Active Electrical Cables: A Full Guide to Retimed Copper When Optics Are Overkill
AECs put DSP retimers in copper OSFP/QSFP-DD assemblies for AI scale-out: Credo ZeroFlap 800G SPAN to 7 m, Marvell Alaska A 1.6T beyond 3 m, and why DAC, ACC, AOC, and CPO are different BOM columns.
At 800G and 1.6T, passive twinax runs out of meters before it runs out of ports. Copper still wins the short hop on cost, power, and flap rate. It just needs silicon in the connector.
That product class is the active electrical cable (AEC): a narrow-gauge copper assembly with DSP retimers (digital signal processors that recover clock and data, then retransmit a clean waveform) at each end. Credo's ZeroFlap 800G SPAN SKUs list 4 m to 7 m rack-to-rack reach. Marvell's Alaska A 800G AEC DSP, demonstrated with cable partners at OFC 2025, targets the same 7 m class at 100G per lane, while its Alaska A 1.6T DSP (8×200G) is built for greater than 3 m in-rack copper at 200G per lane.
This guide covers what AECs are, how they differ from DAC (direct attach copper), ACC (active copper with analog redrivers), and AOC (active optical cable), who ships the retimer silicon and finished cables as of September 2026, and where the honest limits sit. Scale-up fabrics live in the NVLink, InfiniBand, and UALink guide. Switch-package light is co-packaged optics. Device-level waveguides are silicon photonics. Host attach is PCIe 6.0. Do not treat AEC copper as a substitute for any of those columns.
Key Takeaways
- An AEC embeds DSP retimers with clock-and-data recovery at both ends, regenerating the signal instead of only amplifying it (Credo ZeroFlap AEC product family; Marvell Alaska A AEC DSP briefs).
- At 800G (8×112G-class lanes), Credo lists SPAN OSFP AECs at 4–7 m, and Marvell's OFC 2025 note says Alaska A 800G DSP enables up to 7 m copper where passive DAC is estimated near 2 m without AEC silicon.
- At 1.6T (8×200G), Marvell's Alaska A 1.6T PAM4 DSP (announced June 2024; generally available as of OFC 2025) targets greater than 3 m in-rack copper on 32 AWG-class assemblies.
- Credo positions AECs as up to 50% lower power than optical alternatives and up to 75% less volume than DACs on ZeroFlap SPAN SKUs, with a 10-year service-life claim on the 800G SPAN page.
- Honest limits: AECs are not NVLink scale-up, not CPO, and not a 100 m DCI fix. Retimers add watts and nanoseconds; 1.6T copper reach is tighter than 800G; platform qualification still gates every hyperscale BOM.
AEC at a glance
| Attribute | Detail |
|---|---|
| What it is | Copper cable assembly with DSP retimers (and often FEC / gearbox helpers) in the MSA connectors; regenerates PAM4 lanes for short-reach Ethernet or InfiniBand links |
| What it is not | Passive DAC; ACC redriver cable; AOC or pluggable optics; NVLink/UALink scale-up; CPO on a switch package; long-haul DCI |
| Core trick | Recover and retransmit a clean electrical waveform so thinner, longer copper stays within BER budgets after passive twinax runs out of meters |
| Key hardware | Retimer DSPs (Credo, Marvell Alaska A, Broadcom-class SerDes IP in integrator designs); OSFP / OSFP-RHS / QSFP-DD / QSFP112 assemblies; CMIS-managed firmware |
| Reach class (Sep 2026) | 800G: Credo SPAN 4–7 m; Marvell Alaska A 800G up to 7 m. 1.6T: Marvell Alaska A targets >3 m in-rack at 200G/lane |
| Standards / ecosystem | HiWire Consortium founded 2019 by Credo; HiWire Spec v1.0 contributed into Open Compute Project Interconnects / AEC work |
| Biggest unsolved problem | 224G/lane cable manufacturing yield and longer 1.6T reach; multi-vendor interoperability discipline; operators confusing AEC with optics or with GPU scale-up fabrics |
DAC vs ACC vs AEC: the copper ladder
Start with the physics. PAM4 at 112G and 200G per lane loses eye margin fast in twinax. Passive DAC has no electronics: lowest cost, near-zero added power and latency, and the shortest reach. Industry and vendor write-ups in 2025–2026 typically put 800G passive DAC near 2–3 m. Marvell's OFC 2025 note estimates passive copper at 800G near 2 m without AEC DSPs.
ACC inserts analog redrivers (often CTLE-style equalizers) that boost high-frequency content. Power stays low, latency barely moves, and reach stretches into the mid-single-digit meters. Noise is amplified with the signal. There is no full clock-and-data recovery.
AEC puts a DSP retimer at each end. The chip samples the impaired waveform, recovers clock and data, and drives a fresh PAM4 stream onto the next segment. That is why AECs can hold 5–7 m class links at 800G on published Credo SPAN and Marvell Alaska A 800G claims, and why 1.6T AECs exist as a product category at all.
| Class | Electronics | 800G reach class | 1.6T note | Power character |
|---|---|---|---|---|
| DAC | None (passive twinax) | ~2–3 m (Marvell cites ~2 m without AEC DSP) | Compresses further; often ~1 m class | ~0 W added |
| ACC | Analog redriver / CTLE | Mid-single-digit meters (vendor-dependent) | Low-watt analog path still shipping in demos | ~1–2 W/end class (secondary guides) |
| AEC | DSP retimer + CDR | Credo SPAN 4–7 m; Marvell Alaska A up to 7 m | Marvell Alaska A 1.6T >3 m in-rack (32 AWG) | Mid watts/end; Credo claims up to 50% less than optical on SPAN |
| AOC / pluggable | Optics + DSP | Tens of meters to 100 m+ | Still the long hop | Higher than AEC on Credo's SPAN comparison |
Why AI racks buy AECs
GPU clusters multiply short links. Each accelerator needs scale-out Ethernet or InfiniBand bandwidth to a leaf or middle-of-row switch. When the hop is inside one or two racks, optics are often overkill on power, cost, and optics failure modes. Copper stays attractive if it can clear the meter count.
Credo markets ZeroFlap AECs specifically for lossless backend RDMA fabrics and "zero soft link flaps," with the 800G family enabling host-to-switch connectivity up to 7 m. The SPAN OSFP-to-OSFP-RHS page adds the commercial claims operators actually care about: up to 50% less power than optical, 75% less volume than DACs, and a 10-year service-life statement.
Marvell frames AECs as copper plus DSP technology originally matured for optical modules, now applied to short-reach AI racks. At OFC 2025 it cited 650 Group's December 13, 2024 Active Electrical Cable report projecting AEC silicon growing 61% per year to $1.3 billion by 2029. Treat that as a market forecast, not a purchase order.
Form factors and SKU shapes
Finished AECs plug into the same cages as pluggables: OSFP, OSFP-RHS (riding heat sink), QSFP-DD, and QSFP112. Credo's catalog separates CLOS (often 1–3 m host/switch), SPAN (longer rack-to-rack, including 4–7 m 800G OSFP SPAN), and SHIFT (breakout / speed-shift assemblies such as 800G to dual 400G-class ends).
Firmware and management matter as much as copper gauge. Production AECs expose CMIS-style management, FEC status, and SerDes telemetry so a cluster can treat the cable like a module. That is also why hyperscale quals take months: you are qualifying silicon, firmware, and mechanicals as one assembly.
| SKU shape | Typical use | Example published reach |
|---|---|---|
| CLOS (short) | In-rack NIC/GPU to ToR | Credo 800G CLOS OSFP: 1–3 m |
| SPAN (long copper) | Rack-to-rack / mid-row | Credo 800G SPAN OSFP–OSFP-RHS: 4–7 m |
| SHIFT (breakout) | Fan-out / speed conversion | Credo 800G SHIFT OSFP to dual OSFP-RHS or QSFP112: up to 7 m on listed SKUs |
| 1.6T AEC | 200G/lane in-rack | Marvell Alaska A 1.6T: >3 m (32 AWG-class) |
Who ships the retimer silicon
Two supply models dominate. Vertically integrated vendors design the retimer and sell finished AECs. System integrators assemble cables around merchant DSPs from Marvell, Broadcom, or similar SerDes houses.
Credo is the loudest integrated AEC brand in Ethernet AI fabrics. Its ZeroFlap pages claim invention of the AEC category and ship 100G through 1.6T product families with in-connector retimers, gearboxes, and FEC.
Marvell sells the Alaska A PAM4 DSP family into cable OEMs. The Alaska A 1.6T part (MV-CHA1600NV class) is an 8×200G retimer announced June 27, 2024 as the industry's first 1.6T AEC DSP, with cable-partner sampling called out for August 2024. As of the April 1, 2025 OFC release, Marvell said the 1.6T AEC DSPs are generally available, with ecosystem demos from 3M, Amphenol Communications Solutions, Broadex, Luxshare-Tech, and TE Connectivity. The same release states Alaska A 800G (100G/lane) delivers connectivity for up to 7 meters.
Integrator brands (Amphenol, TE, Molex, Luxshare-Tech, Broadex, and others) matter for volume and regional manufacturing. The DSP choice still sets the electrical ceiling.
HiWire, OCP, and what "standard" means
Credo founded the HiWire Consortium in September 2019 to specify and certify AECs as a plug-and-play class on top of IEEE and MSA building blocks. HiWire Spec v1.0 later moved under the Open Compute Project Interconnects / AEC workstream so hyperscalers and ODMs could extend speeds and test methods in the open.
That history matters for RFPs. "AEC" on a quote sheet is not enough. Ask which HiWire/OCP profile, which MSA cage, which lane rate, which FEC mode, and which CMIS revision the cable was qualified against on your switch and NIC silicon.
Honest limits
AECs shrink the optics bill. They do not delete optics from the cluster.
- Distance ceiling. Published 800G copper SPAN is a 7 m class story. 1.6T published reach on Marvell's Alaska A is a greater-than-3 m in-rack story. Leaf-to-spine and longer rows still want AOC or discrete pluggables, and package-level optics are a different design (see the CPO guide).
- Power is not free. Retimers draw watts at both ends. Credo's "up to 50% less than optical" claim is relative to optical alternatives on SPAN SKUs, not a zero-watt DAC.
- Latency budget. Each retimer adds a small, finite delay (nanoseconds to low hundreds of nanoseconds depending on design). Fine for most Ethernet scale-out. Wrong tool if you needed NVLink-class scale-up.
- Qualification wall. Hyperscale AECs pass months of unit, system, and pilot testing. A cable that works in a lab tray can still fail a fleet BER or CMIS policy.
- Category confusion. ACC is not AEC. DAC is not AEC. Buying "active copper" without a retimer part number is how teams invent outages.
What happens next
Through late 2026 the fight is 200G-per-lane yield: thinner twinax, 3 nm-class retimer power, and who can push 1.6T AECs past the short in-rack hop without falling back to optics. Operators should measure real tray-to-tray lengths, demand DSP part numbers on every active-copper line item, and keep AEC, ACC, DAC, AOC, and CPO in separate BOM columns.
Inside Deep Tech's take: treat AECs as the default middle distance for AI Ethernet and InfiniBand scale-out when the tape measure says you are still inside copper's honest window. Past that window, stop arguing with physics and buy light.
FAQ
What is an active electrical cable (AEC)?
An AEC is a copper interconnect with DSP retimers at each end that recover and regenerate the electrical signal.
That regeneration is what separates AECs from passive DAC twinax and from ACC cables that only apply analog equalization.
How is an AEC different from a DAC or ACC?
DAC is passive copper with no electronics. ACC adds analog redrivers. AEC adds full DSP retiming with clock-and-data recovery.
In practice that ladder maps to short, medium, and longer copper reaches at 800G and 1.6T before optics take over.
How far can an 800G AEC reach?
Published Credo ZeroFlap 800G SPAN OSFP SKUs list 4 m to 7 m for rack-to-rack links.
Marvell's OFC 2025 materials likewise describe Alaska A 800G AEC DSP copper connectivity up to 7 meters, versus roughly 2 meters for passive copper without AEC silicon in Marvell's estimate.
What changes at 1.6T?
Lane rates move to 200G-class PAM4, so copper reach compresses versus 800G.
Marvell's Alaska A 1.6T AEC DSP (generally available as of OFC 2025) targets greater than 3 meters for in-rack 32 AWG-class assemblies.
Who makes AEC retimers and finished cables?
Credo sells vertically integrated ZeroFlap AECs. Marvell sells Alaska A DSPs into cable OEMs such as Amphenol, TE, Luxshare-Tech, Broadex, and 3M partners shown at OFC 2025.
Broadcom-class SerDes IP also appears in integrator designs; always ask which DSP is inside the assembly you are buying.
Should I use AECs instead of optics in an AI cluster?
Use AECs when the measured hop fits the qualified copper envelope and you want lower power and cost than optics on that hop.
Keep optics (AOC, pluggables, or CPO) for longer rows, structured fiber, and package-level switch I/O; keep NVLink-class fabrics for GPU scale-up.
Are AECs standardized?
Yes, in the sense that HiWire Consortium work (founded 2019) produced Spec v1.0 that moved into Open Compute Project Interconnects / AEC activities.
You still need platform-specific qualification against your switch, NIC, FEC, and CMIS stack.
Primary sources for this guide include Credo's ZeroFlap AEC product pages (including the 800G SPAN OSFP-to-OSFP-RHS datasheet claims) and Marvell's June 27, 2024 Alaska A 1.6T announcement plus April 1, 2025 OFC 2025 AEC ecosystem release. Market sizing cited via Marvell's reference to 650 Group's December 13, 2024 Active Electrical Cable report. Figures are dated as of September 2026 editorial review.
