DAC vs AOC vs AEC: Which Cable Goes Where in an AI Rack? (2026 Buyer's Guide)

Short version: passive DAC for links inside a rack or between adjacent racks, active copper (ACC/AEC) for the runs that land just past copper's ceiling, AOC to cross rows and halls, and separate transceivers plus structured fiber once the fabric is large enough that someone has to repair it at 3 a.m. The decision is driven by distance, power, airflow and serviceability — not by bandwidth. A DAC and an AOC built to the same standard carry exactly the same data rate.

Key takeaways

• Speed is not the variable. Same standard, same rate. Distance, power, airflow and cable management are what differ.
• Passive copper's ceiling shrinks as the lane rate climbs — roughly 7 m at 10G, 3–5 m at 25G, ~3 m at 100G and 200G, and only 1–2 m at 400G and 800G.
• Active copper (ACC/AEC) is the tier most buyers don't know exists, and it covers exactly the 3–7 m runs where passive copper has already failed and optics are still overkill.
• AOC buys distance and airflow at the cost of a few watts per end and a sealed, fixed-length assembly you cannot re-terminate.
• Transceivers plus fiber lose on unit price and win on repair. Past a few dozen links that trade starts to matter more than the invoice.
• Most production clusters use several of these, matched per link, rather than standardizing on one.

The four options at a glance

Type	Practical distance	Power	Relative cost	Best for
Passive DAC	~1–7 m, falling as the rate rises	Near zero	Lowest	In-rack, adjacent rack
ACC / LACC	~3–5 m	~1–2 W per end	Moderate	Runs that just miss passive reach
AEC	~5–7 m	~3–5 W per end	Moderate–high	Power-conscious builds past ACC
AOC	10 m to 100+ m	~8–15 W per cable	High	Row-to-row, hall-to-hall, dense racks
Transceivers + fiber	100 m MMF, 500 m DR4 SMF, 10 km LR4	~8.5 W per 400G module	Highest up front	Large fabrics, structured cabling, repairability

Naming caution: the market is not consistent about the middle two. NVIDIA's LinkX catalog uses ACC and LACC; much of the industry says AEC for anything retimed. ACC generally means a re-driver or linear amplifier in the plug; AEC generally means a full retimer DSP. Before trusting a reach figure, find out which one is actually in the connector.

And if you came here comparing these against a form factor — OSFP, QSFP-DD, QSFP112 — that is a separate axis entirely, and collapsing the two is the most expensive interconnect mistake there is. Start with OSFP vs DAC vs AOC: You're Comparing Two Different Things.

How far passive copper really goes

This is the table most buying mistakes come down to. Treat it as planning guidance, not specification — real reach depends on conductor gauge, shielding quality, bend radius and how forgiving the host's SerDes is.

Link rate	Passive DAC, typical	Active copper	AOC
10GbE (SFP+)	up to ~7 m	—	100+ m
25GbE (SFP28)	~3–5 m	to ~7 m	100+ m
100GbE (QSFP28)	~3 m	~5 m	100+ m
200G HDR (QSFP56)	~3 m	~5 m	100+ m
400G NDR (QSFP-DD / OSFP / QSFP112)	~1–2 m, sometimes 3 m	~3–5 m	to ~100 m
800G (OSFP / QSFP-DD800)	~1–2 m	~3–5 m	to ~100 m
1.6T (OSFP-XD)	~1 m or less	limited	optics, effectively

The practical consequence at the top of that table: at 400G and 800G, copper is an intra-rack medium. If a design has a 4 m copper run in it, the design needs changing, not the cable. Measure the real cable path — including slack and the trip through the cable manager — not the straight line between racks.

Why the ceiling exists

A passive DAC is twinax copper with a connector housing molded onto each end and no signal regeneration anywhere in between. As the electrical signal travels, it loses energy and picks up noise — insertion loss. The faster the per-lane signalling, the less margin the receiver has left to tell a 1 from a 0. That is the whole reason the usable length keeps shrinking as rates climb from 10G through PAM4 at 400G and 800G. Nothing about the cable got worse; the budget got tighter.

What it looks like when you push past the limit

An over-length passive DAC rarely fails cleanly, which is exactly what makes it expensive to diagnose. Symptoms:

• The module is detected but the link never comes up at all
• The link comes up and then flaps — often under load, or as rack temperature changes through the day
• The link trains at a reduced negotiated speed instead of its rated speed
• Elevated CRC errors and retransmits even while the link looks stable

Before blaming the cable, check FEC settings too — a mismatch produces very similar symptoms. We cover that in DAC Cable No Link? Fix 25G/100G FEC Mismatches.

Passive DAC — what it's good at

• Lowest cost. No optical engine, no separate transceivers. At hundreds or thousands of links the difference compounds.
• Near-zero power. Nothing to feed beyond the host interfaces — and nothing to cool.
• Low latency. No optical conversion, though in practice the difference against an AOC is nanoseconds and rarely changes application behaviour.
• Few failure modes. No active components means little to go wrong on a fixed short run.

Where it starts to hurt: copper is physically bulky, and it gets bulkier at higher lane counts because shielding and gauge have to grow. In a rack with a handful of links that is irrelevant. In a dense AI rack with dozens of 400G links, thick bundles occupy routing space, block access to ports, make service slower and complicate the next upgrade.

Active copper — ACC and AEC

Signal conditioning in the plug buys back the metres that insertion loss took away, at a fraction of the power an optical engine needs. It is the right answer for the run that measures 3.5 m when your cable is rated for 2 — which, in practice, is a very common situation and the one people most often try to solve by buying a longer passive cable and hoping.

ACC and AEC at 400G and 800G are built and coded to the specific host-and-switch pair rather than stocked generically, because the coding and the retimer configuration have to match both ends. Send us the two endpoints and the length and we will quote the correct part instead of selling you a passive cable that will not train. Request a quote →

AOC — active optical

An AOC has optical transceivers bonded into each end of a fibre assembly. Electrical in, light across, electrical out. It installs as simply as a DAC — no separate modules to source or insert — and it removes the copper distance penalty completely.

• Distance. 10 m to 100+ m without a loss-budget calculation.
• Bulk and airflow. Dramatically thinner and lighter than equivalent copper, so more links fit through a cable manager without congestion.
• Serviceability, up to a point. Easier to route, label and trace than heavy copper.

The trade-offs are real: a few watts per end, a fixed factory length you cannot adjust in the field, and a sealed assembly — when one optical engine fails you replace the entire cable, not a module.

AOC or transceivers and fibre?

This is the question that decides large fabrics, and the answer inverts with scale.

An AOC is one part number, factory-matched, cheaper than two modules plus a patch cord. Excellent at eight links. At eight hundred, the sealed assembly becomes the problem: one bad end means pulling the whole cable out of the cable management and running a new one, through a populated rack, with the fabric degraded. Transceivers plus structured fibre cost more up front and are repaired one module at a time in ninety seconds with the fibre left in the tray.

If the cluster has a growth plan, price mean-time-to-repair next to the cable. It never appears in the cost-per-link tables the industry publishes, and past a certain size it dominates them.

Airflow and thermal density

AI racks are the reason this section exists. GPU chassis draw several kilowatts each, and dense racks run from 40 kW to over 100 kW. Cooling becomes a primary design constraint rather than an afterthought, and while cables are only one part of the airflow equation, bundles do influence how air moves at the front and rear of a populated rack — and how quickly a technician can get to a port.

Optics move the heat rather than removing it: roughly 8.5 W for a 400G OSFP module, around 15 W for a twin-port 800G. Populate all 32 cages of a Quantum-2 QM9700 with twin-port optics and you have added something like 480 W to a single rack unit, paid for twice — once at the PSU and once at the CRAC. Copper adds zero. That is the real reason NVIDIA keeps copper inside the rack, not cost-cutting — see what GTC 2026 said about copper.

Coding: this applies to cables, not just optics

A common and expensive assumption is that EEPROM coding is a transceiver problem. It isn't. DACs and AOCs carry EEPROM at each end, and a switch will reject an unrecognised cable as fast as it rejects an unrecognised module. "Any 25G AOC" is not a purchasing specification — the coding has to match the platform at both ends, and the two ends are often different platforms.

Four misconceptions worth clearing up

"DAC is always the best value." Lowest invoice, not always lowest cost. In a dense rack, routing and service time can justify optical.

"AOC is only for long runs." Plenty of teams deploy AOC at 3 m purely for bulk and airflow reasons.

"Fibre is too fragile for a data centre." Modern assemblies are built for this environment. Respect the minimum bend radius and they are dependable.

"Optical power draw makes AOC impractical." A few watts per end against a multi-kilowatt GPU chassis is not the constraint. At several hundred ports it becomes a real line item — which is an argument for counting, not for avoiding optics.

Which should you choose?

Passive DAC if the link stays inside a rack, the run is comfortably inside the table above, purchase cost and power efficiency matter, and the layout is stable.

ACC or AEC if the run lands just past passive reach and you would rather not add optical power and heat to the rack.

AOC if you need real distance, rack density is high, congestion is already a problem, or simpler maintenance and future flexibility are worth the premium.

Transceivers and structured fibre if the fabric is large, will grow, spans rooms or floors, or will be maintained by people who are not you.

Most real deployments use three of the four. Match the interconnect to the link, not to a standard you picked once.

FAQ

Is DAC faster than AOC?

No. Built to the same Ethernet or InfiniBand standard, they carry the same data rate. The choice is about distance, power, airflow and serviceability.

Is latency lower with DAC?

Marginally — a passive DAC performs no optical conversion. The difference is a few nanoseconds and is unlikely to affect application performance in a production AI network.

How far can a passive DAC run at 25G? Can I use one at 10 m?

Plan on 3–5 m at 25G depending on vendor. At 10 m, use an AOC or a transceiver pair — a passive cable at that length will either fail to train or flap under load.

How far can an 800G passive DAC run?

Plan on 1–2 m. At 800G copper is an intra-rack medium.

Does a longer passive DAC draw more power?

No. Passive copper draws negligible power at any supported length. The limit is signal integrity, not power.

Can I mix DAC, AOC and AEC in the same cluster?

Yes, and most production clusters do — matching cable type per link rather than standardizing throughout.

Is AEC compatible with the same ports as DAC and AOC?

Generally yes, in the same QSFP and OSFP cages, but confirm the specific switch and NIC pair before ordering — active cables need matching coding at both ends.

What is the difference between ACC and AEC?

Both are active copper. ACC typically uses a re-driver or linear amplifier and reaches roughly 3–5 m; AEC uses a full retimer DSP and reaches roughly 5–7 m at higher power and cost. NVIDIA's LinkX catalog says ACC and LACC; much of the market says AEC for anything retimed.

What is the difference between an AOC and a transceiver with fibre?

An AOC is a fixed-length assembly with optics permanently bonded to both ends, so it cannot be re-terminated and a single failed end means replacing the whole cable. Separate transceivers plus a patch cord let you change lengths and swap modules individually, at higher up-front cost.

Do DAC and AOC cables need to be coded for my switch?

Yes. Both carry EEPROM at each end and switches reject unrecognised cables exactly as they reject unrecognised optics. We program and link-test every assembly to the platform you specify.

Which is cheaper, DAC or AOC?

DAC, significantly, per link at purchase. Over the life of a large fabric the gap narrows or reverses once repair time is counted — and AOC's reach isn't something DAC can substitute for anyway.

How much do DAC and AOC cables cost?

At Resilient Tec: the MCP1600-C001 100G DAC (1 m) is $59, the MCP1660-W001E30 400G DAC (1 m) is $119, the MCP1650-H001E30 200G DAC (1 m) is $87.45, the MFA2P10-A010 25G AOC (10 m) is $59, the MFA1A00-C003 100G AOC (3 m) is $140 and the MFA1A00-C010 100G AOC (10 m) is $249.90. All new compatible cables with a lifetime warranty. (Prices checked September 28, 2026.)

---

Ready to buy

Tested and shipping from Knoxville, Tennessee. Free U.S. shipping, lifetime warranty, POs and net terms available. Prices and stock checked September 28, 2026.

Passive copper — in rack

• MCP1600-C001 — 100G QSFP28 passive DAC, 1 m — $59.00
• MCP1650-H001E30 — 200G QSFP56 passive DAC, 1 m — $87.45
• MCP1660-W001E30 — 400G QSFP-DD passive DAC, 1 m — $119.00
• MCP4Y10-N001 — OSFP 400G NDR passive DAC, 1 m, DGX H100/H200 — $195.00

Active optical — past copper reach

• MFA2P10-A010 — 25G SFP28 AOC, 10 m — $59.00
• MFA2P10-A015 — 25G SFP28 AOC, 15 m — $63.00
• MFA1A00-C003 — 100G QSFP28 AOC, 3 m — $140.00
• MFA1A00-C010 — 100G QSFP28 AOC, 10 m — $249.90
• MFS1S50-H005E — 200G QSFP56 AOC, 5 m — $989.89

Active copper (ACC / AEC) — built and coded to order. Request a quote →

Longer, shorter, or a platform pair that isn't listed? We build double-programmed DAC and AOC to order — including different coding on each end.

Shop by what you concluded

• DAC Cables — passive copper for in-rack and adjacent-rack links, from $29
• AOC Cables — active optical for cross-row runs, from $59
• Transceivers — for structured fibre and runs past AOC's ceiling
• Breakout Cables — when the spine runs faster than the server NICs
• Fiber — patch cords and MPO trunks
• All Interconnects

Related reading

• OSFP vs DAC vs AOC: You're Comparing Two Different Things — form factor, medium and line rate as three separate decisions
• OSFP DAC Cables: What They Are, When to Use One, and When They'll Quietly Ruin Your Week — the 400G/800G copper detail
• DAC Cable No Link? Fix 25G/100G FEC Mismatches — when the length isn't the problem
• 800G AI Fabric Buying Guide: Copper Where You Can, Optics Where You Must
• Lossless Ethernet for RoCE: Why Cable Quality Affects Tail Latency
• Why the Same Transceiver Works in One Switch and Not Another
• DAC Cables Guide · AOC Cables Guide · Fiber Optic Cables Guide · Breakout Cables Guide · Network Switches Guide
• NVIDIA OSFP Part-Number Crosswalk — moving past QSFP56 to OSFP

Not sure which spec fits your exact switch and NIC combination? Send us both endpoints and the rack distance — we will confirm the part before you order. Shop DAC Cables | Shop AOC Cables | Talk to a Hardware Specialist