A 100G-to-400G upgrade rarely requires swapping every switch, transceiver, and fiber at once. In most brownfield networks the practical move is to upgrade the congested layer first, keep suitable 100G QSFP28 links in service, and replace only the components that would otherwise become a bottleneck. This guide shows how to make that decision as an engineering exercise rather than a transceiver purchase.
Before you commit budget, five questions decide almost everything: is 100G creating a measurable constraint; which existing assets are genuinely reusable; which 400G optical architecture matches your current fiber plant; does the target platform support the speed, breakout, and FEC modes you need; and can the change be tested and rolled back safely? The rest of this article works through each in turn, adds a scenario-to-solution matrix, a worked link budget, a computable TCO model, and a representative migration walkthrough with the failure modes teams actually hit.
What You Can Reuse, Validate, or Must Replace
Every reliable migration plan sorts existing assets into three buckets. Use this as your first pass, then confirm each item against the sections below and your switch vendor's compatibility list.
| Asset | Usual verdict | What decides it |
|---|---|---|
| QSFP28 transceivers (SR4, CWDM4, LR4, DAC, AOC) | Validate, often reuse at 100G | Exact part number, vendor coding, power class, and whether the QSFP-DD port qualifies that module |
| Existing 100G switches | Reuse as-is (still 100G) | The ASIC and electrical interface set the rate; a new optic does not turn a 100G switch into a 400G switch |
| Duplex LC single-mode fiber | Validate | Fiber type, connector polish, insertion loss, distance, cleanliness, and the exact endpoint modules |
| MPO-12 multimode trunks (used for 100G SR4) | Usually replace or re-architect | SR4 uses 8 active fibers; 400G SR8 needs 16 and normally MPO-16 |
| Patch panels, monitoring, spares | Validate | Connector, polarity, loss budget, and whether tooling can read modern module telemetry |
Can You Run 400G and 100G From the Same Switch?
Yes, in many designs. A switch with QSFP-DD ports can operate native 400G modules in some ports while carrying existing 100G QSFP28 optics in others, which lets you replace a constrained spine without immediately touching every leaf switch or server link. But "it fits" and "it works" are two different tests, and both have to pass.
Form-factor compatibility means a QSFP28 module physically seats in a QSFP-DD cage. The QSFP-DD Multi-Source Agreement designed the cage to remain backward compatible with the QSFP family, so no mechanical adapter is required. OSFP platforms can also accept QSFP modules, but through an adapter rather than native compatibility, and the port still has to be configured to the right speed.
Platform compatibility is the harder test: the switch ASIC, network operating system, firmware, port group, and transceiver policy all have to support that module at the required speed. A mechanically valid module can still be rejected, run without full diagnostics, need a manual port-speed setting, or be unsupported in a particular port group. This is why the vendor hardware compatibility list belongs in your design, not in troubleshooting. If you want the underlying detail on how the cage and its management model are defined, this overview of the QSFP-DD form factor is a useful companion.
The same "familiar shape ≠ compatible" caution applies to cabling: an LC or MPO connector you recognize does not guarantee the cable behind it supports the new optical interface.
Does Your 100G Network Actually Need 400G?
The availability of 400G is not, by itself, a reason to migrate. Base the decision on capacity, application growth, port availability, equipment lifecycle, and operational readiness together.
Look past average utilization. A link with a moderate daily average can still suffer congestion, queue buildup, packet loss, or latency during backup windows, distributed-storage operations, or east-west bursts. The evidence that justifies a migration comes from peak and percentile utilization, queue drops and buffer occupancy, congestion-related retransmissions, latency during busy periods, oversubscription at the leaf-spine boundary, and the rate at which demand is climbing. The case is strong when you have a repeatable constraint that traffic engineering, link aggregation, or a smaller topology change cannot fix.
Sometimes the pressure is ports, not bandwidth. A 400G spine can deliver a given aggregate capacity in fewer ports and free up connectivity for new leaf switches. Server refresh plans matter too: if most servers will stay on 25G or 50G for years, replacing every leaf immediately adds little value, and upgrading the spine first is more economical.
AI, storage, and east-west growth expose 100G bottlenecks earlier than traditional enterprise traffic. The right response is not to convert every link to 400G, but to find where the workload is constrained and upgrade the smallest layer that removes the bottleneck.
Staying on QSFP28 is a legitimate, defensible decision when the network still has headroom, current switches have useful life left, the next server refresh won't need faster access, the fiber plant is due for a separate redesign, the operations team isn't ready to support PAM4/FEC/module-management changes, or a 400G upgrade would create power and cooling problems without solving a current business need. Waiting is valid when it is backed by measurements and a capacity forecast.
QSFP-DD vs OSFP vs Staying on QSFP28
There is no universally superior form factor. The best choice depends on the target platform, your installed optics base, thermal headroom, and roadmap.
| Path | Main advantage | Main limitation | Typical fit |
|---|---|---|---|
| QSFP-DD | Native QSFP-family compatibility and high port density | Platform qualification and thermal limits still apply | Brownfield networks with reusable QSFP28 assets |
| OSFP | Larger body with more thermal design margin | QSFP modules need an adapter and platform validation | Platforms standardized on OSFP, including many AI/HPC designs |
| Stay on QSFP28 | No migration cost or operational change | Adds no capacity or higher-speed density | Networks without a current 100G constraint |
Choose QSFP-DD when you hold a large inventory of qualified QSFP28 modules or want to run 100G and 400G from the same switch generation. Choose OSFP when the switch ecosystem, thermal design, or long-term roadmap is built around it; if you plan to reuse QSFP optics behind an OSFP-to-QSFP adapter, qualify the adapter for the specific module, speed, airflow, latch access, and management interface. Stay on QSFP28 when the network meets its performance targets and the expected benefit of 400G is smaller than the cost, disruption, and remaining value of the current gear.
Scenario-to-Solution
Most teams don't need a form-factor lecture; they need to know what to do with the plant they already have. Use this to narrow options quickly, then validate against the actual platform.
| Current situation | Goal | Primary option to evaluate | Watch out for |
|---|---|---|---|
| 100G SR4 over MPO-12 OM4 | Native 400G on the same trunk | Bidirectional MMF option, or re-cable to MPO-16 for SR8, or move to single-mode | SR4 (8 fibers) cannot become SR8 (16 fibers) on the same MPO-12 |
| 100G LR4 over duplex OS2 | 400G over the same duplex pair | 400G FR4 or LR4 | Recheck the full link budget and endpoint interoperability |
| Keep 100G leaf, upgrade spine only | New 400G spine, existing leaf | QSFP28 in QSFP-DD at 100G, or 400G DR4 broken out to 4×100G | Port-group and breakout support on the new switch |
| New spine is OSFP, inventory is QSFP28 | Reuse existing modules | Vendor-qualified OSFP-to-QSFP adapter | Port must be configured to 100G; validate the exact module |
| In-rack 100G on DAC | Short 400G-to-4×100G links | Qualified 400G-to-4×100G DAC/AOC/AEC | Host lane rate, gearbox behavior, and FEC must match |
What You Can Actually Reuse
QSFP28 transceivers
QSFP28 modules can often run in a supported QSFP-DD port at 100G, but validate by part number rather than by family. SR4, CWDM4, LR4, DAC, and AOC products don't all behave identically; vendor coding, EEPROM behavior, power class, firmware, and diagnostics all affect acceptance. Weigh the commercial side too: a module may be technically compatible yet close to end-of-support or a poor fit for your new sparing strategy.
Duplex LC single-mode fiber
Existing duplex single-mode fiber is frequently reusable for 400G FR4 or LR4 when both endpoints, connector type, link loss, and optical budget line up. Before approving reuse, confirm fiber type, connector polish, total insertion loss, number of patch points, polarity, cleanliness, distance, and the exact transmitter/receiver specs. Two modules sharing a duplex LC face does not guarantee they interoperate. If you're unsure whether your plant is the right single-mode grade for the reach, this comparison of OS1 vs OS2 single-mode fiber is worth a look.
MPO-12 multimode trunks
A conventional 100G SR4 link uses four transmit and four receive fibers - eight active fibers inside an MPO-12. A 400G SR8 link uses eight transmit and eight receive fibers - sixteen active fibers, normally on MPO-16. A standard 100G SR4 MPO-12 channel therefore cannot be treated directly as a 400G SR8 channel. Your realistic options are new MPO-16 parallel multimode cabling, a supported 400G bidirectional multimode solution that fits the existing plant, or a move to single-mode. Which one wins depends on the switch and transceiver portfolio, not only on the cable already in the ground. Because reach is grade-dependent, it's worth confirming your fiber against the published OM1–OM5 multimode distance limits before committing.
Patch panels, monitoring, and spares
Panels are reusable if they support the required connector, polarity, density, and loss budget. Monitoring needs a closer look: modern high-speed modules expose telemetry through CMIS (the Common Management Interface Specification, maintained by the OIF), and your tools must correctly read module state, alarms, temperature, per-lane data, and diagnostics. Finally, redesign spares - a mixed network may need both 100G and 400G optics, several connector types, breakout assemblies, adapters, and cleaning kits.
Choosing the Right 400G Optical Architecture
"400G" is an aggregate Ethernet rate defined under the IEEE 802.3 Ethernet standards, not a single optical interface. Different 400G modules use different lane counts, fiber types, connectors, reaches, and breakout methods.
| Link need | Common 400G approach | Connector | Migration consideration |
|---|---|---|---|
| Short parallel MMF | SR8 | MPO-16 | Needs 16 active fibers |
| Parallel SMF, 4×100G breakout | DR4 | MPO-12 | Well suited to four single-lambda 100G links when supported |
| Duplex SMF, data-center/campus reach | FR4 | Duplex LC | May reuse suitable duplex SMF after budget validation |
| Longer duplex SMF | LR4 | Duplex LC | Validate distance, loss, host support, interoperability |
| Existing MMF plant | Bidirectional MMF option | LC or MPO (varies) | Must be supported at both endpoints; naming varies |
SR8 spreads 400G across eight channels for short, high-density parallel multimode links; don't patch an MPO-12 SR4 trunk into an MPO-16 SR8 module. DR4 uses four parallel single-mode lanes over MPO-12, and its key advantage is breaking one 400G port into four 100G single-lambda links when the modules and switch agree on the optical interface and FEC - but treat standard DR4-to-100GBASE-DR breakout as distinct from 4×100G FR/LR products, and confirm the exact SKU pairing rather than assuming any DR4 breaks out to any 100G DR/FR/LR. FR4 multiplexes wavelengths onto duplex single-mode, attractive when you already run duplex LC and the distance fits; it cuts fiber count but generally does not give four independent 100G paths. LR4 extends duplex single-mode reach; as with FR4, verify the specific module and platform rather than trusting the name.
Bidirectional multimode options are designed to preserve lower-fiber-count MMF, and can be genuinely useful - but they are not universal drop-ins for SR4 or SR8. Confirm the optical standard, wavelength scheme, connector polish, fiber grade, and the exact far-end module; if the concept is new to you, here's how bidirectional transceivers work.
Connecting New 400G Ports to Existing 100G Equipment
Mixed-speed links are the bridge between the old and new network.
400G DR4 to four 100G links
A qualified 400G DR4 or parallel single-mode module can feed four compatible 100G single-lambda modules through an MPO-to-four-duplex breakout - one 400G platform providing four independent 100G connections to existing switches, with the option to move to a native 400G link later. Validate the port's breakout support, the far-end 100G optical type, the FEC mode, MPO polarity, lane mapping, distance, and the switch configuration syntax. For the cabling side, this MPO breakout cable guide helps you match the assembly to the interface, and vendor-built MPO-to-LC breakout assemblies cover the common polarities.
DAC, AOC, and AEC for short links
For short runs, a switch may support 400G-to-4×100G DAC, AOC, or AEC assemblies - and they are not interchangeable. DAC is passive or active copper for short distances; AOC integrates optics into a fixed cable; AEC adds active electronics and may perform signal conditioning or rate conversion depending on design. Don't buy on connector shape and aggregate speed alone; confirm the host interface, lane rate, gearbox behavior, FEC, cable length, and qualified platforms.
FEC and lane mapping - where "correct" links still fail
PAM4 (four-level pulse-amplitude modulation, the signaling used across 400G) is error-prone enough that these links depend on FEC - Forward Error Correction, which repairs bit errors in flight. Both ends must use compatible FEC behavior, and it matters which side runs it and whether the module handles it internally. A link can stay down even when both modules are recognized, the fiber is connected correctly, and the nominal speeds appear to match. Check pre-FEC and post-FEC error counters during qualification, and confirm that lane mapping - which electrical lane carries which optical channel - survives reboots, firmware changes, and breakout reconfiguration.
Port-group limitations
Not every 400G port supports every breakout mode. Some ASICs configure ports in groups, so changing one port's lane rate can affect its neighbors; others restrict breakout to specific cages, transceiver types, or software releases. Treat the switch configuration guide and hardware compatibility tool as part of the optical design, not an afterthought.
A Representative Migration Walkthrough
The following is an illustrative brownfield scenario, not a specific customer project, but it mirrors how these migrations actually unfold - including the parts that go wrong.
Starting point. Two 100G spine switches, leaf uplinks running 100G SR4 over MPO-12 OM4 trunks. During nightly storage replication, spine uplinks saturate, queues build, and replication windows overrun. Average utilization looks fine; percentile and queue-drop data tell the real story.
Decision. The MPO-12 OM4 trunks can't carry 400G SR8 directly (8 fibers vs 16). Rather than re-cable the whole plant at once, the team specifies a new QSFP-DD spine and two parallel tactics: keep leaf uplinks at 100G by running existing QSFP28 optics in QSFP-DD ports, and add 400G spine-to-spine plus selected 400G DR4 ports broken out to 4×100G-DR where new single-mode is being pulled anyway.
Lab findings (the useful part). Four issues surfaced before production, each cheap to fix in a lab and expensive in a change window:
- A 400G-to-4×100G link came up physically but stayed down - the far-end 100G module defaulted to a different FEC mode than the breakout expected.
- After a warm reboot, one port group lost its breakout configuration and reverted to a single 400G port, taking four leaf links with it.
- An MPO trunk with the wrong polarity brought up only half the lanes; the module reported present and healthy while several channels stayed dark.
- A reused QSFP28 module seated and passed traffic but exposed no temperature or DOM telemetry, because that part number wasn't on the platform's qualified list.
Outcome. With FEC modes pinned, breakout configuration made persistent across reboot, polarity corrected, and the unqualified module swapped for a listed equivalent, the spine cut over in a single window. Leaf switches stayed in place, and the access layer was scheduled to move only when its own refresh cycle arrived.
FAQ
Q: Can QSFP28 modules run in QSFP-DD ports?
A: A QSFP-DD cage accepts QSFP28 modules without a mechanical adapter, but real operation depends on the ASIC, software, port configuration, and vendor qualification. Check the compatibility list and test the exact module first.
Q: Can an MPO-12 cable used for 100G SR4 support 400G SR8?
A: Not as a direct swap. 100G SR4 uses eight active fibers in MPO-12; 400G SR8 uses sixteen and normally MPO-16. You need new MPO-16 cabling, a supported bidirectional MMF option, or a move to single-mode.
Q: Can existing duplex LC fiber support 400G?
A: Potentially. FR4 and LR4 commonly use duplex LC single-mode, but reuse depends on fiber grade, polish, distance, loss budget, cleanliness, and interoperability. Validate the whole channel, not just the connector.
Q: Do all 400G ports support 4×100G breakout?
A: No. Breakout depends on the ASIC, NOS, port group, transceiver or cable, lane mapping, and FEC. Verify the exact configuration before ordering.
Q: Should you skip 400G and move straight to 800G?
A: Only when the target architecture, equipment lifecycle, capacity forecast, and platform economics support it. Compare available platforms, optics, cabling, operational readiness, and the capacity you actually need over the evaluation period - don't adopt or skip a speed simply because it's newer.
Bottom Line
A successful 100G-to-400G migration starts with an inventory and a measurable requirement, not a transceiver order. For a brownfield QSFP28 network, the lowest-risk path is usually to identify the constrained layer, choose a platform that supports the 100G and 400G modes you need, map each existing link to a specific 400G optic, validate optics, cabling, breakout, FEC, thermals, and monitoring in a lab, upgrade the constrained layer first, and migrate the edge on its own lifecycle. The goal is not to reuse everything - it's to reuse what stays technically suitable and replace what would otherwise become a future capacity, reliability, or operational problem.
