800G and 1.6T Ethernet: what your data center cabling needs before 2027

Your next spine switch is going to have 64 ports of 800G, not 32 ports of 400G. That decision was already made at Broadcom, NVIDIA, and Cisco — and published between late 2025 and mid-2026. The question that remains open for your data center in Mexico is whether your current cabling plant can carry that transition, or whether you need to recable before 2027. If your last major migration was to 100G or 400G FR4 over OM4, this article is going to sting a bit: you are going to have to re-evaluate fiber, MPO connectors, topology, and even your patch racks. The good news: if you plan it out with time, the cost per Gbps of 800G is already within about 5% of 400G, and the cost per Gbps of 1.6T will follow the same curve as previous generations.

Why 400G is no longer enough for AI workloads

The number you need to remember is 40–70x. Training workloads for large models (LLM, foundation models) generate between 40 and 70 times more east-west traffic than a traditional compute workload of equivalent size. On a classic spine-leaf architecture with 64 ports of 400G on the spine, that means 25.6 Tbps of bisectional bandwidth — and a modern training job saturates that during all-reduce operations. When you double the port from 400G to 800G without adding switches, you double the bisectional bandwidth without doubling rack space or cabling. That is exactly what hyperscalers and large colos are doing today. The 2026 projection for global 800G transceiver shipments exceeds 40 million units, driven almost entirely by AI infrastructure buildouts at hyperscalers and large colos.

800G: two form factors, one decision

This is where it gets complicated. 800G comes in two main physical formats, and the choice you make defines your roadmap to 1.6T:

Practical decision rule for a data center in Mexico:

Greenfield AI cluster with NVIDIA H100/H200/GB200 GPUs: OSFP (integrated heat sink handles 15–18W per port better).

Cloud DC upgrade with existing QSFP28 cabling: QSFP-DD800 (lets you reuse existing MPO-12/APC fiber and migrate gradually).

No hyperscaler or major colo has picked OSFP as an absolute standard. NVIDIA Spectrum-4 and Arista 7800R3 ship in both versions. Always verify the exact SKU of the switch before buying optics.

1.6T: the roadmap that is no longer vaporware

1.6T is already in early commercial rollout at hyperscalers since mid-2026. It uses 8 lanes of 200G PAM4 (224 Gb/s electrical class). The required switch ASICs already exist: Broadcom Tomahawk 6 (102.4 Tbps), Cisco Silicon One G300 (25.6–51.2 Tbps), NVIDIA Spectrum-6, Marvell Teralynx 10. The optical module comes in three main formats:

The QSFP-DD1600 vs OSFP1600 decision is mainly thermal and backward compatibility. QSFP-DD1600 wins on compatibility (the cage accepts QSFP-DD800, QSFP-DD 400G, QSFP28 100G). OSFP1600/OSFP-XD wins on thermal headroom and coherent optics density.

If your operator tells you “you need 1.6T in 2027”, the conversation with them should be about coherent vs direct-detect and link distance, not about the format.

The IEEE 802.3df standard and why it matters

IEEE P802.3df was published as a standard in 2024 and covers 200 GbE, 400 GbE, 800 GbE, and 1.6 TbE over the same physical architecture (8 lanes with variable lane speed). The key for your RFP:

200 GbE per lane: each lane runs at 200G PAM4 over single-mode fiber (SMF). Reaches: 500m (DR1), 2km (FR1), 10km (LR1).

400 GbE: 4 lanes of 100G PAM4 or 2 lanes of 200G PAM4. Same reaches as 200G with a different lane count factor.

800 GbE: 8 lanes of 100G PAM4 or 4 lanes of 200G PAM4. SMF at 500m/2km/10km available today.

1.6 TbE: 8 lanes of 200G PAM4 (OSFP1600) or 8 lanes of 224G PAM4 (electrical class). SMF 500m/2km/10km.

Ethernet Alliance 2025–2030 projections: yearly switch port volume moves from 150–160M in 2025 to 300M+ in 2030. Switch size goes from 51.2T to 204.8T+. Power per rack goes from 100 kW to 1 MW. If your DC is not designing for those numbers, it will be obsolete before 2030.

What your structured cabling needs before 2027

Here is the operational impact. If your current cabling plant is OM3 or OM4 with MPO-12, you have a problem. If it is OM5 with MPO-16, you have the correct base for 800G and 1.6T.

Fiber: OM5 or OS2, not OM3 or OM4

OM3 and OM4 are already at the end of their distance range for 400G parallel optics, and they do not support 800G SWDM8. OM5 wideband multimode supports 400G, 800G, and 1.6T without recabling — that is why hyperscalers specify OM5 for new greenfields. If you are on SMF (OS2), the critical path to 10km is already covered; you only need to validate MPO-16 vs MPO-12.

Connectors: MPO-16 vs MPO-12

The connector roadmap is moving to Base-16 (16 fibers, MPO-16). Today 800G DR8 uses dual MPO-12 (16 fibers total). 1.6T will consolidate to native MPO-16 to reduce connector count and simplify patch panels. Specifying MPO-16 APC now avoids an expensive re-patch in 2–3 years.

Topology: spine-leaf with oversubscription ≤ 3:1

East-west AI workloads saturate any topology with high oversubscription. Your spine-leaf design should target 1:1 to 3:1 oversubscription for AI fabrics. If your current DC is at 6:1 or worse (typical of pre-AI designs), upgrading to 800G will not save you — you need to rebuild the topology.

High-density patch panels and racks

VSFF (Very Small Form Factor) connectors like MMC and SN-MT are entering hyperscale deployments. If your current patch panel is standard MPO-12, you will need to migrate to higher-density panels to handle the doubled fiber count per rack that 1.6T implies.

The checklist for your 2026–2027 RFP

Before signing the purchase of your next spine switch, require from your vendor:

ASIC and cage details: Tomahawk 5/6, Spectrum-4/6, Silicon One G200/G300, Marvell Teralynx — confirm the exact model and whether it supports native 800G or requires breakout.

Certified optics platform: list of OEM-certified optics for that specific switch. Do not accept “compatible” without proof in production.

Form factor and backward compatibility: confirm whether the cage accepts QSFP-DD800, OSFP800, or both. Confirm also whether it will accept 1.6T in the same cage with optics upgrade (without changing hardware).

1.6T roadmap: does the manufacturer have 1.6T modules in development for that switch? When? OSFP1600 or QSFP-DD1600?

Phased migration plan: if you have 400G installed, the vendor must have a playbook such as spine-first with 2×400G breakout to 800G, then native 800G.

Are you quoting your upgrade to 800G/1.6T?

If your data center is evaluating an upgrade to 800G now, or a greenfield that has to reach 1.6T in 2027–2028, Noxtel designs the structured cabling specification (OM5/OS2 + MPO-16 APC), validates optics-vs-switch compatibility against the preselected vendors, and delivers a phased migration plan that fits your budget.

Quote your cabling upgrade to 800G/1.6T → [link to leads form]

Sources

Vitex — “800G OSFP vs QSFP-DD Deployment Guide” (Aug 2026). https://vitextech.com/blogs/blog/osfp-vs-qsfp-dd-800g-deployment-guide

Wolontek — “OSFP vs QSFP-DD vs SFP-DD: The Ultimate 400G/800G Guide (2026)”. https://wolontek.com/osfp-vs-qsfp-dd-vs-sfp-dd-guide

Fibermall — “1.6T OSFP: The Complete Guide to Next-Generation Data Center Connectivity”. https://fibermall.com/blog/1-6t-osfp-complete-guide.htm

AscentOptics — “OSFP 1.6T Optical Transceiver: Complete Technical Guide” (May 2026). https://ascentoptics.com/blog/osfp-1-6t-optical-transceiver-guide

Fibermall — “1.6T QSFP Roadmap: QSFP-DD1600, OSFP, and the Path to 1.6T Ethernet”. https://fibermall.com/blog/1-6t-qsfp-roadmap-qsfp-dd1600-guide.htm

IEEE P802.3df Task Force — “End-to-end FEC for 200G per lane based 200GbE, 400GbE, 800GbE and 1.6TbE” (Nov 2022). https://grouper.ieee.org/groups/802/3/df/public/22_11/lu_3df_01a_2211.pdf