AI clusters have changed what a high-speed interconnect has to do. In a traditional enterprise network, bandwidth grows slowly and most traffic flows north-south, between users and servers. In an AI training fabric, thousands of GPUs exchange gradients, parameters, and checkpoints with each other almost continuously, and a single slow or unstable link can stall an entire job. That east-west pressure, not user demand, is what pushes racks to 800G.
So the real question is no longer "do we need 800G?" It is more specific and more expensive to get wrong: where should you run 800G DAC, where does 800G AOC earn its place, and where do 800G optical transceivers make more sense? 800G is not a single product. It is a family of direct attach copper cables, active optical cables, and pluggable optical modules, and each one solves a different link problem. This guide maps those products to real positions in an AI data center and gives a selection framework that network architects, system integrators, and procurement teams can act on.
Why AI Data Centers Are Moving From 400G to 800G
During distributed training, GPU servers run collective operations such as all-reduce across the cluster. The result is bursty, high-volume east-west traffic between accelerators that has very little to do with external users. Three requirements follow from that:
- Bandwidth density. A single rack can hold several GPU servers, each with multiple NICs that need high-speed uplinks. Port count per rack-unit matters as much as raw speed.
- Low, predictable latency. Congestion or link flaps reduce GPU utilization, and idle GPUs are the most expensive thing in the building. Stable links protect training throughput.
- A migration path. Many sites start at 400G and grow into 800G, then plan for 1.6T. The cabling decisions you make today either enable that path or block it.
The standards have already caught up to this. IEEE 802.3df-2024 formally defines Ethernet operation at 400 Gb/s and 800 Gb/s using eight-lane, 100 Gb/s-per-lane signaling. Looking forward, the Ethernet Alliance 2026 Roadmap shows hyperscalers leaning on 100G–800G interconnects today, with 1.6T and beyond on the horizon, and it flags power per watt and cooling as the next limiting factors rather than speed alone.
Which 800G Interconnect Should You Use?
If you only read one section, read this one. The table maps the most common link types to a first-choice product. Treat it as a starting point, then validate against distance, power, airflow, port form factor, and your upgrade plan.
| Link scenario | First-choice 800G product | Why it fits |
|---|---|---|
| Same-rack server-to-switch | 800G DAC (passive or active copper) | Lowest cost, lowest latency, near-zero added power on passive copper |
| Short-to-medium rack-to-rack | 800G AOC | Lighter and more flexible than copper; easier to route in dense trays |
| Leaf-spine and switch-to-switch fabric | 800G optical transceiver | Longer reach options, cleaner fiber management, easier expansion |
| 400G-to-800G migration | 800G optical module with breakout, or breakout cable | Supports staged upgrades and mixed-speed ports |
| High-density AI backbone | 800G OSFP or QSFP-DD optical module | High port density with structured fiber instead of bundles of fixed cable |
800G DAC: Best for Short In-Rack Links
An 800G direct attach copper (DAC) cable carries the signal over copper conductors with no optical conversion. That simplicity is the whole point: passive DAC adds almost no power, almost no latency, and almost no cost compared with an optical link. The classic use case is an 8-GPU server one or two rack-units away from its top-of-rack switch, a 1–2 meter run that never leaves the rack.
Reach is the constraint. Passive 800G DAC is practical to roughly 2–3 meters; beyond that, active copper cables (ACC/AEC) add equalization to extend the distance a few more meters at the cost of some power. Past that point, copper stops making sense and you should move to AOC or optics.
Reach for 800G DAC when the server and switch share a rack, the path is short and direct, latency matters more than distance, and you can manage cable stiffness and bend radius. Avoid it when the run is long or routes through congested trays, when airflow is already tight, or when the link will need frequent reconfiguration.
Field note: at 800G the copper bundle is thick and stiff. In a crowded GPU rack the limiting factor is usually not bandwidth but bend radius and airflow. A DAC that is one size too long does not just look untidy; it can block exhaust, force a sharp bend near the cage, and make module swaps painful months later.
800G AOC: Flexible Short-to-Medium Reach
An 800G active optical cable (AOC) bonds the transceivers and fiber into one sealed assembly. You get a plug-and-play cable like a DAC, but with the reach and routability of optics. Typical 800G AOC lengths run from a few meters up to around 30–50 meters, with some assemblies reaching 100 meters, which covers most rack-to-rack and within-row links that are too long for copper.
Compared with DAC, an AOC is thinner, lighter, and far easier to dress through dense cable trays, and it sidesteps the electromagnetic coupling concerns of high-speed copper. Its drawback is that it is a fixed assembly: if either end or the fiber is damaged, the whole cable gets replaced, and you cannot reuse the optics or re-route the fiber independently the way you can with modules plus structured cabling. AOC is a strong fit when flexibility matters and the endpoints are stable, and a weaker fit when you expect to re-patch the link often.
800G Optical Transceivers: Built for Scalable AI Fabrics
Pluggable 800G optical modules are the right tool once you are interconnecting switches, rows, and halls through structured fiber. Because the transceiver and the fiber are separate, you choose module type, fiber type, connector, and route independently, and you can service or upgrade each piece on its own. That separation is what makes optics scale where fixed assemblies do not.
Match the module to the reach:
- SR-type (e.g. 800G-SR8) over OM4/OM5 multimode fiber for short reaches of roughly 30–100 meters, common inside a hall.
- DR-type (e.g. 800G-DR8) over parallel single-mode fiber for about 500 meters, suited to longer in-building links.
- FR-type (e.g. 2x400G-FR4 or 800G-FR4) over duplex single-mode fiber for about 2 kilometers, with LR-type extending further for campus and inter-hall runs.
Optical modules also expose digital diagnostics (DDM/DOM), so you can read per-lane optical power, temperature, and bias in operation instead of guessing. For the cabling behind them, single-mode plant such as OS2 single-mode fiber is the usual choice when you want headroom for future speeds. The exact module choice should follow switch compatibility, the reach budget, and the fiber you already have in the ground.
800G Interconnect Map: Where Each Type Fits in an AI Data Center
Generic "it depends" advice does not help a buyer issue a purchase order. Here is how the three product types actually distribute across an AI fabric, with a typical distance and the main risk for each layer.
GPU Server to Top-of-Rack Switch
This is the densest, highest-volume layer, and most of it is short. When the server sits in the same rack as the ToR or end-of-row switch, 800G DAC at 1–3 meters is the cheapest, lowest-latency option and the default. If the run stretches across racks or the tray is congested, an 800G AOC removes copper bulk and routes more cleanly. If you are deploying structured fiber from day one, 800G optical modules give you the most long-term flexibility. The risk here is purely physical: too much copper in a hot rack costs you airflow.
Leaf-Spine Switch Fabric
Leaf-spine and similar fabrics keep bandwidth predictable as the cluster grows, and these switch-to-switch links want distance, clean fiber management, and easy expansion, so 800G optical modules dominate. Plan this layer around more than transceiver speed: check port density per switch, the thermal envelope, connector and fiber type, and whether the cabling can absorb the next speed bump without a rip-and-replace.
Storage and Backend Network Links
Training data, checkpoints, and distributed storage traffic put sustained pressure on the back-end network, and these links are the ones you least want to take down for maintenance. Optical modules over structured fiber win here because you can monitor them with DDM/DOM and replace a single module or patch cord without disturbing the rest of the link, which a sealed cable assembly does not allow. For high-fiber-count runs between storage and compute, pre-terminated MPO/MTP trunk cabling keeps the back-end tidy and serviceable.
400G-to-800G Migration and Breakout
Almost nobody upgrades every link at once. A realistic interim state mixes 400G servers, 800G switches, and breakout connections, and breakout is exactly where installs go wrong. An 800G port commonly breaks out as 2x400G, 4x200G, or 8x100G, but the port has to support the mode you want, the switch operating system has to be configured for it, and the cable or module has to match the correct lane mapping. Plan the optical side with MTP/MPO breakout cables that fit your target ratio. A wrong breakout assumption does not fail loudly; it quietly stalls an installation while someone re-cables a rack at 2 a.m.
800G DAC vs AOC vs Optical Modules
| Factor | 800G DAC | 800G AOC | 800G Optical Module |
|---|---|---|---|
| Best use case | Very short same-rack links | Flexible short-to-medium links | Scalable fabric and structured cabling |
| Typical reach | ~2–3 m passive; a few more with active copper | ~3–50 m, some to 100 m | ~100 m (SR) to ~500 m (DR) to ~2 km (FR) and beyond |
| Relative cost | Usually lowest | Medium | Higher, varies by module type and volume |
| Added power | Near zero (passive) | Higher than DAC | Highest; can be ~15 W or more per module |
| Serviceability | Replace the whole cable | Replace the whole cable | Replace module or fiber separately |
| Future expansion | Limited | Moderate | Strong |
| Main buyer concern | Length, stiffness, airflow | Fixed assembly, replacement | Compatibility, fiber type, power, thermals |
There is no single best 800G product. Use DAC where the link is short and the path is simple. Use AOC where you need flexibility but the endpoints stay put. Use optical modules where the network has to scale, stay structured, and upgrade cleanly. The cost figures above are relative and shift with reach, vendor, module type, and order volume, so price the specific link, not the category.
OSFP vs QSFP-DD: Form Factor and Cabling Checks
Before you order anything, confirm the physical and electrical interface. This is where compatibility problems hide.
OSFP vs QSFP-DD
800G modules ship in two main form factors. The OSFP MSA defines OSFP as an eight-lane pluggable supporting 400G, 800G, and 1.6T, with a slightly larger body and integrated heat sink that suits higher-power, thermally demanding modules. The QSFP-DD MSA defines QSFP-DD800 as an eight-lane interface running 100 Gb/s PAM4 per lane, with backward compatibility to the existing QSFP family as a key advantage. For a deeper look at the latter, see this QSFP-DD technical overview.
The practical rule is blunt: OSFP and QSFP-DD are not interchangeable. The switch port, NIC port, module housing, thermal design, and cable type all have to match. Confirm which ecosystem your switch uses before you commit to a module line.
Connectors and Fiber: MTP/MPO, LC, OM4, OS2
Connector and fiber type matter as much as the module. Parallel optics such as SR8 and DR8 use MTP/MPO connectors, while duplex variants such as FR4 use LC, depending on the optical standard and reach. On the fiber side, multimode like OM4/OM5 carries short-reach links, and the practical distances are easy to underestimate, so it is worth checking the multimode fiber distance limits against your actual runs. Single-mode covers longer reach and future headroom. Polarity is the quiet trap with parallel optics, so confirm it early; this engineer's guide to choosing between MTP and MPO covers the connector and polarity decisions before you buy in bulk. Verify the existing fiber plant, connector polish (UPC vs APC), and patch panel layout against the module you intend to use.
Thermal and Airflow
800G ports run hotter than anything that came before them, and high-speed optical modules can draw on the order of 15 W or more each. Do not check the data rate alone. Check module power, airflow direction, heat-sink type, and the switch's operating temperature window. In a high-density rack, poor airflow turns a perfectly correct optical choice into an operational fault, and the symptoms (intermittent flaps, FEC errors that climb with temperature) are easy to misdiagnose as bad optics.
Common 800G Selection Mistakes to Avoid
Choosing DAC only because it is cheaper. DAC is excellent for short links, but force a long copper run into a dense GPU rack and you get a stiff bundle that blocks exhaust, sits at a sharp bend near the cage, and complicates every later module swap. The cable saved a few dollars and cost airflow.
Ignoring form factor compatibility. OSFP and QSFP-DD both deliver 800G, but they are physically and electrically different ecosystems. Order the wrong one and it will not seat in the port.
Forgetting future migration. A 400G-to-800G plan has to account for breakout support, fiber count, patch panel density, and 1.6T readiness, or the next upgrade becomes a rebuild.
Underestimating thermal design. High-speed modules need stable airflow and active monitoring. Heat-related faults masquerade as optics faults and waste days of troubleshooting.
Buying without validation. At 800G, interoperability testing is not optional. Test one link end to end, then scale.
FAQ: 800G DAC, AOC, and Optical Transceivers
Q: Is 800G DAC better than AOC?
A: Neither is universally better; they fit different distances. DAC is cheaper, lower latency, and adds almost no power, but it is practical only to about 2–3 meters. AOC reaches roughly 30–50 meters and routes more easily in dense racks. Use DAC inside a rack and AOC for short-to-medium rack-to-rack links.
Q: How far can an 800G DAC reach?
A: Passive 800G DAC is practical to about 2–3 meters. Active copper cables (ACC/AEC) add equalization to extend that by a few meters while drawing some power. Beyond roughly that range, move to AOC or optical modules.
Q: When should I use 800G optical transceivers?
A: Use optical modules when links cross racks, rows, or halls, when you need DDM/DOM monitoring, or when the design relies on structured fiber that has to scale and upgrade over time. SR-type covers about 100 meters on multimode, DR-type about 500 meters, and FR-type about 2 kilometers on single-mode.
Q: Are OSFP and QSFP-DD interchangeable?
A: No. They are separate form factors with different mechanics, thermals, and cages. The switch port, NIC, module, and cable must all be the same type. Confirm which form factor your switch uses before selecting modules.
Q: What is the best 800G interconnect for AI clusters?
A: It depends on the layer. DAC for short in-rack GPU-to-ToR links, AOC for flexible short-to-medium runs, and optical modules for leaf-spine, backend, and storage fabrics where reach, monitoring, and clean upgrades matter most. A real cluster uses all three.
Q: What is the difference between 800G SR8, DR8, and FR4?
A: They differ by fiber and reach. SR8 runs over OM4/OM5 multimode for short distances (roughly 30–100 meters), DR8 runs over parallel single-mode for about 500 meters, and FR4 runs over duplex single-mode for about 2 kilometers. Choose based on the distance and the fiber already installed.
Final Thoughts
800G modules and cables are core building blocks of AI data center networks, but each type plays a distinct role. For short same-rack links, 800G DAC is the most cost-effective choice. For flexible short-to-medium runs, 800G AOC simplifies deployment. For scalable fabrics, leaf-spine networks, and future-ready structured cabling, 800G optical modules are the better long-term option. Before you select a product, define the link position, the real measured distance, the form factor, the fiber type, the thermal budget, and the upgrade path. Decisions made in that order reduce installation risk, keep the network scalable, and leave room for the next jump to 1.6T.
