Heading to 800G: Projects Fail on the Cabling, Not the Technology
3 min read
2026 is the year data centres shift up a gear: 400G remains the installed base for server access, 800G becomes the reference standard for fabric interconnects, especially in high-density GPU AI clusters, and 1.6T is appearing on the horizon. The story always gets told through transceivers and switches. But those who build and cable networks know better: most 800G projects don’t fail on the technology, they fail on cabling planning — optical choices, breakout strategy and fibre management decided too late. This is true at hyperscale, and it is true, at a smaller scale, for the corporate data hall that needs to last the next ten years.
The three variables that decide everything (and must be decided first)
- Fibre type: OM4, OM5 or single-mode. On multimode, OM4 and OM5 support 400G and 800G, but with a 100-metre limit — beyond that, you either move to single-mode or rethink the topology. The choice isn’t “let’s take the best one”: it’s about sizing the hall’s actual distances before laying cable, because changing fibre type afterwards means redoing the runs.
- Fibre density: 8, 12 or 16-fibre MPO. This is where the costly mistake lies. The 8-fibre multimode infrastructure that supports 400G today needs double the fibre to reach 800G; 16-fibre connectivity (MPO-16) provides the right density for the next step up. Those who install today thinking only of tomorrow’s 400G end up recabling; those who design ducting and panels for the target density save themselves an entire follow-up job.
- The optical budget, in decibels. Margins are tightening: the allowed insertion loss is falling (roughly ~1.8-1.9 dB for 800GBASE-SR8), which makes every connector and every splice a critical factor. At these speeds a dirty connector or a poor cross-connect doesn’t degrade the link: it takes it down. That’s why OTDR certification of every run stops being a formality and becomes the difference between a hall that powers up cleanly and one plagued by untraceable intermittent errors.
Why it matters for the corporate data hall too
You don’t need to be a hyperscaler. With the boom in data centre investment and the adoption of on-premise AI, even mid-sized corporate data halls now face the same choices in miniature: internal backbones that must withstand future upgrades, breakout from fabric to servers, fibre management that doesn’t turn into an unmanageable tangle by the third expansion. Anyone renovating a data centre in 2026 and cabling it “for today” will be recabling in 2028. The rule is the same at every scale: the fibre outlasts the equipment, so you design for the speed that’s coming, not the one you need right now.
The checklist before laying a single metre
- Map the hall’s actual distances: above or below 100 metres decides multimode vs single-mode.
- Choose the MPO density for the target, not the present: if the horizon is 800G/1.6T, 16 fibres avoids recabling.
- Define the breakout strategy (fabric → server) before installation: it’s the choice that, if postponed, causes projects to fail.
- Set the optical budget and insist on documented bidirectional testing of every run and cross-connect.
- Design fibre management: routing, bend radii, labelling, room for expansion. The hall is only as strong as its weakest link — and by the third addition, the tangle is the weakest link.
800G technology is mature and available. What is in short supply is installation carried out with the right horizon in mind: it’s the work we do, to the highest standard and with test results in hand, from site survey to documented handover.
Are you planning or renovating a data hall and want to cable it for the next ten years, not just for today? Talk to a technician: free site survey and quote.