The copper-versus-fiber question is no longer theoretical — AI workloads are forcing a tier-by-tier audit of every link in the data center

Question 1: The Honest Answer — Copper Is Not Dead, But Its Territory Is Shrinking

Let's start with the question exactly as customers ask it: "We have Cat6A horizontal runs, DAC in-rack links, and some old Cat5e patches. AI workloads are coming. Do we need to replace it all?"

The honest, engineer's answer is: it depends on the link, not the facility. The AI era is not a sudden cliff where copper stops working everywhere. It is a steady squeeze — each speed generation shrinks copper's usable reach, and each new GPU density level pushes more links past copper's comfort zone. The result is that copper and fiber are no longer competing; they are coexisting by tier, and the boundary between them is moving inward.

This is not AMPCOM's opinion alone. It is the consensus that emerged at NVIDIA GTC 2026, where Jensen Huang stated: "For the next 5–10 years, we will still use a massive amount of copper, and also a massive amount of optics." Broadcom's CEO Hock Tan reinforced the same position, noting that copper can scale from 100G to 400G per lane within a rack before advanced optical solutions become mandatory. The industry calls this "copper where you can, optics when you must" — and the "must" threshold is what this article helps you identify.

The physics behind the squeeze are straightforward. Copper attenuates high-frequency signals — the higher the data rate, the faster the signal degrades. At 10G NRZ, a passive DAC works to roughly 7 meters. At 100G PAM4, that drops to about 3 meters. At 800G (8×112G PAM4), passive DAC reaches only 1–2 meters. And at the coming 1.6T generation (8×224G PAM4), the practical passive copper span falls to roughly 1 meter or less — a distance that barely covers a single in-rack GPU-to-switch hop. This is what the industry calls the "Copper Wall": the point where equalization, retiming, and signal conditioning consume more power and cost than simply switching to fiber.

Speed Generation	SerDes Lane Rate	Passive DAC Max Reach	What It Means for Your Copper
10G	10G NRZ	~7 m	Copper is fine for almost any in-room link
25G / 40G	25G NRZ	~5 m	Copper still covers most rack-to-rack in a row
100G	100G PAM4 (4 lanes)	~3 m	Copper is limited to in-rack or adjacent-rack
400G	100G PAM4 (8 lanes)	~2–3 m	Copper is in-rack only; AEC extends to ~5 m
800G	112G PAM4 (8 lanes)	~1–2 m	Copper barely covers GPU-to-leaf; optics take over row-level
1.6T	224G PAM4 (8 lanes)	~1 m (estimated)	Passive copper is nearly impractical; AEC/AOC mandatory

So: copper is not dead. But the territory where it is the right choice is shrinking with every speed generation. If your facility is staying at 10G/25G for management and access-layer links, copper remains the correct answer. If you are building AI training clusters at 400G/800G, copper's role contracts to the shortest in-rack hops, and fiber takes over everything else.

Copper still handles the shortest, most cost-sensitive links — but its reach shrinks dramatically at 800G and beyond

Question 2: Where Copper Still Wins — The In-Rack Stronghold

Inside a single rack — the domain NVIDIA calls "scale-up" — copper remains the optimal choice in 2026, and the industry expects it to stay that way through at least 2027–2028. Here is why.

The physics still favor copper at short reach

• Power: A passive DAC consumes ~0.1 W per end — roughly 1/50th the power of a 800G optical module (20–25 W). In a rack with 48+ high-speed ports, that difference is measured in kilowatts.
• Latency: Passive copper adds essentially zero latency — no electro-optical conversion, no retiming. For NVLink and scale-up fabrics where every nanosecond matters, this is a genuine advantage.
• Cost: A passive DAC is typically 2–5× cheaper than an equivalent AOC, and 3–5× cheaper than separate transceivers plus fiber. For links under 2 meters, the cost penalty of optics buys nothing.
• Reliability: No lasers, no photodetectors, no DSP — fewer failure modes. NVIDIA's own NVL72 rack uses active copper cables for GPU-to-NVSwitch connections, validating copper's role in the most demanding AI scale-up designs.

Active copper extends the boundary

When passive DAC's reach is insufficient — typically at 400G/800G where the GPU-to-leaf run is 2–5 meters — active copper (ACC and AEC) fills the gap without jumping to optics:

Copper Type	Technology	800G Reach	Power per End	Best For
Passive DAC	No active components	~1–2 m	~0.1 W	Shortest GPU-to-leaf hops
ACC (Active Copper)	Analog redriver / equalizer	~4–5 m	~1.5 W	In-rack and adjacent-rack at 800G
AEC (Active Electrical)	Digital retimer with CDR	~5–7 m	~4–8 W	Server-to-switch rows where optics are overkill

Astera Labs reports that AECs deliver approximately 25–50% lower power consumption compared to AOCs, and multiple hyperscalers have standardized on AEC technology for 800G in-rack and adjacent-rack links. Credo Technology, the dominant AEC retimer supplier, has seen deployments where AECs replaced both passive DAC (which could not reach) and AOC (which was overkill). The practical takeaway: copper's territory has expanded slightly thanks to active technology, even as passive copper's reach has shrunk.

Inside the rack, copper's power and latency advantages remain decisive — active copper (AEC) extends that boundary to 5–7 m at 800G

Question 3: Where Copper Already Loses — Anything Beyond the Rack

The moment a link leaves the rack — crossing an aisle, reaching a spine switch, connecting to another row or hall — copper's advantages evaporate. This is the tier NVIDIA calls "scale-out", and it is where fiber is not just preferred but physically mandatory at AI speeds.

Why copper fails at row-level and beyond

• Reach: At 800G, passive DAC reaches 1–2 m. Even AEC tops out at ~7 m. Row-level links are typically 10–30 m, and spine connections can span 50–100 m+. Copper simply cannot carry 800G that far.
• Signal integrity: At 224G PAM4 (1.6T), copper channel loss becomes so severe that the equalization circuitry consumes more power than the data path itself. The industry calls this the "Copper Wall" — the frequency threshold where no amount of DSP can economically recover the signal.
• Cable bulk: High-speed copper bundles are thick, stiff, and heavy. In a dense AI hall with hundreds of 800G links per row, copper's physical mass becomes a cable management and airflow nightmare.
• EMI susceptibility: AI racks push 100–200 kW. The electromagnetic environment around power distribution, cooling systems, and thousands of GPUs makes copper's EMI vulnerability a real reliability risk at multi-meter distances.
• Heat: Copper cable resistance generates heat at the connector interface. In a rack already operating at 120+ kW, every watt of cable heat adds to an already stressed cooling system.

Where fiber is already the default

In every AI cluster AMPCOM engineers surveyed in 2026, the following link types were 100% fiber:

Link Type	Typical Distance	Speed	Why Fiber Is Mandatory
Leaf-to-spine (in-row)	10–20 m	400G/800G	Beyond AEC reach; EMI and cable bulk
Spine-to-spine (cross-hall)	30–100 m	800G	Copper cannot carry 800G past ~7 m
Inter-building / DCI	100 m – 10 km+	800G/1.6T	Only singlemode fiber reaches these distances
Breakout trunking	5–50 m	800G→2×400G	Density and flexibility; MPO fiber required

The fiber choice for these links is almost always OS2 / G.657.A2 singlemode on MPO-16/MPO-24 trunking. Singlemode carries both 800G and 1.6T with no reach penalty, so the next generation becomes a transceiver swap rather than a recable. For shorter row-level links under 100 m, OM4/OM5 multimode is a cost-optimized alternative. Our guide to choosing the right fiber type: singlemode vs multimode explains the trade-off in depth.

Once a link leaves the rack, fiber is the only option at AI speeds — and an OS2 singlemode plant carries multiple generations of optics

Question 4: The Copper-to-Fiber Tipping Point — How to Audit Your Plant

So how do you decide which of your existing links to keep and which to replace? The answer is a link-by-link audit against three criteria: distance, speed, and upgrade horizon.

The three-question audit

For every high-speed link in your facility, ask:

1. What is the actual physical distance? Measure it — do not estimate. Under 2 m at 800G: passive DAC. 2–7 m: AEC. 7–30 m: AOC. Beyond 30 m: pluggable optics + fiber.
2. What speed will this link run at in 12–18 months? If 800G today and 1.6T planned, passive DAC's 1–2 m reach shrinks to ~1 m. AEC may become necessary even for in-rack links. Plan fiber for anything that might exceed copper's 1.6T envelope.
3. Will this link need breakout or reconfiguration? If yes, choose pluggable optics + fiber — not AOC or DAC. The flexibility of separate modules and structured fiber pays for itself the first time a layout changes.

Existing Copper Link	Current Speed	Distance	Verdict for AI Workloads
Cat5e horizontal	1G	≤100 m	Keep — fine for management/access; not in AI's path
Cat6A horizontal	10G	≤100 m	Keep — still the standard for PoE++ and Wi-Fi 7
Cat8 short-link	25G/40G	≤30 m	Keep — adequate for non-AI interconnects
Passive DAC (25G/100G)	25–100G	≤3 m	Keep — still within copper's reach
Passive DAC (400G)	400G	≤2–3 m	Keep for now — plan AEC/AOC if upgrading to 800G
Passive DAC (800G)	800G	≤1–2 m	Keep only if ≤2 m — otherwise replace with AEC/AOC
Copper row-level links (>5 m)	100G+	5–30 m	Replace with fiber — already past copper's reach at AI speeds
Copper inter-hall / DCI	Any	>30 m	Replace immediately — should have been fiber already

A link-by-link audit typically finds that 80–90% of copper is fine — only the high-speed fabric tier needs fiber migration

Question 5: The 5-Step Migration Plan — Keep What Works, Fiber What Doesn't

Based on hundreds of AMPCOM customer audits, here is the practical 5-step plan for preparing an existing cabling plant for AI workloads — without a full recable.

Step 1: Segment your links by tier

Map every high-speed link into three tiers: in-rack (≤5 m), row-level (5–30 m), and backbone/DCI (>30 m). This determines which medium each link should use. Document actual measured distances — not estimates.

Step 2: Retain copper where physics still favor it

Keep passive DAC for all in-rack links at 25G–400G under 3 meters. Keep Cat6A for all access-layer and PoE links. Keep Cat8 for short 25G/40G data center links. Do not touch what works — the goal is targeted migration, not blanket replacement.

Step 3: Migrate row-level copper to AEC or AOC

For links of 3–7 m at 400G/800G, replace passive DAC with AEC — active copper with retimers that extend reach to 5–7 m while staying cooler and cheaper than fiber. For links of 7–30 m, replace with AOC or pre-terminated OM4/OM5 MPO assemblies with pluggable optics.

Step 4: Build the backbone on OS2 singlemode MPO

All spine-to-spine, inter-hall, and DCI links should use OS2 / G.657.A2 singlemode fiber on MPO-16/MPO-24 trunking. This single fiber type carries 400G, 800G, and 1.6T with only module changes. Install spare dark fiber in every trunk — AI racks grow faster than anyone plans. As our guide to MPO fiber solutions explains, choosing the right fiber count upfront prevents costly retrofits.

Step 5: Document, test, and plan for the next generation

Document polarity, length, speed, and link type per port — using a TIA-606-style labeling system. Test every fiber link with OTDR and endface inspection before first power-on. Test every copper link with a Fluke DSX-class tester. And plan your 1.6T upgrade path now: identify which DAC links will need AEC at 1.6T, which AOC links will need pluggable optics, and which fiber trunks have enough dark fiber for the next expansion.

Whatever the link type — copper or fiber — certify it before first power-on and document every port for the next upgrade cycle

Key Takeaways

Question	Bottom Line	AMPCOM Recommendation
Is all copper obsolete?	No — copper's territory is shrinking, not disappearing	Audit by link, not by facility
In-rack copper?	Still the best choice at ≤5–7 m for 800G	Keep passive DAC / AEC in-rack
Row-level copper?	Already past copper's reach at AI speeds	Migrate to AOC or OM4 MPO fiber
Backbone / DCI?	Fiber has been mandatory for years	OS2 singlemode MPO, 1.6T-ready
Access layer (Cat6A)?	AI does not affect 10G/PoE horizontal cabling	Keep — no action needed
1.6T planning?	Passive copper reach drops to ~1 m; plan AEC/AOC	Build fiber-rich trunks now