Data center GPU network optics and cabling: the AI cluster field guide

The optics and cabling layer, and why an AI cluster is a cabling project

An AI/GPU cluster is as much a cabling project as a compute project. The GPUs only earn their cost when they train as one machine, and that means every GPU has to talk to every other GPU over a high-speed, low-latency network called the back-end fabric. That fabric runs at 400G, 800G, and now the first 1.6T links, almost all of it over fiber with pluggable optics plugged into the switches and the network cards. On top of the back-end there is a front-end network for storage and the outside world, and a separate management network. Put together, that is a very large count of high-speed links, and the links are the part that comes up last and breaks first.

The compute and the topology are covered elsewhere. The spine-leaf and rail-optimized fabric design, the switch tiers, and the oversubscription are in the network-architecture guide, and the single-mode versus multimode decision and the OM and OS grades are in the fiber-types guide. This guide is the layer in between: the transceivers, the MPO fiber, the loss budget, and the install discipline that turn a fabric design on paper into links that pass light.

Get this layer wrong and the schedule and the reliability both suffer. A dirty connector, a flipped polarity, an optic matched to the wrong reach, or a link that nobody tested does not announce itself at install. It shows up at bring-up when the fabric will not train, and now you are chasing one bad link out of thousands in a powered room. The work that prevents that is unglamorous: match the optic and fiber to the speed and reach, clean and test every connector to the loss budget, and run structured cabling with the polarity sorted out before the first trunk is pulled.

The back-end fabric and the cabling volume

The back-end fabric is the part that makes an AI cluster a cabling job at a scale most data center work never sees. Every GPU connects to the fabric at high speed, and the high-end nodes carry several fabric ports each, so the link count climbs fast with the GPU count. A single GPU server commonly needs eight to sixteen fiber connections just for the compute fabric, before the storage and management networks are counted. Multiply that across hundreds or thousands of GPUs and the fabric alone is tens of thousands of fiber connections.

The published platform numbers make the volume concrete. NVIDIA's recent rack-scale platforms have been cited at on the order of 144 high-speed optical modules per rack on the GPU side, and a leaf switch can carry 64 or 72 high-speed ports, each of which may break out further. Every one of those ports is a transceiver at one end, a length of fiber, a patch, and a transceiver at the other end, and every junction is a connector that has to be clean and within loss budget. The exact counts depend on the platform, the topology, and the oversubscription the design chose, so size the cabling scope off the actual port map, not a rule of thumb.

Treat the link count as the project driver it is. The architecture guide covers how spine-leaf and rail-optimized designs set the port counts and the oversubscription ratio. The number to carry away here is that the cabling volume of an AI cluster is large enough that the optics and fiber become a major line item, a long-lead procurement, and a schedule risk in their own right.

Front-end, back-end, and management: the different networks

An AI cluster runs more than one network, and they do different jobs at different speeds, so the optics and cabling differ by network. The back-end GPU fabric is the big, fast, latency-sensitive one. It carries the traffic between GPUs during training, it runs at the highest rates in the building, and it dominates the link count. It is built as a dedicated fabric, often InfiniBand or a tuned Ethernet, and it is the network the rest of this guide is mostly about.

The front-end network connects the GPU nodes to storage, to the provisioning systems, and out to the rest of the data center and the world. It still runs fast, often 100G to 400G per node, but it does not carry the all-to-all training traffic, so its link count and its latency demands are lower. The management network is lower still, often copper for in-band and out-of-band control, and it is the one people under-plan because it is not the headline number.

Keep the three separate in the design and the labeling. Mixing a management run into the fabric trunking, or pulling front-end and back-end on the same poorly labeled bundle, is how a tech patches the wrong port at bring-up. The networks have different speeds, different optics, and different criticality, and the as-built has to make that obvious to the next person in the room.

What speeds do AI cluster links run at?

AI cluster fabrics run at 400G and 800G today, with the first 1.6T links arriving in 2026, and the per-link rate is built up from electrical and optical lanes rather than one fast signal. An 800G link, for example, is commonly eight lanes of 100G, written 8x100G. The jump to 1.6T keeps eight lanes and roughly doubles the lane rate toward 200G per lane. That lane structure is why a single high-speed port can break out into several lower-rate links, and why the fiber count per port is what it is.

The progression is fast and it is set by IEEE 802.3 for Ethernet and by the InfiniBand roadmap for the dedicated fabrics. 400G was the AI default a generation ago, 800G is the default for new high-end builds now, and 1.6T is entering production for the largest training clusters first. Each step roughly doubles the bandwidth per port and tightens the optical and cleanliness margins that come with it.

Design for the rate the cluster will actually run, and confirm it against the switch and NIC platform and the IEEE or InfiniBand specification for that rate, because the lane structure and the supported optic types are defined there. Do not assume a fiber plant pulled for one rate carries the next one. The reach a given fiber supports shrinks as the rate climbs, which is the single most common way a cabling plant ages out before the building does.

What is a transceiver, and which form factors does the fabric use?

A transceiver, or pluggable optic, is the module that converts the switch's electrical signal into light for the fiber and back again. It plugs into a cage on the switch or the network card, the fiber connects to its front, and it does the work of putting the lanes onto the glass. The fabric is full of them, two per link, and they are among the more expensive and more power-hungry parts of the cabling scope.

Two form factors dominate the high-speed fabric. QSFP-DD and OSFP are the cages you will see on 400G and 800G switches and NICs. OSFP is slightly larger and was designed with more thermal headroom, which matters as module power climbs; QSFP-DD keeps backward compatibility with the older QSFP family. Which one a port uses is set by the switch and NIC platform, and the module has to match the cage, the host, and the link partner, so the optic is selected against the platform's supported list, not just the rate.

Cost and power are real and they add up. High-speed modules are not cheap, and they draw real watts each, so the transceiver count drives both the bill of materials and the thermal load. Buy them against the platform compatibility matrix, because a module that is electrically the right rate can still be unsupported on a given host, and an unsupported optic that mostly works is a bring-up problem waiting for the worst moment.

Which optic do you use for which distance?

You match the optic's reach type to the link distance, and the named reach classes tell you roughly how far each will go. SR, the short-reach type, uses multimode fiber and 850 nm lasers for the shortest links, commonly inside a row at up to around 100 m depending on the rate and fiber grade. DR and FR are single-mode types at 1310 nm for longer links: DR is the shorter single-mode reach, often around 500 m, and FR reaches farther still, into the low kilometers. LR goes farther again where a long run demands it.

The reach figures move with the rate and the standard, so treat the numbers as the IEEE-defined ballpark and confirm the exact reach for the specific optic against the manufacturer datasheet and the IEEE 802.3 clause for that type. An 800GBASE-SR8 module is built for short multimode hops; an 800GBASE-DR8 for the single-mode runs across a hall. The point is to size the optic to the longest link it has to serve, not the average, and not the shortest.

The expensive error is matching the optic to the wrong distance in either direction. Put a short-reach multimode optic on a run that exceeds its reach and the link is marginal or dead, and the fix is new optics and possibly new fiber in a live room. Put an expensive long-reach single-mode optic on a 3 m in-rack hop and you have paid for reach and power you will never use. The reach class is a budget decision as much as a physics one.

Single-mode and multimode at high speed

The single-mode versus multimode choice is covered in depth in the fiber-types guide, but the part that matters for high-speed AI links is the shift that happens as the rate climbs. Multimode with short-reach optics is cheaper on the optic and fine for the short in-row hops, but the distance it supports shrinks at every rate step. The reach that was comfortable at 100G is tight at 400G and tighter at 800G, so a multimode plant that served an older build may not carry the new one over the same runs.

Single-mode does not shrink the same way. At 400G and above, more of the fabric is moving to single-mode with DR and FR optics, because the fiber carries the higher rates over the distances an AI hall actually spans, and because it gives the plant a longer life across rate steps. The trade is the optic cost, which has historically run higher for single-mode, though that gap has narrowed at the high rates.

The practical call is to use multimode short-reach only where the runs are genuinely short and the rate fits its reach, and to default to single-mode for anything that has to cross the room or survive the next rate step. The fiber-types guide has the grade-by-grade detail; the discipline here is to pick the glass for where the speeds are heading, because re-pulling fiber in a populated AI hall is the outcome you are trying to avoid.

What is an MPO connector, and why the fabric is full of them

An MPO connector is a single ferrule that carries many fibers at once, which is what makes the parallel optics in an AI fabric practical. MTP is a high-performance brand of MPO; the terms get used together. Instead of a separate duplex connection per fiber pair, one MPO ferrule lands 8, 12, or 16 fibers in a single push-on connector, which is the only sane way to handle the fiber count an AI cluster carries.

The fiber count in the connector follows the optic. MPO-12 has been the long-standing workhorse for 40G and 100G parallel optics, with eight of its twelve fibers used. MPO-8 appears where only eight fibers are needed. MPO-16 has become the connector for the 8-lane parallel optics behind 400G SR8 and 800G SR8, because eight transmit and eight receive fibers fit a single 16-fiber row without the waste of pairing two 12-fiber connectors. The optic's lane and fiber map decides which connector the link uses, so confirm it against the transceiver datasheet.

MPO is high density and it is unforgiving. The ferrule is small, the fibers are exposed in a flat row, and a single dirty or damaged fiber in the array fails the whole link. That density is the reason cleanliness and inspection, covered below, matter more on an MPO plant than on the duplex connections most techs cut their teeth on.

MPO polarity: get it right or the link does not come up

MPO polarity is the rule that makes sure each transmit fiber at one end lands on the right receive fiber at the other end, and it is the classic way an MPO link fails to come up despite clean connectors and good optics. Because the connector carries a whole array, the order the fibers run through the trunk, the cassettes, and the patch cords has to be managed end to end, or transmit meets transmit and nothing links.

TIA defines three methods, commonly called polarity A, B, and C. Type A is straight-through, with the fiber order preserved. Type B is flipped, with the order reversed end to end. Type C flips fibers in pairs. The whole channel, trunk, cassette, and jumper, has to follow one consistent method, and parallel optics commonly use the flipped Type B method so the array maps correctly. The detail is less important than the discipline: pick the method for the optic, document it, and do not mix components from different polarity schemes in one channel.

This is where bring-up time disappears. A polarity mismatch passes a continuity check and a clean inspection, then refuses to link, and a tech burns hours swapping optics and re-cleaning before someone thinks to check the polarity of the channel. Sort the polarity in the design, label it on the components, and verify it as part of certification, because finding it at install costs minutes and finding it at bring-up costs the schedule.

Breakout cabling

Breakout cabling splits one high-speed port into several lower-rate links, and it is how an AI fabric matches a high-density switch to the nodes below it. An 800G port can break out to eight 100G links or two 400G links; a 400G port to four 100G links. On the fiber side that is one MPO trunk fanning out to several duplex or smaller-MPO connections, done with a breakout cable or a cassette in a patch panel.

There are two ways to build it. Point-to-point breakout cables run a single assembly from the switch port straight to the endpoints, which is cheap and fast for a small count but turns into an unmanageable tangle at fabric scale. Structured breakout puts the fan-out in cassettes at a patch field, so the trunk runs panel to panel and the breakout happens at a documented, serviceable point. For anything beyond a handful of links, the structured approach is what keeps the plant maintainable.

Match the breakout to the design's port map and the optic. A breakout assembly carries an assumed polarity and fiber map, and the wrong assembly is another way to get transmit and receive crossed. Confirm the breakout scheme against the switch configuration and the optic's lane map before you order at volume, because a breakout mistake repeats across every port that uses it.

DAC, AOC, or transceivers and fiber?

For the short links, you choose between direct-attach copper, active optical cable, and a pair of transceivers on fiber, and the right answer depends on the distance and the count. DAC is a fixed copper cable with the connectors built on, cheap and very low power, but limited to short reaches, commonly a couple of meters at 400G and shorter still at 800G. It suits in-rack and adjacent-rack links where the run is short and the count is small.

AOC is a fixed assembly with the optics built onto both ends and fiber in between, so it reaches farther than DAC, up to around 100 m on multimode, at a power draw between DAC and a full transceiver pair. It saves the cost and the cleanliness exposure of field-terminated connectors, but because the ends are fixed you cannot patch through a panel or swap one optic, which limits how it fits a structured plant. The third option, separate transceivers on a structured fiber link, is the flexible one: it patches, it serves any reach the optic supports, and it is what most of the fabric uses beyond the short hops.

The common pattern is a mix. DAC for the in-rack short hops where it fits, AOC for some row-level runs, and transceivers on structured fiber for the rest of the fabric. Confirm the reach and the host support for any DAC or AOC against the platform list, because the copper reach in particular is short and shrinks at every rate step.

The power and heat the optics add

The optics consume real power and throw real heat, and the count multiplies it into a load people forget to budget. A single high-speed module commonly draws on the order of 14 to 20 watts or more, with the higher figures on the longer-reach single-mode types and the digital signal processor inside the module accounting for a large share of that. One module is nothing. Tens of thousands of them is a different conversation.

Do the multiplication at the switch and the rack. A fully populated high-port-count 800G switch can dissipate several hundred watts from its optics alone, on top of the switch silicon, and that heat lands in the same cabinet and the same cooling envelope as everything else. The form factor matters here too: OSFP was designed with more thermal headroom than QSFP-DD, because the heat is the constraint that limits how densely the modules can run.

Account for the optics in the power and cooling budget from the start, not as an afterthought when the cabinet runs hot. The optic power is part of why module efficiency and the emerging approaches below get so much attention. Confirm the per-module power against the manufacturer datasheet for the specific optic, because it varies by reach type and generation, and a budget built on the cheapest short-reach figure will fall short once the single-mode long-reach modules go in.

Structured cabling versus point-to-point

Structured cabling routes the fabric through patch panels and a documented hierarchy instead of running every link as its own point-to-point cable, and at AI-cluster scale it is the difference between a serviceable plant and a tangle nobody can fix. TIA-942 lays out the spaces and the hierarchical-star structure for a data center: the main distribution area, the intermediate and horizontal distribution areas, and the equipment distribution area at the cabinet. The fabric trunks run between those distribution areas, and the breakout and the patching happen at the panels.

The point-to-point alternative looks faster on day one and costs you for the life of the room. A floor of individual cables run switch-to-node has no patch field to test against, no place to isolate a bad link, and no way to re-home a connection without pulling a new cable across a live hall. The spaghetti also blocks airflow and buries the labeling. Structured cabling trades a little more up-front material and a patch field for the ability to test, patch, and trace every link at a known point.

The architecture guide covers how the fabric design maps onto these spaces. The discipline here is to land the high-speed links on patch panels with the polarity and the fiber map documented, so that bring-up and every later move work against a structured plant instead of a field of guesses.

What is the optical loss budget?

The optical loss budget is the total light the link can afford to lose between the two transceivers and still pass data, and at high speed that budget is small enough that every connector counts. The transmitter launches a certain power, the receiver needs a certain minimum, and the difference is the budget. Everything in the path spends it: the fiber's own attenuation, each connector, each splice, and any patch. Spend more than the budget and the receiver sees too little light, and the link errors or fails.

The headroom shrinks as the rate climbs. High-speed links at 400G and above can run with link loss budgets on the order of only 1 to 2 dB, which is a hard number to live inside once you count connectors. A single mated connector pair commonly costs a few tenths of a dB, with IEC-grade single-mode connectors typically specified at a maximum around 0.3 to 0.5 dB each. Stack two or three connections in a structured channel and a sloppy plant eats the whole budget before the fiber length even matters.

Budget the link before you build it and verify it after. Add up the fiber loss for the length, plus the connector and splice count times their rated loss, and confirm the total fits under the transceiver's published budget for that reach. The exact connector and link loss limits come from the IEC and TIA standards and the transceiver datasheet, so design to those numbers, not to a generic figure. The tighter the budget, the less room there is for a marginal connector to hide.

Connector cleanliness is the number one fiber failure

A dirty connector is the most common cause of a failed or marginal fiber link, and on a high-speed plant with a 1 to 2 dB budget it is the failure that hides the longest. A speck of dust on a ferrule, a fingerprint, an oily residue from a careless touch, any of it adds loss, scatters light, and can scratch the end-face when two connectors are mated dirty. On an MPO array, one bad fiber out of sixteen fails the whole link, so the density makes cleanliness less forgiving, not more.

Inspect and clean every connector, every time, before mating. The discipline is inspect, clean if it fails, inspect again, then mate, and it applies to the bulkhead side you cannot see as much as the jumper in your hand. A fiber scope shows the end-face against an acceptance standard; IEC 61300-3-35 is the common end-face inspection and grading standard, and it tells you whether the contamination or the defect is in a zone and a size that fails. Do not eyeball it and do not assume a factory-terminated jumper is clean out of the bag.

This is the discipline that separates a fabric that comes up clean from one that fights you for weeks. The cluster has tens of thousands of connectors, and the math is unforgiving: a small percentage of dirty connectors across that many links is still a large pile of intermittent faults to chase in a powered room. Clean before you mate, or you own the bring-up.

How do you test and certify every link?

You test every link, you do not assume any of them, and you keep the results. The base certification is an insertion-loss measurement of the finished channel with an optical loss test set, commonly called Tier 1, which reads the total end-to-end loss including every connector and splice and confirms the length and the polarity. That loss number is then compared against the link's loss budget; if it passes, the link is certified, and if it fails, you have caught it before bring-up instead of during it.

Where a link fails or needs diagnosis, an OTDR trace, commonly called Tier 2, shows where the loss is along the fiber, so you can tell a bad connector from a tight bend or a bad splice. TIA treats the loss test as the certification of record and the OTDR as the diagnostic and optional layer, so the loss test is the one that must pass on every link, and the OTDR is what you reach for to find the fault. The limits and the methods come from the TIA and ISO/IEC fiber standards, so certify against the standard the project specifies.

Test all of it, label the passes, and file the results. Skipping links to save time is a false economy on a cluster this size, because the link you did not test is the one that brings the fabric down at training, and now you are testing it anyway, live, under pressure. Certify cold, document the loss margin, and hand over a record that proves every link met budget.

Labeling, cable management, and airflow

Labeling and cable management are what make a fabric of this density serviceable, and they are the first things sacrificed under schedule pressure. Every trunk, every patch, and every port needs a label that ties back to the design's naming scheme, at both ends, so a tech at bring-up can trace a link without ringing it out. With tens of thousands of connections, an unlabeled or inconsistently labeled plant is unworkable, and the cost of fixing it later is far more than the cost of doing it during the pull.

Cable management is not cosmetic at this density. The fiber and the breakout assemblies have to be routed and supported so that bend radius is respected, weight is not hanging on connectors, and the bundles do not block the airflow the cabinets need. Fiber has a minimum bend radius, and a cable cinched too tight or kinked behind a panel adds loss or breaks later, which shows up as a link that degrades over time rather than failing clean.

Airflow is the part cabling people underweight. A dense, badly dressed bundle in front of an exhaust or a switch intake turns into a thermal problem, and in an AI hall where the cabinets already run hot, that matters. Dress the bundles to keep the air paths open, support the weight off the connectors, and keep the labeling legible after the panel is full.

Link length, latency, and the reach limits

Keep the fabric physically tight, because length costs you in two ways: latency and reach. Light takes time to travel the fiber, and on a latency-sensitive training fabric where the GPUs synchronize constantly, longer links add delay that the workload feels. The design keeps the hot back-end links short for that reason, which is part of why rail-optimized and tightly packed layouts are common in AI halls.

Length also runs into the optic's reach limit. Every optic type has a maximum distance, and the shorter, cheaper types have the least room. A run that creeps past a short-reach optic's distance during install, because the route went the long way around the room, turns a planned cheap link into a marginal one or forces a more expensive optic. The routed length is what counts, not the straight-line plan distance, the same lesson the electrical trades learn on voltage drop.

Plan the link lengths against the optic reaches and confirm the routed distance before committing the optic. Where a link has to be long, size the optic for it from the start. The exact reach a given optic supports is set by the IEEE clause and the manufacturer datasheet for that part, so design the lengths to those limits rather than discovering them when a long link will not stay up.

The install volume an AI cluster carries

The sheer install scope is the part that surprises teams new to AI builds. Tens of thousands of fiber connections, every one inspected, cleaned, mated, and tested, is a large labor effort with a long critical path, and it tends to land late in the schedule when the room is already crowded with racks and pressure. Underestimate it and the cabling becomes the thing the whole cluster waits on.

The biggest schedule lever is pre-terminated versus field-terminated fiber. Pre-terminated trunks and assemblies arrive with factory connectors tested to spec, which cuts field labor and removes most of the field-termination quality risk, at the cost of needing accurate lengths up front and less flexibility if a run changes. Field termination is more forgiving of length surprises but slower and more exposed to workmanship faults. Most large AI plants lean heavily on pre-terminated assemblies for the trunking and reserve field work for the exceptions.

Stage the work so the cabling is not the bottleneck. Pull and test the trunks and the patch fields before the switches and nodes arrive where the sequence allows it, so that when the equipment lands the links are already certified and waiting. The plant that was staged and tested ahead comes up on schedule; the one that started cabling after the racks were energized is the one explaining the slip.

Co-packaged optics and linear-drive: the emerging direction

Two emerging approaches aim at the optic power problem, and both are early enough that you design around the pluggable plant you can buy today and watch the rest. Linear-drive pluggable optics, LPO, remove the power-hungry digital signal processor from the module and let the switch silicon drive the optics more directly, which can cut module power and latency. Co-packaged optics, CPO, go further and integrate the optical engines next to the switch chip instead of in a pluggable cage, trading the field-replaceable module for a large gain in power efficiency per port.

The reported power gains are real but the maturity is not settled. LPO is described as roughly halving module power against a conventional pluggable, and CPO as delivering several times the power efficiency per port, which is why both get so much attention as rates climb to 1.6T and the optic power becomes a limiting load. Against that, CPO gives up the easy swap of a failed pluggable, and the ecosystem, the multi-vendor supply, and the serviceability are still maturing.

For a build you are cabling now, the practical stance is to plan the structured fiber plant around standard pluggable optics, which have multi-vendor supply and a known service model, and to treat CPO and LPO as a direction to track rather than a default. The figures and the timelines here move quickly, so verify the current state against the vendor roadmaps and the standards bodies before betting a design on either.

Commissioning and fabric validation

Commissioning is where the certified links become a working fabric, and it is the acceptance step that proves the cabling did its job. Link bring-up takes the certified channels, plugs the optics, and confirms each link trains to its rate with no errors. A link that passed its insertion-loss test cold should come up clean; one that certified but will not train points you back at polarity, an optic mismatch, or a marginal connection the loss test was too close to catch.

Validate the fabric, not just the links. Once the individual links are up, the fabric has to be checked as a whole for error rates, for the expected topology, and for the all-to-all reachability the training workload depends on. This is where a single bad link out of thousands shows itself as a hot spot or a re-transmission, and where the value of having tested and labeled every link cold pays off, because you can isolate the offender fast instead of bisecting a powered fabric.

Treat commissioning as a gate, not a formality. The cluster is not accepted because the racks are powered; it is accepted because the fabric trains clean and the link records prove it. Run the loss and error checks, validate the fabric, and capture the results as the acceptance record, because that record is what you defend the install with when a link degrades months later.

What to keep: the cable plant record and as-built

The cable plant is only as good as the record that documents it, and on a fabric this size the as-built is what makes the next change survivable. Capture the link records: the endpoints, the optic type at each end, the fiber type and the polarity method, the connector count, the measured insertion loss and its margin against budget, the test date, and who certified it. That set is what answers the question, six months out, of whether a degrading link was ever within budget or just barely passed.

Keep the labeling scheme, the patch schedule, and the polarity convention in the same record as the test results, so the physical plant and the documentation agree. A photo of the labeled panel, the inspection and test results, and the as-built map captured at the point of work, in the field, beats a drawing reconstructed from memory weeks later. A field tool such as FieldOS is built for exactly this: capture the link test, the connector inspection, and the panel photo at the cabinet, tie them to the link ID, and the as-built assembles itself instead of being rebuilt from a pile of notes.

Hand over a record that lets the next person work the plant without ringing out every link. That is the difference between a fabric that takes a change in stride and one where every move is an investigation.

Common mistakes

• Mating dirty fiber connectors without inspecting and cleaning, the single most common fiber failure.
• Crossing MPO polarity so transmit meets transmit and the link will not come up despite clean optics.
• Matching the optic to the wrong reach, a short-reach multimode optic on a run that exceeds its distance.
• Building a link over its optical loss budget by stacking too many connectors on a tight high-speed budget.
• Skipping the insertion-loss test on links to save time, then finding the bad one live at bring-up.
• Running point-to-point spaghetti instead of structured cabling, leaving no patch field to test or trace against.
• Ignoring the optics power and heat, so the cabinet runs hot once the modules are populated.
• Using the straight-line plan distance instead of the routed length and overrunning a short-reach optic.

Field checklist

Standards and references

IEEE 802.3 defines the Ethernet rates and the optical types behind them, including the 400G, 800G, and 1.6T families and their lane structures and reach classes such as SR, DR, FR, and LR. The InfiniBand specification covers the equivalent rates on the dedicated AI fabrics. The speeds, the reaches, and the supported optic types come from those documents and from the transceiver datasheet for the specific part, so confirm the exact rate, reach, and lane map against them rather than a rule of thumb.

The transceiver form factors follow multi-source agreements, the QSFP-DD MSA and the OSFP MSA, which define the module and cage so optics from different vendors fit the same host. Structured cabling for the data center follows TIA-942 for the spaces and the hierarchical-star design, with the MPO polarity methods and the channel insertion-loss limits set in the TIA-568 fiber standards and their ISO/IEC counterparts. Connector end-face inspection and grading follows IEC 61300-3-35.

The figures in this guide are typical at the time of the 2026 review, and this field moves fast. Hedge the speeds, the reaches, the loss limits, and the optic power to the IEEE or InfiniBand specification, the TIA and IEC standards, and the manufacturer datasheet, and let the project design and specification control the choice. The three habits that hold across all of it: match the optic and fiber to the speed and reach, clean and test every connector to the loss budget, and run structured cabling with the MPO polarity sorted before the first trunk is pulled.

What to document

A cluster of this link count cannot be operated from memory, so the documentation is part of the deliverable. Capture each element, its spec, and the note that lets the next person reproduce or defend the decision.

The record that matters most is the per-link certification tied to the link ID, because that is what proves the fabric met budget and what you check first when a link degrades.

Units and terms

The optics and cabling layer carries its own vocabulary, and the same idea reads differently across a switch datasheet, a cabling submittal, and a fiber test report.

Link rates run in gigabits and terabits per second (400G, 800G, 1.6T), built from lanes at a per-lane rate. Optical loss is in decibels (dB), where a few tenths of a dB per connector decides whether a tight link fits its budget. Reach is in meters and kilometers, set by the optic type. Fiber size is in microns of core, roughly 9 for single-mode and 50 for multimode, with the grades named OM for multimode and OS for single-mode.

Back-end fabric

The high-speed, low-latency network connecting the GPUs for training; the largest link count in the cluster

Transceiver / pluggable optic

The module that converts the switch's electrical signal to light and back, in a QSFP-DD or OSFP cage

MPO / MTP

A multi-fiber push-on connector carrying 8, 12, or 16 fibers in one ferrule; MTP is a high-performance MPO brand

Polarity

The end-to-end fiber mapping (methods A, B, C) that ensures transmit lands on receive; wrong polarity, no link

Loss budget

The total optical loss a link can afford between transceivers, spent by fiber, connectors, and splices, in dB

DAC / AOC

Direct-attach copper (short, cheap, low power) and active optical cable (farther, fixed-end optics on fiber)

SR / DR / FR / LR

Optic reach classes: SR short multimode, DR and FR longer single-mode, LR longer still

Single-mode / multimode

Single-mode (~9 micron core) carries one path far; multimode (~50 micron core) carries many paths a shorter way

Related tools

Calculators and readiness checks for this work

Compare your options

FAQ

People also ask

• What is a spine-leaf network?
• What is the difference between single-mode and multimode fiber?
• What is direct-to-chip liquid cooling?
• How much power does an AI rack use?
• Why does AI need fast storage?