In high-speed data centers, “use DAC for short, AOC for long” isn’t a rule of thumb pulled from thin air—it’s what the physics of signal transmission demands. Direct Attach Copper (DAC) moves electrical bits over twinax with almost zero link power and vanishing latency, which is why server uplinks to a top-of-rack switch and short switch-to-switch jumps are so cost-effective with copper. The catch is distance: as line rates climb, the same piece of copper that sailed at 10G starts running out of headroom at 100G. That’s not vendor superstition; it’s a consequence of how high-frequency signals behave in metal, and why a carefully scoped short-reach DAC design keeps BOM and power budgets lean without risking link flakiness.

Three mechanisms set the limits. First is plain attenuation from resistance: push current through any conductor and you lose amplitude as heat, and the farther you go the more you lose. Second is skin effect: at higher frequencies the current crowds toward the conductor’s surface, shrinking the effective cross-section and raising resistance right when you can least afford it. Third is dielectric loss in the cable’s insulation; the rapidly oscillating electromagnetic field polarizes the material, and a slice of your signal energy gets turned into heat with every centimeter. The result is a closing eye diagram as you add meters, especially at 25/50 Gb/s per lane, which is why practical copper reach at 100G is measured in just a handful of meters in real racks, not on an empty lab bench.

Add crosstalk and reflections to the mix and you see why distance erodes faster at higher data rates. Adjacent pairs couple capacitively and inductively, so energy “leaks” into neighbors; longer bundles mean more opportunities for that leakage to accumulate, and return-loss from impedance discontinuities further eats your margin. Copper also hears the room. Big power supplies, high-current whips and RF-noisy bays raise the floor of electromagnetic interference (EMI), and a long electrical run becomes a longer antenna. AOC dodges all of this because fiber is a dielectric and the signal is light; there’s no skin effect, dielectric loss is tiny over tens of meters, and EMI doesn’t apply. That immunity is why engineers lean on AOC distance & EMI margin whenever links leave the rack, cross aisles, or share trays with power.

None of this makes DAC obsolete; it just defines its sweet spot. Inside a rack or to the neighbor, passive DAC is brutally efficient: cheapest per port, no optics to power, nearly zero latency, and easy to stock in a few standard lengths. Keep copper close to the switch and life stays simple; push it across the room and you start paying in rework hours, airflow penalties from thick bundles, and head-scratching “works in the lab, fails in the row” tickets. Fiber flips that trade: the 2–3 mm jacket routes cleanly, ladder racks stop overflowing, and you can move a storage shelf or add a GPU tray without discovering you’re one meter short. When your roadmap points at 200/400G, those same pathways keep working because the signal-integrity math still favors light over long electrical runs.

Total cost of ownership is where the hybrid model wins. Yes, an AOC link costs more up front and you’ll budget a watt or two per end, but cooler intakes and slower fans claw back energy in dense rows, and predictable routing saves change-window time. Meanwhile DAC keeps the majority of short hops cheap and power-free, which is why it continues to carry most intra-rack traffic and keeps the BOM streamlined. The art is in drawing a clear boundary—copper for the short, clean paths that never leave the rack; fiber for anything long, noisy, or operationally awkward. If you standardize SKUs, label lanes once, and validate port modes and firmware in a small pilot, the network stops being fragile and starts behaving like infrastructure.

For readers who want to see the physics tied to practice, think of frequency as the rate your bitstream forces voltage to swing. At tens of gigabits per second per lane, those swings are fast enough that metal behaves like a narrower conductor (skin effect), the insulator around it starts sipping energy ,dielectric loss, and the fields couple to neighbors (crosstalk). Shorten the path and you suppress all three; that’s DAC’s domain. Lengthen the path and the penalties compound; that’s where optics win. Design to the map, not the wish list: measure real routes, add honest slack, and decide up front which medium owns each segment. If you want a deeper primer to share with your team, park a simple reference on signal integrity fundamentals so nobody has to relearn this under pressure at 2 a.m.