Copper vs Fiber Interconnects 2026 — DAC, AOC & Optical Transceivers for GPU and NVMe Edge Storage Fabrics

Nobody budgets for the cables until the bill arrives. You spec the GPU, the 存储, the CPU, the cooling — and then discover that the interconnect layer quietly costs more than a mid-range server, and gets it wrong means your expensive GPUs idle waiting on a bottleneck you never saw coming. Interconnects are the nervous system of every AI node: the DACs and optical cables that link GPUs to switches, and the copper risers and retimers that connect NVMe drives to the CPU. This guide lays out the copper-versus-fiber decision across the speeds that matter in 2026 — 100G, 200G, 400G, and 800G — and the PCIe Gen5/Gen6 cabling underneath it.

The Three Interconnect Types, Defined

• DAC — Direct Attach Copper. A passive (or lightly active) copper twinax cable with transceivers permanently molded onto each end. Cheapest per link, lowest latency, lowest power, but limited reach (typically 2–5 m passive, up to ~7 m active).
• AOC — Active Optical Cable. Fiber with the optics built into each connector, so it plugs in like a DAC but travels like fiber. Reaches 30–100 m, immune to EMI, costs more and draws more power than DAC.
• Optical transceiver + fiber. A pluggable module (QSFP28/QSFP-DD/OSFP) into a separate fiber jumper. The most flexible — you can swap optics and fiber independently — and the only option for structured, patch-panel, or long-haul runs beyond ~100 m.

The core trade is simple and worth memorizing: copper is cheaper, lower-latency, and lower-power, but short; fiber reaches far, ignores electrical noise, and is the only path past a few meters. Everything else is deciding which side of that trade your deployment sits on.

DAC vs AOC vs Optical Transceiver — Side by Side

Where the Decision Actually Gets Made

Inside the rack, use copper (DAC). The vast majority of GPU-to-switch links are within 3 meters, and that is exactly where DAC wins: a 400G OSFP DAC costs a fraction of an equivalent AOC or transceiver pair, adds near-zero latency, and draws almost no power — which matters when you are already fighting the thermal budget of a dense GPU chassis. The rule of thumb that has held for years: if it is in the same rack, it is DAC; if it crosses a rack or leaves the cabinet, it is fiber.

Cross-rack and structured cabling, use AOC or transceivers. Once a link spans rows, you need fiber for reach and for immunity to the electrical noise of a factory floor or substation. AOC is the plug-and-play choice for a handful of cross-rack links; a real transceiver-plus-fiber plant is worth it the moment you are installing patch panels, running a backbone, or think you might change optics vendors down the line. In an industrial setting the EMI argument tips things further toward fiber than it would in a clean data center — copper DACs near a VFD or welder can pick up noise that fiber simply ignores.

For GPU fabrics specifically (NVLink / InfiniBand / 400G+ Ethernet), watch the power math. At 400G and 800G, optical module power becomes a first-order cost: a rack of 800G links can burn kilowatts in optics alone, and that heat has to go somewhere. DACs sidestep almost all of it within the rack — another reason hyperscalers keep intra-rack links on copper and reserve optics for the spine.

PCIe Gen5/Gen6 and the NVMe Side of the Same Coin

The interconnect decision is not only about Ethernet and InfiniBand. Underneath your 存储 lives a parallel cabling question: how the NVMe drives reach the CPU. PCIe Gen5 and the emerging Gen6 raise the signal-integrity bar to the point where copper traces and risers need retimers — active signal regenerators that clean up the eye diagram over any meaningful distance. The same copper-versus-fiber logic applies in miniature: short, direct board-to-board or drive-to-backplane runs stay on copper (MCIO, OCuLink, slimSAS), while anything crossing an enclosure or a rack turns to optical or active copper. For an edge server with 4–8 hot-swap U.2/E1.S NVMe bays, copper risers plus a good backplane are the standard; for disaggregated storage where drives live in a separate shelf from the compute, you step up to retimed copper or optical links.

Reference Pricing (Q3 2026, indicative street)

The gap widens with speed: at 100G the copper advantage is modest, but at 400G and 800G a DAC can cost less than a third of the optical equivalent while drawing a tenth of the power. That is why the "copper inside the rack, fiber between racks" rule only becomes more true as speeds climb.

Who Should Buy What

You are building a single GPU edge server or a small cluster within one rack — buy DACs for every link under 3 meters and stop there; the optical premium buys you nothing inside a cabinet. You are wiring a multi-rack cluster, a factory floor, or an outdoor/electrically-noisy site — run a fiber plant with transceivers (or AOCs for quick cross-rack links) and keep copper strictly for the rack-local hops. You are disaggregating storage from compute — treat the NVMe fabric as a second interconnect decision and size retimers or optical links for the enclosure-to-enclosure distance. In every case, buy cables that are coded and tested for your switch and NIC vendor — a bargain DAC that your switch refuses to bring up is not a bargain.

Pricing and availability as of August 2026, QSCompute distribution channel. Bulk and long-term-agreement pricing available for integrators and multi-site fleets.