800G Data Center Interconnect Selection Guide 2026
DAC · ACC · AEC · AOC · Optical Transceivers — the complete engineer's framework for choosing the right interconnect for every link in your AI data center. 800G · AI Interconnects · NVIDIA · Updated February 2026.
📋 Table of Contents
10 comprehensive sections — jump to any topic
⚡ 1. Why 800G Broke the Old Playbook
At 400G, interconnect selection was a two-step process: measure the distance, pick copper or fiber. Passive copper comfortably reached 3–5 meters. Multimode fiber handled everything from the rack to the end of the row. Done.
800G changed the underlying physics. Each of the eight lanes now runs at 112 Gbps using PAM4 signaling — four voltage levels instead of two — pushing the Nyquist frequency to approximately 28 GHz. At that frequency, copper losses from skin effect and dielectric absorption roughly double compared to 400G's 56G PAM4 per lane. The result: passive copper cables maxed out at approximately 2 meters, down from 3–5 meters at 400G.
That single change rippled through every interconnect decision in the data center and created a new problem: a 3–7 meter "dead zone" too far for passive copper but too close to justify the cost and power of optics. Into that gap emerged Active Electrical Cables (AEC) — a technology that barely existed at 400G but is rapidly becoming the most consequential interconnect category at 800G.
112G PAM4 per lane doubles copper loss — passive DAC maxes out at ~2 m versus 3–5 m at 400G
3–7 m links: too far for passive copper, too close and costly for optics — AEC fills this critical gap
RS(544,514) KP4 FEC on every 800G link adds 50–100 ns latency per hop — always check if specs include this
Digital retimer cables cover 5–7 m reliably — the new default for inter-rack connectivity at 800G
The physics also forced a rethinking of Forward Error Correction. FEC is now mandatory on every 800G link. The standard RS(544,514) KP4 code corrects a raw bit error rate of 2.4×10⁻⁴ down to less than 1×10⁻¹⁵ — but adds 50–100 nanoseconds of latency per hop. When you see latency comparisons between cable types, always check whether they include FEC processing or just the cable propagation delay.
📊 2. The Five Interconnect Types at a Glance
Before going deep on each type, here's the complete picture. This is the unified comparison that covers all five 800G interconnect types across the metrics that drive real deployment decisions.
| Attribute | Passive DAC | ACC / LACC | AEC ⭐ | AOC | Optical Xcvr |
|---|---|---|---|---|---|
| How it works | Pure copper, no electronics | Copper + analog equalizer (CTLE) | Copper + digital retimer (full CDR) | VCSEL TX → MMF → PIN RX | Pluggable module + separate fiber |
| Max reach | ~2 m | 3–5 m | 5–7 m (9 m demo'd) | 30–100 m | 100 m – 10 km |
| Power/link | <0.15 W | 1.5–3 W | ~10 W | 12–17 W | 14–18 W/module |
| Added latency | ~5 ns/m | <15 ns | 50–100 ns | 100–200 ns | 100–200 ns |
| Cable diameter | 10–14 mm | 8–12 mm | 5–7 mm | 3–4 mm | 3 mm (fiber) |
| Weight | ~45 g/m | ~35 g/m | ~20 g/m | ~9 g/m | ~5 g/m (fiber) |
| Bend radius | 100–140 mm | 80–120 mm | 30–50 mm | 30–40 mm | 15–30 mm |
| Relative cost | 1× | ~2× | 2–3× | 4–8× | 5–15× |
| MTBF | ∞ (no electronics) | Very high | 100M hours (Credo) | 400K–900K hours | 400K–900K hours |
| Serviceability | Replace cable | Replace cable | Replace cable | Replace cable | Swap module, keep fiber |
Quick Notes on Each Type
Passive DAC is the workhorse for intra-rack links under 2 meters. Zero power, lowest cost, lowest latency (~5 ns/m). IEEE 802.3ck specifies 2 m maximum at 112G PAM4 — real deployments typically cap at 1.5 m for margin. The other challenge is physical: at rack scale with dozens of cables, the 10–14 mm outer diameter creates real airflow obstruction.
ACC / LACC adds lightweight analog equalization (CTLE circuits) inside each connector, extending reach to 3–5 m while adding only 1.5–3 W per link. NVIDIA markets their active copper as "LACC" (Linear Active Copper Cable). Unlike AEC, ACC uses analog signal conditioning — not digital retiming — providing near-zero added latency with no FEC interaction. ACC occupies a narrow but important niche: links just beyond passive DAC reach that don't need full AEC regeneration.
AOC embeds VCSEL transmitters and PIN receivers inside each connector, converting to 850 nm multimode light for 30–100 m reach. However, AOC at 800G draws 12–17 W, uses aging-prone VCSELs (practical lifespan 5–7 years in clean environments), and requires full cable replacement on failure. AOC is not offered for 100G-PAM4 800G OSFP twin-port on NVIDIA platforms — check your qualified optics list. As AEC extends copper to 7–9 m, AOC's primary remaining use case is links exceeding 10 meters or environments requiring EMI immunity.
Optical Transceivers + Fiber is the modular approach — a pluggable module in the OSFP cage connects to a separate fiber patch cord. If a transceiver fails, swap the module, keep the fiber. This modularity is the primary advantage: the OS2 single-mode fiber you install for 800G DR8 today will support 1.6T and beyond without replacement. Module options span SR8 (50–100 m on multimode), DR8 (500 m on single-mode, with breakout to 2×400G/4×200G/8×100G), 2×FR4 (2 km), and 2×LR4 (10 km).
🔌 3. AEC: The 800G Breakout Story
If there's a headline in 800G interconnects, it's the emergence of Active Electrical Cables as a distinct and increasingly dominant category. Understanding why requires understanding what's inside each cable type.
Inside the Cable: How Each Type Works
The key distinction is in the electronics. Passive DAC has none — signal quality degrades with every millimeter of copper. ACC adds analog equalization (CTLE) that amplifies and reshapes the signal, boosting reach to 3–5 m without fully regenerating it. But AEC goes further: it uses digital retimer chips with full Clock and Data Recovery (CDR) at each end, completely regenerating a clean signal. This compensates for over 38–40 dB of channel loss, enabling reliable links up to 7 meters — and recently, Marvell and Infraeo demonstrated a 9-meter 800G AEC at the OCP Global Summit.
Market Momentum
- $1.3B projected AEC silicon market by 2029 — 61% CAGR (650 Group)
- 272% Credo YoY revenue growth, Q2 FY2026 — driven by hyperscaler AEC demand
- 100M hours MTBF claimed by Credo for AEC — ~100× better than optical transceivers
- ~Cat6 AEC cable diameter (5–7 mm) vs 10–14 mm for passive DAC
- Credo Q2 FY2026 revenue: $268M — powered by four hyperscaler customers each >10% of revenue
Key Silicon Vendors
- Credo — HiWire ZeroFlap: zero soft link flaps, critical for RDMA/lossless AI networks
- Marvell — Alaska A DSP: first 1.6T AEC DSP, generally available
- Broadcom — BCM87850/87854 retimer
- Semtech — CopperEdge: 1.6T active copper, sub-100 ps latency
- xAI endorses Credo ZeroFlap AECs for 100,000+ GPU clusters — eliminates soft link flaps
Credo's financials tell the story of the market: Q2 FY2026 revenue hit $268 million — up 272% year-over-year — powered by AEC demand from four hyperscaler customers each contributing over 10% of revenue. Their CEO stated publicly that AECs are becoming the de facto standard for inter-rack connectivity and are displacing optical connections up to 7 meters.
Linear Pluggable Optics: The Near-Term Power Win
While AEC is the breakout story for copper, LPO is the emerging story for optics. Linear Pluggable Optics remove the DSP from the transceiver module, relying instead on the host switch ASIC's built-in SerDes. The result: LPO versions of 800G DR8 consume as little as 8.5 watts versus 14–17 watts for conventional modules — a 40–50% power reduction. Latency drops from 8–10 ns per hop (DSP processing) to under 3 ns.
The LPO MSA has over 50 members, and the 100G-DR-LPO spec was completed ahead of OFC 2025. Broadcom's Tomahawk 5+ and switches from Arista and Juniper support LPO natively. Adoption projections vary — some analysts predict LPO could represent over a third of 800G transceiver shipments by 2027. At 1,000 ports, the annual energy cost difference between DSP-based and LPO modules exceeds $50,000.
🗺️ 4. Where Each Cable Goes in an AI Cluster
Every production AI cluster uses multiple interconnect types. The question isn't "which cable should I use" — it's "which cable goes where." Here's how the hyperscalers actually deploy, based on published architectures from NVIDIA (GB200 NVL72), Meta (24K-GPU cluster, SIGCOMM 2024), and xAI (Colossus, 200K GPUs).
What the Hyperscalers Confirmed
NVIDIA GB200 NVL72: Over 5,000 copper NVLink cables inside each rack (scale-up). Optical transceivers — 400G SR4 with ConnectX-7, upgrading to 800G DR4 with ConnectX-8 — for the scale-out network (rack-to-rack). The pattern: copper inside, optics outside.
Meta's 24K-GPU cluster (SIGCOMM 2024): Copper DAC for all intra-rack connections, single-mode fiber with pluggable transceivers for inter-rack. Meta chose copper for intra-rack specifically because passive DAC has significantly better MTBF than optical — in a system where one link failure stalls thousands of GPUs, per-link reliability is paramount.
xAI Colossus (200K GPUs): Uses NVIDIA Spectrum-X Ethernet with Credo ZeroFlap AEC for inter-rack connectivity. The largest known GPU cluster chose AEC over AOC for the 3–7 m zone.
Breakout Cables: Bridging 800G Switches to 400G NICs
The most common deployment today doesn't use pure 800G-to-800G links. Instead, an 800G OSFP twin-port switch port (containing two independent 400G links sharing a single OSFP cage) breaks out to two 400G NIC ports. This is the standard topology for DGX H100 and H200 SuperPODs, where each QM9700 switch port connects to two ConnectX-7 NICs.
These breakout cables require different connector types at each end — IHS (finned-top) at the switch, RHS (flat-top) or QSFP112 at the NIC. Available configurations include passive DAC breakout (≤2 m), ACC/LACC breakout (≤5 m), and AEC breakout (≤7 m). The breakout approach reduces effective per-port cost by roughly 40% compared to discrete 400G links. As ConnectX-8 deployments increase through 2026–2027, the topology will shift toward native 800G-to-800G connections.
Testing and Validation: What Every 800G Link Needs
Every 800G link requires FEC-aware testing before production traffic flows. Pre-deployment testing covers BER testing using PRBS13Q or PRBS31Q pattern generators and eye diagram analysis to verify signal quality margin. Baseline DDM readings (transceiver temperature, TX/RX power, laser bias) become your reference for ongoing health monitoring.
Connector inspection is non-negotiable for optical links. A single dust particle on an MPO ferrule can cause 0.2–0.5 dB of additional attenuation — enough to push a marginal link into failure. Industry data indicates that 70% of fiber optic link failures trace to contamination or connector damage, not hardware defects. Always inspect and clean MPO/APC connectors before mating.
For ongoing monitoring: set DDM alarm thresholds in your NMS. Rising TX bias current indicates VCSEL aging. Declining RX power with steady TX power indicates fiber degradation or connector contamination. These early warning signals let you schedule replacements during maintenance windows rather than responding to failures that stall GPU training runs.
🧭 5. The Visual Decision Framework
Distance is the primary decision driver — it eliminates options immediately. After distance, the decision branches into power budget, scale, and migration timeline. Here's the framework for deployment planning.
| Distance Zone | Cable Type | Power / Link | Approx. Cost / Link | Primary Use |
|---|---|---|---|---|
| 0–2 m (intra-rack) | Passive DAC | <0.15 W | $105–$300 | GPU-to-switch within the same rack |
| 2–5 m (near-rack) | ACC / LACC | 1.5–3 W | ~$600 | Links just beyond passive DAC reach |
| 3–7 m (adjacent racks) ⭐ | AEC | ~10 W | $500–$1,000 | Inter-rack — the 800G sweet spot |
| 10–100 m (cross-row) | AOC or SR8 | 12–17 W | $2,700+ | Row-to-row or end-of-aisle runs |
| 100 m – 10 km (campus / DCI) | DR8 / FR4 / LR4 | 14–18 W | $900+ (3P) / $6,319 (OEM) | Inter-building, DCI, hyperscale spine |
Secondary Decision Factors
Power Budget
- DAC: ~$1.5K/yr per 1,000 ports (3-yr cost)
- AEC: ~$98K/yr per 1,000 ports (3-yr cost)
- Optical: ~$167K/yr per 1,000 ports (3-yr cost)
- 3-yr AEC vs Optical delta: $500K+ at scale
- Evaluate LPO if platform supports — 40–50% power savings
Platform & 1.6T Checklist
- OSFP IHS (switch) or RHS (NIC) at each port? They are NOT interchangeable
- NDR = DAC/MMF ok; XDR = no DAC, no MMF
- Check vendor qualified optics list before ordering
- Upgrading within 3 years? Install 16-fiber MPO trunks now
- Prefer OSFP over QSFP-DD — fiber survives 1.6T, copper replaces
💰 6. TCO: The Numbers Nobody Publishes
Purchase price is the number everyone compares. It's also the least useful number for deployment decisions. Total cost of ownership includes capital cost, power consumption, cooling overhead (multiply by PUE), sparing requirements, and failure-driven downtime.
Capital Cost: What You'll Actually Pay
| Interconnect | Street Price / Link | Relative | Notes |
|---|---|---|---|
| 800G Passive DAC (0.5–2 m) | $105–$300 | 1× | Third-party pricing |
| 800G ACC (3–5 m) | ~$600 | ~2× | Third-party pricing |
| 800G AEC (3–7 m) | $500–$1,000 | 2–3× | Retimer vendor dependent |
| 800G AOC (1–100 m) | $2,700+ | 4–8× | Third-party pricing |
| 800G DR8 (third-party) | $900–$1,200 | Module only | Add fiber cost separately |
| 800G DR8 (OEM-branded) | Up to $6,319 | 2–5× third-party | NVIDIA/Cisco coded |
The OEM premium is enormous. Brand-name transceivers cost 2–5× their third-party equivalents. At hyperscale, SemiAnalysis found this delta can account for nearly 10% of a cluster's total cost of ownership. Third-party optics validated against your platform's qualified optics list deliver identical performance at a fraction of the cost.
Power Cost: The Number That Compounds
At $0.08/kWh and PUE 1.4, the 3-year power and cooling cost per 1,000 ports tells a stark story: Passive DAC (<0.15 W) runs ~$4,500 total. ACC (~3 W) reaches ~$87,000. AEC (~10 W) runs ~$294,000. AOC (~14 W) reaches ~$411,000. Full optical (~34 W per module pair) tops $1,000,000. The AEC vs AOC gap — both covering 3–10 m — shows a $117,000 three-year power advantage for AEC per 1,000 ports. The cheapest cable is not always the cheapest deployment.
Reliability also factors into TCO. Passive DAC has effectively unlimited MTBF (no electronics). Credo claims AECs achieve 100 million hours MTBF — roughly 100× better than optical transceivers, which typically rate 400,000–900,000 hours. Standard practice is to stock 2–3% of deployed optical transceivers as spares. For 1,000 optical ports, that's 20–30 spare modules at $900–$1,200 each — $18K–$36K in additional inventory.
🧵 7. Cable Management at 10,000-GPU Scale
Cable management in AI clusters is an entirely different discipline than enterprise data centers. An enterprise deployment with 100 racks might have 5,000 cables. An AI training cluster with the same rack count can exceed 70,000.
The Cable Counts Get Serious Fast
NVIDIA's DGX SuperPOD Design Guide provides official numbers: a single Scalable Unit (248 GPUs) requires 508 compute network cables plus hundreds more for storage and management. At 4 Scalable Units (1,016 GPUs), that's 2,044 compute cables. Analysis by Enfabrica showed that a 32,256-GPU three-tier cluster requires approximately 73,728 total cables, and at 100,000+ GPUs, cable termination points reach 4.8 million.
Physical Cable Comparison
- Passive DAC: 10–14 mm OD · ~45 g/m · bend radius 100–140 mm
- AEC: 5–7 mm OD · ~20 g/m · bend radius 30–50 mm
- AOC / Fiber: 3–4 mm OD · ~9 g/m · bend radius 30 mm
- At 1,000 cables: DAC = 90 kg total weight; AEC = 40 kg; Fiber = 15 kg
- DAC bundles can add 1,000–1,500 lbs per GPU rack + airflow obstruction
Airflow & Thermal Impact
- Modern AI racks run at 40–140 kW — dense DAC bundles obstruct critical airflow paths
- Meta achieves PUE 1.1 with overhead cable routing — works best with thinner cables
- NVIDIA recommends ≤50% cable tray fill and 12-inch separation between data and power trays
- AEC at 5–7 mm is the practical alternative to DAC at dense port counts
- Every percentage point of utilization improvement is worth $80K–$120K annually per 512-GPU cluster
⚠️ 8. Six Mistakes That Delay Deployments
These are the errors that show up repeatedly in real procurement cycles. Each one has cost real teams real time and money.
🔮 9. Your 800G Decision Is a 1.6T Decision
The 800G-to-1.6T migration will happen approximately twice as fast as previous generational upgrades, with volume 1.6T switch deployments projected for late 2026. What you install today directly impacts your upgrade path.
What Survives the Transition — and What Doesn't
What Survives
- Your single-mode fiber plant survives. OS2 fiber installed for 800G DR8 will support 1.6T. Install 16-fiber MPO trunks now even if you're only lighting 8 fibers — avoids re-cabling.
- Your OSFP cage infrastructure mostly survives. OSFP1600 (8×200G lanes) fits existing OSFP cages.
- AEC silicon is already there. Marvell announced the first 1.6T AEC DSP (200G/lane) in April 2025 — generally available. AEC vendor ecosystem and deployment patterns carry directly into 1.6T.
- The alternative OSFP-XD form factor does NOT fit existing cages — it requires new switch hardware.
What Doesn't Survive
- Your copper cables will not survive. All DAC, ACC, and AEC cables are designed for 100G/lane. 1.6T requires 200G/lane — entirely different electrical characteristics.
- Every copper cable replaces at the 1.6T transition. This is normal — copper has always been generation-specific.
- Budget for full copper replacement as part of your 1.6T migration plan.
- Invest in structured single-mode fiber for any link designed to last more than one generation.
✅ 10. Selection Checklist
Use this for every 800G interconnect procurement decision. Walk through in order — by the bottom, your cable choice will be clear.
Putting It All Together
The 800G interconnect landscape is more complex than any previous generation — five cable types, two incompatible OSFP form factors, a generational shift from multimode to single-mode fiber, and an AEC category that barely existed three years ago. But complexity doesn't mean confusion when you follow a structured process.
Start with distance. Then check platform compatibility. Factor in power at your specific scale. Think one generation ahead. The most successful deployments plan interconnect infrastructure as a first-class workstream alongside compute and power — not as an afterthought. They use multiple cable types matched to distance zones. And they order early, because cables that cost hundreds of dollars become bottlenecks for equipment worth millions.
