OCP Global Summit 2025 | Credo | Wired for Success: Enhancing AI Cluster Reliability with Copper Cables

Introduction

At OCP Global Summit 2025, Credo brought a topic that seems old-fashioned but is being re-valued in the AI era — the return of copper in AI data centers.

While the industry focuses on optical interconnect and CPO, Credo used real deployment data and reliability analysis to argue: for short-reach connections in GPU-dense clusters, copper is not only still viable but a key option for improving reliability.

Content

1. The AI-Era Network Bottleneck: Reliability Matters More Than Bandwidth

As AI training clusters grow from 16K GPUs to 3M GPUs and more than 5 GW of power, network reliability becomes a decisive factor in system operation.

Credo proposed a new metric:

MTBF-LF (Mean Time Between Link Flap) — the average interval between momentary link drops.

At this scale, even a link flap lasting just a few seconds can interrupt a training job across thousands of GPUs and force a rollback to the last checkpoint, at a steep cost in time and money.

Credo emphasized: as optical modules grow more complex and temperature and connector variables multiply, copper actually provides a more stable transmission foundation.

2. AI Network Tiers: Scale-Up vs. Scale-Out

Credo divides the modern AI network into two tiers:

• Scale-Up Network: high-speed GPU-to-GPU interconnect within the rack (e.g., NVLink), using TCP-based communication with high reliability.

• Scale-Out Network (Backend Fabric): GPU communication across racks, mostly UDP-based transport — fast, but without self-healing mechanisms.

When a link flap occurs in the scale-out tier, the entire GPU cluster can briefly lose contact, causing compute efficiency to plunge.

Credo's view: deploying high-reliability copper (Active Electrical Cable, AEC) on the "first hop" closest to the GPU can significantly reduce these interruptions.

3. Why Is Copper Coming Back in AI Networks?

In the past, copper cables were seen as bulky, short-reach, and hard to manage.

But Credo's AEC (Active Electrical Cable) changes that:

• AEC structure: retimer chips are embedded at both ends of a traditional passive copper cable (DACC), splitting a long link into three short channels and reducing signal degradation.

• Performance and characteristics:

  • Reach of up to 7 meters, enough to cover rack-to-rack links.

  • MTBF > 100M hours (optical modules: about 10M hours).

  • Power consumption only about half that of optical modules.

  • Cable diameter as thin as 6mm, with a small bend radius and flexible routing.

  • Deployments already exceed 2 million AECs, with 6 billion cumulative operating hours and no flaps.

Credo's signature purple AEC cables have been adopted by multiple cloud customers and have become one of the standard cabling options for high-reliability GPU clusters.

4. 1.6T Copper and Future Architectures

At this year's summit, Credo showcased its next-generation 1.6T AEC product line (AEC 1.60):

• Supports 200G/lane, 8W@800G / 13W@1.6T.

• Reach of up to 6 meters, while keeping a thin, Cat6-class cable diameter.

• Two configurations are offered:

  • 1.60-to-1.60: for the NVIDIA CX9 NIC.

  • 1.60-to-2x800G: for the CX8 NIC.

AEC 1.60 is designed specifically for the NVIDIA "Reuben" architecture (Blackwell's successor) and is expected to support higher-density GPU racks and liquid-cooled architectures.

5. GPU Density and Liquid Cooling Drive Copper's Next Evolution

As GPU rack density rises and liquid cooling spreads, rack-to-rack distances shrink — an ideal environment for copper.

Generation	GPU Model	GPUs per Rack	Cooling	Copper Reach Needed
Hopper generation	NVL72	72 GPUs	Air cooling + partial liquid cooling	7 m
Blackwell generation	NVL144	144 GPUs	Liquid cooling	5 m
Reuben Ultra generation	NVL576	576 GPUs	100% liquid cooling	1.5–2 m

Credo noted that advances in GPU density and liquid cooling re-amplify copper's advantages of short reach, low power, and high reliability.

Reuben Ultra (2027) is set to become the first copper-centric "1.6T-class liquid-cooled cluster" architecture.

6. Hardware Demos and Ecosystem Integration

On the OCP show floor, Credo also showcased:

• NVL404 and NVL576 GPU rack prototypes, fully cabled with purple Credo AECs.

• Complete cable routing, bend radius, and space-utilization designs, demonstrating that high-density GPU racks can stay tidy and reliable with copper.

• Early cabling designs for the Reuben architecture, shown together with partners including NVIDIA, Jabil, and Supermicro.

Conclusion

Credo's OCP 2025 talk prompts a fresh look at the value of copper in AI data centers.

While optical interconnect focuses on long-range, high-speed transmission, copper still plays an irreplaceable role in the "reliability layer of short-reach, critical links."

As GPU architectures evolve toward high density and liquid cooling, AEC's low power, high reliability, and low maintenance make it copper's second spring in the AI era.

Extended Perspective

1. Technical takeaways

   • Credo's view reflects an "era of division of labor between optics and copper":

     Optical modules handle long-reach links (scale-out), while copper focuses on short-reach high speed (scale-up).

   • With AEC integrating DSP/SerDes, electrical signal-quality control has entered the silicon era, making copper cables "smart" too.

2. Supply chain observations

   • Credo has its own DSP, SerDes, manufacturing, and test labs, forming a fully in-house, vertically integrated chain — key to its product reliability.

   • The spread of AECs also makes it easier for ODM system makers (e.g., Supermicro, Jabil) to integrate multi-vendor GPU architectures, reducing module dependency.

3. Market trends

   • In future AI data centers, networks will take on a two-tier structure: "Optical for Scale-Out, Copper for Scale-Up."

   • This means optical packaging and copper cable products will no longer compete but develop in parallel, together forming high-reliability, high-efficiency AI infrastructure.