CPO and NPO fall under the "on-board/integrated within the package" route, representing a "light engine moving from the panel to the chip" integration spectrum (plug-and-play → LPO → NPO → CPO → OIO).
While XPO is the ultimate improvement version of the "panel plug-and-play" route (initiated by Arista in March 2026 with 45 enterprises forming the XPO MSA to introduce, it is an alternative upgrade to OSFP, not a synonym for traditional modules).
Abbr. | Full name | Position of the light engine | Key meaning |
| XPO | eXtra-dense Pluggable Optics | Front panel of the switch (panel position) | P = Pluggable, emphasizing hot-swappable capability |
| NPO | Near-Packaged Optics | The motherboard is close to the ASIC, approximately 1-5 cm away. | The optical engine is independent and can be replaced separately |
| CPO | Co-Packaged Optics | Same package or substrate as ASIC, < 1 mm | Deep integration, non-disconnectable |
Why do AI data centers require these solutions?
The core contradiction can be summed up in one sentence: Bandwidth density and energy efficiency are increasing, but maintainability and ecology must not be sacrificed.
The data rate has changed from 800G to 1.6T to 3.2T, and the single-channel data rate has changed from 100G to 200G to 400G;
The high-speed signals of the traditional panel module need to travel over 100 mm on the PCB, resulting in severe deterioration of loss, crosstalk, and reflection;
The most power-consuming part in the module is the DSP (accounting for about half of the module's power consumption), and the higher the rate, the more difficult it is to withstand;
The power density of a single cabinet is approaching 100 kW, and the air cooling has reached its limit. The cooling method has been forced to be restructured;
The GPU cluster has gone from thousands of cards to hundreds of thousands of cards. Scale-up (within a rack/super node) and Scale-out (between racks/clusters) have completely different requirements for bandwidth, latency, and reliability.
Therefore, an industry rule has emerged: The closer the optical engine is to the chip, the lower the power consumption and latency, but the maintainability and ecological openness are worse. The three routes essentially take three different points on this curve.
Comparison of the characteristics of the three schemes:
| Dimension | XPO (Plug-in Limitation) | NPO (Pragmatic Moderate) | CPO (Deep Integration) |
Electrical link length | Centimeters to decimeters (up to) | 1-5 cm | <1 mm |
Bandwidth of a single module/engine | 12.8T (64×200G), roadmap 25.6T | 3.2T mainstream, 6.4T under development | 3.2T/6.4T starting point, single machine can reach 100T |
| System density | 16 modules in 10U → 204.8T per rack unit, approx. 4 times that of OSFP | Mid-high (with many ports but low single-port rate) | The highest (by bypassing panel space limitations) |
Power consumption | Approximately 400W per module. Must be equipped with native liquid cooling. | The 3.2T optical engine is approx. 20W (linear direct drive without DSP) and has a LGA socket that can be replaced individually | 30%-70% lower than traditional solutions, optimal per-bit energy efficiency |
| Maintainability | Front panel hot-swappable, single-module replacement. | The optical engine does not require a motherboard replacement. | The optical engine is integrated with ASIC, and faults usually require replacement of the entire board or the entire machine. |
The respective roles they play: not as substitutes, but as layers and divisions of labor.
CPO—Addressing "The Limits of Bandwidth Density and Energy Efficiency"
Role: For ultra-high bandwidth networks above 51.2T with single-channel 200G/400G, it is the ultimate direction for Scale-up scenarios (interconnection of GPUs/switch chips within the rack).
Mechanism of action: Reduce the electrical link to the millimeter level, almost eliminating PCB losses, and eliminating the need for timing and DSP, making it the only technology that can simultaneously meet the "power consumption + density" indicators in the 3.2T era.
Penetration rate: Approximately 0.5% in 2026, and TrendForce predicts about 35% in 2030.
Cost: Non-hot-swappable → The operation model must be restructured; the cost of a single optical engine is still high (some estimates range from 3.5–40,000 USD); cooling must be liquid cooling.
NPO—Solving the problem of "Need to increase production immediately"
Role: The greatest common divisor before CPO scale-up, the industrial golden window from 2026 to 2028.
Mechanism of action: The optical engine is placed close to the ASIC (at the centimeter level), significantly reducing the electrical link loss; the optical engine is decoupled from the switching chip, allowing cloud vendors to independently select and replace it, bypassing the barrier of advanced packaging.
Cost: Customized motherboards and cooling designs are still required, and it cannot be directly inserted into existing equipment; Fundamentally, it is still a "transition form", and CPO needs to give way in the long term.
XPO—Solving the problem of "How to quickly increase the port bandwidth of existing data centers"
Role: Plug-and-play ecological defensive innovation. With a brand-new electromechanical structure and native liquid cooling, increase the density of the panel by one level, and extend the lifespan of the plug-and-play route.
Mechanism of action: One module supports 8 1.6T OSFPs, and 10U is reduced from 51.2T to 204.8T, meaning that the number of switch racks for a 100,000-card cluster can be reduced by approximately 75%, directly saving floor space, power distribution, and cabling costs. Supports the full range of DR/FR/LR/SR/ZR, and expansion can be plugged in as needed without shutting down the entire machine.
Cost/Risk:
① The highest power consumption (400W per module), with the unit bandwidth power consumption being approximately twice that of the 1.6T module;
② Highly dependent on liquid cooling, when the power density of the rack is less than 120 kW, the utilization rate of liquid cooling is insufficient and not cost-effective;
③ The electrical path is longer than NPO/CPO, with higher SerDes power consumption and PCB material grade requirements;
④ It is necessary to upgrade the switching chip to 102.4T/204.8T for it to be meaningful;
⑤ The module width exceeds the existing standard, and the cabinet structure needs to be accepted by the owner.
Application Scenario
| Scene | Distance/Feature | Preferred solution | Reason |
| Chip-to-chip / Super-node (Scale-up) | < 1m, copper cable has reached its limit | CPO (short-term) - OIO (long-term) | Ultra-high bandwidth density and energy efficiency, closed ecosystem within the rack for performance enhancement |
Rack-to-rack / Leaf spine (Scale-out) | From tens of meters to 2 km | NPO, LPO/LRO, for high-density scenarios use XPO | To balance cost, maintainability and multi-vendor supply |
| DCI | From several kilometers to 80 km | Traditional plug-in, coherent sinking (ZR/ZR+) | Priority should be given to distance, link budget, and standard compatibility, and integrated solutions are not applicable |
Model Selection Comparison
