800G is no longer a roadmap slide. The optics ship in volume, two form factors have settled into their ecosystems, and the IEEE standard behind the mainstream variants has been published since 2024. What has not settled is the buying process — because an 800G transceiver migration fails or succeeds on details that never mattered at 10G: which cage your platform chose, how much power the faceplate can feed, and whether your fibre plant can present sixteen fibres at an angle-polished connector.
This is the pre-order checklist. Six questions, roughly in the order they bite.
Who actually needs 800G today
Three groups, and it is worth being honest about which one you are in.
AI and GPU clusters. This is where the volume is. GPU-to-leaf and leaf-to-spine fabrics are being built at 400G and 800G per port as standard, and the traffic is real — training jobs saturate links in a way normal enterprise workloads never do.
Spine and DCI upgrades in high-growth data centres. If your spine is already dense 400G and still growing, 800G halves the port count for the same capacity, and the per-bit economics start to favour it.
Everyone else — probably not yet. If your estate runs on 10G, 25G and 100G and the links are not saturated, an 800G migration is a solution looking for a problem. The one reason to read on anyway: 800G decisions being made by hyperscalers are consuming the industry’s optics manufacturing capacity, which affects lead times and pricing on everything below it. Knowing the landscape helps even if you are not buying into it. (And if the estate you actually run is the older one, the more urgent read is transceiver end-of-life planning — the optics below 800G are the ones quietly going last-ship.)
Check 1 — the form factor was chosen for you
800G ships in two form factors: QSFP-DD800 et OSFP. You do not get to pick in the abstract — your switch or NIC picked for you — but you do need to know which, and what each choice implies.
QSFP-DD800 is the evolution of QSFP-DD, from the same MSA group. Its defining advantage is backwards compatibility: the cage accepts QSFP-DD 400G and, with the right configuration, QSFP28 100G modules. For a mixed estate that is a real operational benefit — one spares strategy can span three speed generations in the same cage family.
OSFP is a slightly larger module defined by the OSFP MSA, with more thermal headroom — which is why the GPU-fabric ecosystem standardised on it. It is not backwards compatible with QSFP modules; an adapter exists (Arista’s ADPT-O-Q-100G, for example, lets a 100G QSFP28 run in an OSFP port), but that is a per-port accessory, not a strategy.
And within OSFP there is a trap that catches real orders: the heatsink is part of the part number. Switch ports typically take a finned-top (integrated heatsink) module; NIC cages take a flat-top module that relies on the host’s riding heatsink. Same optic electrically, different physical variant — HPE, for instance, sells its 2x400G OSFP as a “Finned-Top” transceiver and the single-port NIC-side version as “Flat-Top”. Order the wrong one and it either will not seat or will not cool. Every 800G BOM should state finned or flat per line.
Check 2 — which “800G” you are actually buying
The IEEE standard for today’s mainstream 800G is 802.3df, published in 2024, which builds 800G from 8 lanes of 100G. That gives you the parallel variants: 800GBASE-SR8 (100 m on OM4, MPO-16), VR8 (50 m), DR8 (500 m on parallel single-mode, MPO-16) and DR8-2 (2 km).
Alongside those, the market sells duplex-friendly modules built as two 400G engines in one shell: 2×FR4 (2 km, two duplex LC pairs), 2×LR4 (10 km), 2×DR4, 2×PLR4. The “2x” in the part number is literal — the module can run as one 800G port or as two independent 400G ports, which is exactly why they dominate real deployments: they double density without forcing a forklift of the far end.
The next generation, 802.3dj, moves to 200G lanes — 800G over four lanes, and 1.6T over eight. It is still in draft. Which leads to the question to ask any supplier in 2026: is this 800G module 8×100G or 4×200G? The answer determines what it can break out to, and what it will interoperate with.
Check 3 — power and cooling, per port and per chassis
This is the budget people forget, because at 10G it never mattered.
A 100G QSFP28 SR4 draws roughly 2.5 W. A 400G FR4 draws 10–12 W. An 800G module typically draws 14–18 W depending on variant — coherent ZR-class parts more still. Populate a 32-port 800G switch and the optics alone are over half a kilowatt, before the switch itself; a row of them shifts the rack’s power and cooling maths visibly.
Three practical consequences. First, check your platform’s documentation for its supported power class per cage — not every port on every platform feeds maximum power. Second, airflow direction matters more at these densities; an optic run hot ages fast and errors first at the highest lane rates. Third, this is why OSFP exists — if your platform gave you the bigger cage, it did so to buy thermal margin. Use it: leave the blanking panels in, and treat a persistently hot module (DOM will tell you) as a fault in waiting.
Check 4 — breakout is where the money hides
Most 800G ports will spend their early life not running 800G at all. The economics of an 800G transceiver migration usually rest on breakout: one 800G cage feeding 2×400G today, or 8×100G into an existing estate.
Checks before you order:
- Does the platform support the breakout mode you want on that port? Breakout is a switch configuration feature, not just a cable. Some platforms restrict which ports, or how many, can be channelised.
- Does the module support it? A 2×FR4 gives you two 400G duplex links; an SR8 can channelise to 8×100G-SR-class lanes; a 2×DR4 can feed eight 100G-DR links via parallel fibre. These are different cables, different far ends and different part numbers.
- Is the far end coded and configured to match? A breakout link has three parts — module, cable plant, far-end optic — and the FEC must line up end to end. At 100G per lane, RS(544,514) FEC is inherent to the signalling; there is no “turn FEC off” escape hatch when a link misbehaves.
Check 5 — the fibre plant is part of the order
800G multimode and parallel single-mode variants arrive on MPO-16 APC connectors — sixteen fibres, angle-polished. Most installed MPO plant is MPO-12, and single-mode trunk MPO may be either polish. The duplex variants (2×FR4, 2×LR4) arrive on two LC pairs per module.
So the plant audit is simple but unskippable: what connectors do you have, what polish, what polarity method, and what will the new modules need? An 800G order that lands before the MPO-16 trunks do is a box of very expensive shelf stock. If the answer is “our plant is duplex LC everywhere”, that is an argument for the 2×FR4/2×LR4 family — and it is a perfectly good one.
Check 6 — where 1.6T actually is
1.6T modules exist — OSFP with 200G lanes (sometimes called OSFP224), and QSFP-DD1600 on the same trajectory — and first-wave shipments into AI fabrics are underway in 2026. But the IEEE 802.3dj standard behind 200G-lane Ethernet is not finished, and early modules track drafts and MSA specifications.
For most buyers the practical advice is short: do not pay for “1.6T-ready” without asking what it means. If a platform claims it, ask whether the cages are wired for 200G lanes, which FEC and which draft the vendor is tracking, and what the upgrade actually requires. If you are not building GPU fabrics, 1.6T is a 2027-onwards conversation — and by then it will be a standard, not a bet.
A sensible migration sequence
Pulling it together into an order of operations:
- Audit what you have — platforms, cage types, fibre plant (connector, polish, polarity), and current port utilisation. This is a day of work and it de-risks everything after it.
- Let the platform pick the form factor, then standardise your variants: one multimode and one single-mode 800G part where possible, chosen so their breakout modes match your actual far ends.
- Fix the fibre plant first. Trunks and cassettes have longer lead times than optics.
- Pilot one link per variant — module, cable, far end, under traffic — before the volume order. Check per-lane DOM readings while it runs; a marginal lane shows up in minutes.
- Stage the rollout spine-first, keeping breakout as the bridge to the parts of the estate that are not moving yet.
- Size spares at order time, not after the first failure — and buy them coded for the platforms they will actually go into.
Check compatibility before you order anything
Every Carritech Optics transceiver is cross-referenced against the OEM part numbers it replaces, tested in real hardware before it ships, and covered by a lifetime warranty. You can check a part number — or search by the hardware platform you are buying for — in about ten seconds, with no form and no sales conversation:
Check your part number or platform →
And if you are planning a migration and would rather hand over the whole bill of materials: send it. We will cross-reference it line by line — speeds, form factors, breakout plans and all — and come back with what works, what to watch, and what it costs. Talk to us.
