Co-packaged optics vs pluggable transceivers 800G: what comes next for your DC cabling

800G and 1.6T are the cabling speeds dominating the data center roadmap from 2026 to 2030. Two technology routes compete: pluggable transceivers (the traditional route with QSFP-DD or OSFP modules that are hot-swappable) and co-packaged optics (CPO), which integrates the optical functions directly into the switch ASIC. Each route has different impact on port density, electrical consumption, latency, and total CAPEX. This article breaks down the technical criteria that matter for a Mexican data center planning to upgrade to 800G within 24 to 36 months.

What changes with 800G (mandatory context)

The IEEE 802.3df standard (approved in 2025) defines 800G and 1.6T Ethernet using four media types: copper (up to 2 meters over twinax DAC), multimode OM4 and OM5 (up to 50 meters), single-mode (up to 10 kilometers with WDM), and Active Electrical Cable (AEC). The dominant transceiver today is the QSFP-DD800 and OSFP800 pluggable, both modular replacements for previous generations (QSFP28, QSFP56). Electrical consumption per 800G port is 14 to 22 W with current pluggable transceivers (depending on reach and PAM4 200G-per-lambda modulation), comparable to the sum of 8 previous-generation 100G ports.

The fundamental technical change is that 800G ports are four lanes of 200G PAM4 over single-mode or multimode fiber, not eight lanes of 100G NRZ as in previous generations. This reduces the number of optical fibers needed per link but increases the module’s complexity. Operators with Tier III-IV sites in Mexico are evaluating the migration between 2026 and 2028 according to the hyperscaler roadmap (Broadcom Tomahawk, Cisco Silicon One, Marvell, Nvidia Quantum).

What CPO is and where it changes the architecture

Co-packaged optics (CPO) integrates the optical components directly on the substrate of the switch or router ASIC, eliminating the front panel where pluggable transceivers connect. Optical fibers enter the ASIC directly via edge connectors. The main saving is electrical: between 30% and 50% less consumption per port, because the electrical signal path between ASIC and transceiver module is removed. The trade-off is that optics now belong to the switch: if a fiber or optical component fails, the entire switch is replaced, or technical staff with specialized CPO skills is required.

CPO has been in commercial development since 2024. The first operational generations are for 12.8T to 25.6T per chip (Broadcom Tomahawk 5, Cisco Silicon One G200, Marvell Spad). The 2026 to 2028 roadmap targets 51.2T to 102.4T per chip with CPO. For Mexican data centers, the decision between pluggable and CPO is a 3-to-5-year decision.

The 5 operational differences

Five points where the two routes diverge:

1. Port density: pluggable 800G (QSFP-DD/OSFP) allows 36 to 72 800G ports in a single 1U switch; CPO allows 2 to 3 times higher densities because the transceiver socket is eliminated. For data centers with limited space, CPO wins on density.
2. Electrical consumption: CPO consumes between 30% and 50% less per port because it shortens the electrical path. For a site with 200 switches and 14,400 800G ports, the annual energy savings are 200 to 400 MWh (cost under CFE GDMTO: USD 35,000 to USD 70,000/year per 200 switches, depending on the tariff mix).
3. Maintenance: a pluggable transceiver is hot-swapped in 30 to 60 seconds without affecting other ports. In CPO, an optical failure can require replacing the switch or specialized staff for substrate-level repair (microscope tools, laser alignment, etc.). An operator with current maintenance staff benefits from pluggable; an operator that outsources maintenance may choose CPO.
4. Latency: CPO reduces latency by 5 to 15 ns per port because it eliminates the intermediate electrical-to-optical transition. For trading applications, HPC, or distributed AI compute with low-latency networks, CPO adds real performance.
5. Total cost: pluggable 800G transceiver costs between USD 1,200 and USD 4,000 per port (depending on reach and modulation); CPO encapsulates the optical cost in the switch, adding between 15% and 25% to the ASIC price, but eliminating USD 800 to USD 1,500 per port in transceivers. For sites planning 10,000+ 800G ports, the balance tilts to CPO in CAPEX; for sites with 1,000 ports, pluggable remains more economical.

When it IS worth planning the migration to CPO

Three cases where CPO makes sense in the 2026 to 2030 roadmap:

1. New sites starting with 800G (greenfield at 25.6T per chip or above): if you are going to buy new switches, evaluating CPO from the initial specification reduces total cost and leaves margin for future upgrades. Hyperscale operators in Mexico (if AWS, Azure, Google arrive at Tier III) typically choose CPO for greenfields.
2. AI factories with low-latency GPU networks: distributed AI workloads (training of large models, parallel fine-tuning) require 5 to 15 ns less latency per port, and CPO delivers this natively. AI sites in Mexico (KIO in Querétaro, Ascenty under construction) are natural candidates for CPO from the start.
3. Operations with on-site specialized maintenance staff: if you have technical staff capable of substrate-level repair (microscope tools, laser alignment) or contracts with manufacturers for in-situ support, CPO is viable. Operators without these capabilities should stay on pluggable.

When to wait with pluggable

Three cases where pluggable is the right decision in 2026:

1. Sites with existing 400G switches and gradual migration. Pluggable 800G QSFP-DD/OSFP is backwards compatible with 400G sites via existing twinax or fiber cabling. CPO requires a new ASIC incompatible with 400G switches.
2. Operators that outsource technical maintenance (manufacturer switch vendor) and have no in-house staff. Pluggable allows transceiver replacement without opening the switch; CPO requires return-to-vendor.
3. Modest port density (up to 72 800G ports per switch). Pluggable in QSFP-DD800/OSFP800 form factor already covers this density with good price-performance options. CPO brings no advantage until densities above 100 ports per switch.

The 2026 Mexican data point

Mexican Tier III-IV operators are in the evaluation phase, not adoption. KIO Networks has published a 400G use case with QSFP-DD transceivers at some sites; Ascenty has deployed InfiniBand NDR (GPU interconnect links, not 800G Ethernet — different technologies). Triara and ODATA are evaluating for new sites. CPO is a 2027-2028 decision for any Mexican Tier III-IV operator; until then, pluggable 800G QSFP-DD/OSFP is the operational choice. For on-prem enterprise data centers (not colocation), the practical decision in 2026 remains pluggable 400G with migration to pluggable 800G within 24 to 36 months.

Sources

1. TIA: Telecommunications Industry Association (TIA-942-B-2023 for data center cabling) — https://tiaonline.org/
2. IEEE 802.3 Working Group: Ethernet standards including 802.3df for 800G and 1.6T — https://www.ieee802.org/3/
3. IEEE 802.3df Task Force: technical specification for 800G — https://www.ieee802.org/3/df/
4. TIA Standards: current list of standards applicable to cabling — https://www.tiaonline.org/standards/
5. BICSI: standards for telecommunications infrastructure in buildings and data centers — https://www.bicsi.org/