Use direct attach copper (DAC) for any GPU-to-switch link that fits within roughly 3-7 meters at 200G/400G (a commonly cited vendor-spec range for passive twinax at these rates, not a fixed physical constant) — which covers intra-rack and most adjacent-rack topologies. Reach for active optical cables (AOC) or transceiver-based fiber only when the physical distance exceeds DAC’s signal-integrity limit, because past that point copper is not an option regardless of your cost preference.

What actually separates DAC from AOC at the physical layer?

A DAC is a passive twinax copper assembly: the connector housings hold no active electronics, the signal travels as an electrical waveform end to end, and the only conditioning is whatever the host SerDes provides. That makes DAC cheaper per link, lower in power draw (effectively zero cable-side power), and lower in latency because there is no electrical-to-optical-to-electrical conversion in the path. The trade-off is bandwidth-distance: as you push to 200G and 400G per port, the copper channel’s insertion loss and crosstalk climb steeply with length.

AOC embeds transceivers in the connector shells and runs light over fiber between them. You pay for the optics, you burn a few watts per cable, and you add a small conversion latency, but the optical channel does not degrade over the short spans a data center rack aisle demands. The result is a clean division of labor: DAC owns the short reach where its economics dominate, AOC owns everything past where copper stops working reliably.

The important nuance is that these are not interchangeable options across the full range of distances. DAC cables are passive, cheaper, and lower-latency than AOC, but they are distance-limited to roughly 3-7 meters at high data rates before signal integrity degrades. Beyond that distance you need AOC or transceiver-based fiber regardless of what your budget spreadsheet prefers — the copper channel simply will not close the link.

When is the DAC-vs-AOC choice a real trade-off versus a foregone conclusion?

Most of the time it is not a trade-off at all. Within a single rack or between adjacent racks, DAC is almost always the correct default: the distance constraint does not bind, and the cost and latency advantages are genuine and measurable across a large fabric. If every GPU-to-top-of-rack link in your cluster is under 3 meters, spending on AOC there buys you nothing but higher BOM cost and higher power.

The decision only becomes an actual engineering trade-off at the top-of-rack-to-spine distances, where DAC’s reach limit is the binding constraint rather than a preference. That is the layer where cable runs commonly exceed 5-7 meters, where you are aggregating many links, and where the physics forces the choice. Below that layer, treat DAC as the default and require justification to deviate. At that layer, treat the reach requirement as the deciding input and let cost fall out of it.

How do you decide per link? A decision rubric

Evaluate each link class (GPU-to-ToR, ToR-to-spine, spine-to-spine) independently rather than picking one cable type for the whole fabric:

Condition	Recommended interconnect	Reasoning
Link length < 3 m, 200G/400G	DAC (passive)	Well inside copper reach; lowest cost, lowest latency, zero cable power
Link length 3-5 m, 200G/400G	DAC if the specific cable is rated and validated for that length; otherwise AOC	Marginal zone — signal integrity depends on host SerDes and cable spec
Link length 5-7 m, 200G/400G	AOC or transceiver + fiber	Copper reach is at or past its practical limit at high rates
Link length > 7 m	AOC or transceiver + fiber	DAC is not viable; cost preference is irrelevant
Latency-critical, short reach (e.g. collective-heavy training)	DAC	Removes E/O/E conversion latency; every nanosecond on all-reduce paths compounds
Power/thermal budget tight at rack level	DAC where reach allows	Passive cable adds no per-link power
Cable management for long structured runs	Fiber (thinner, longer, easier to route)	DAC twinax is thick and stiff at length

Two operational checks before you commit: confirm the cable’s rated length matches the actual routed path (not the straight-line rack distance — slack, service loops, and vertical runs add up), and verify the cable is on the switch and NIC vendor’s qualified compatibility list. In the fabric builds we’ve reviewed, an unvalidated DAC at the edge of its reach is the single most common source of intermittent link flaps that are painful to diagnose in a live training run.

Does the rate matter more than the distance?

They are coupled, and rate is what shrinks the usable copper distance. A DAC that comfortably runs 5 meters at 100G may fail at 400G over the same length because the higher signaling rate widens the loss budget the channel has to satisfy. When you upgrade a fabric from 200G to 400G, re-run the reach analysis; links that were solidly DAC-appropriate can cross into the marginal or AOC-required zone without the physical layout changing at all.

This is why “we standardized on DAC last generation” is not a safe default when rates increase. Treat every rate transition as a trigger to re-audit the marginal-length links, since the binding constraint moves inward as the data rate goes up.

Interconnect choice is a small line item next to the rest of a GPU cluster build, but it is one we see teams get wrong precisely because it looks trivial. For the broader question of where these decisions sit — whether you are building on-premise accelerator racks or renting cloud GPU capacity where the fabric is abstracted away — see cloud GPU vs on-premise AI accelerators: a total cost analysis. For the hardware-tier decisions that determine how many links you need in each distance class, see server GPU for AI inference: why hardware tier matters in production.

Frequently Asked Questions

What is the maximum distance for a DAC cable at 400G?

Passive DAC is practically limited to roughly 3-7 meters at 400G before signal integrity degrades, with the usable length depending on the specific cable spec and the host SerDes quality. Past that range you need active optical cables or transceiver-based fiber. Always validate against the vendor’s qualified length rather than assuming the theoretical maximum.

Is DAC lower latency than AOC?

Yes. DAC is a passive electrical assembly with no electrical-to-optical-to-electrical conversion, so it avoids the small transceiver latency that AOC adds. On latency-sensitive, collective-heavy training workloads this difference compounds across the fabric, which is one reason DAC is preferred for short intra-rack GPU links.

Can I use DAC for top-of-rack to spine connections?

Usually not, because ToR-to-spine runs commonly exceed DAC’s 3-7 meter reach at high data rates. This is exactly the layer where the choice becomes a real trade-off and AOC or fiber typically wins on reach. Measure the actual routed cable path, including slack and vertical runs, before deciding.

Does moving from 200G to 400G change my cable choice?

It can. Higher signaling rates shrink the usable copper distance, so a link that was safely DAC at 200G may fall into the marginal or AOC-required zone at 400G despite no physical layout change. Re-audit all marginal-length links whenever you upgrade the fabric rate.