyond the I/O Wall: NPO vs CPO for Scaling AI Interconnects Across Racks

Large language models are pushing toward ten-trillion-parameter scale, and compute growth is outpacing interconnect bandwidth, creating a pronounced I/O Wall. While Scale-Out and Scale-Across layers have already transitioned from copper to optics, the Scale-Up domain — historically copper-dominated due to short reach and extreme sensitivity to power and cost — is now expanding from single-rack to multi-rack configurations, exposing copper to three compounding pressures.

Three Pressures on Copper Scale-Up Interconnects

Transmission distance: NVIDIA's roadmap from Blackwell to Vera Rubin doubles per-GPU bidirectional NVLink bandwidth from 1.8 TB/s to 3.6 TB/s and domain aggregate bandwidth from ~130 TB/s to 260 TB/s, with Rubin Ultra targeting 576-GPU multi-rack domains. An 8×72-GPU deployment stretches inter-rack Scale-Up links beyond copper's practical reach.

Signal integrity: Higher per-lane rates demand better equalization and retiming. In a front-panel pluggable architecture, electrical signals traverse long board traces and connectors. NVIDIA's high-speed electrical channel example shows 22 dB loss for pluggable modules versus only 4 dB for CPO.

Power consumption: Copper requires stronger signal compensation at high speeds, causing power to spike. At 1.6 Tb/s, a pluggable interface consumes ~30 W while a CPO implementation consumes ~9 W — a ~70 % reduction.

NPO (Near-Package Optics): Pragmatic Near-Term Solution

NPO places an optical engine near the main ASIC package, connected by short electrical channels. This reduces channel loss and compensation overhead, lowering power and improving layout density. Compared with CPO, NPO offers limited integration, bandwidth density, and latency gains, but it is thermally friendly, highly maintainable, and supports separate manufacturing, test, and replacement — making its deployment conditions relatively mature. It serves as the current transitional mainstream for Scale-Up optical interconnects. In March 2026, Zhongji Innolight's subsidiary TeraHop demonstrated a 6.4T 4×DR8 NPO module at OFC 2026 targeting AI Scale-Up interconnects; Tianfu Communications is concurrently engaging multiple customers on custom NPO development.

CPO (Co-Packaged Optics): Ultimate Long-Term Direction

CPO integrates the switch ASIC and optical engines on a common package substrate, shrinking the electrical path to millimeter scale. Using an external laser source (ELS) silicon-photonics example: the ELS supplies continuous-wave light; silicon-photonic modulators in the optical engine imprint data onto the optical carrier; receivers use photodetectors and transimpedance amplifiers for optical-to-electrical conversion. The external laser simplifies thermal management and replacement.

Power & density: By shortening high-speed electrical connections and reducing equalization/retimer needs, CPO cuts interface power. Meta's measured data on Broadcom Bailly (51.2T) shows per-800G optical-engine-plus-laser power dropping from ~15 W (pluggable) to ~5.4 W (CPO), saving ~65 % and ~600 W per system.

Front-panel relief: Moving electro-optic conversion next to the switch ASIC frees front-panel space for optical fiber connectors, enabling higher port density. Broadcom's October 2025 Tomahawk 6–Davisson platform co-packages sixteen 6.4 Tb/s optical engines with the Tomahawk 6 switch die, delivering 102.4 Tb/s switching capacity.

Co-design challenges: CPO demands tighter chip-package-optics-thermal co-design, raising the bar for yield and assembly.

Industry Chain & Key Components

CPO emphasizes package-level co-design and yield; NPO retains more flexible board-level integration. Both require higher-precision optical coupling and higher-density interconnect solutions. Critical elements for CPO/NPO switches include: switch ASIC, optical engine and packaging, fiber arrays and coupling, in-chassis fiber routing and connectors — with some products already incorporating external laser sources and fiber-shuffle components.

Outlook: Co-Evolution, Not Replacement

Short term — NPO leads: Mature deployment, thermal friendliness, and maintainability make NPO the pragmatic choice for bridging today's multi-rack Scale-Up gap.

Long term — CPO wins: Superior power/bandwidth/latency/density balance, front-panel space savings, and potential >50 % BOM cost reduction position CPO as the eventual mainstream. TSMC's COUPE platform, using SoIC hybrid bonding to stack EIC and PIC face-to-face with multiple coupling options, exemplifies the scalable integration path.

Synergy: NPO builds manufacturing and test experience for optical engines, fiber arrays, and ELS that CPO later leverages; CPO pushes the supply chain toward higher integration, lower power, and higher bandwidth density.