OFC 2026 - The Battle for AI Cluster Interconnect: The Paradigm Shift from XPO to CPO and THz Dielectric Waveguides - Microsoft, Arista, Marvell, AttoTude, nEye

Introduction

At the 2026 Optical Fiber Communication Conference (OFC 2026), the evolution of AI cluster interconnects became the central issue in the compute race. AI model size is growing 3x every two years, while network and memory bandwidth grows only 1.5x every two years, creating a serious "I/O wall" challenge. The conference showed how interconnect technology is seeking a new balance among energy efficiency, thermal management and physical limits as the industry moves from NVIDIA's Blackwell architecture in 2025 (120 kW per rack) to Rubin Ultra in 2028 (600 kW per rack).

Deep Dive: Core Technologies and Big-Player Strategies

1. Microsoft: Defining System-Level Metrics for the Scale-up Era

Microsoft stressed that optics is moving beyond traditional scale-out (data center level) into scale-up (accelerator rack level).

• System-level integration: Azure's strategy is "Silicon to Systems," including its in-house Azure Cobalt CPU and Maia AI accelerator.

  
  

• Hard requirements for scale-up: Fotini Karinou laid out what scale-up interconnect must deliver: reach > 20 m, unidirectional bandwidth > 50 Tbps, round-trip latency < 1000 ns, and energy efficiency below 4 pJ/b.

  
  

• Driving standardization: Microsoft is pushing standards such as OCI-MSA (focused on low-power WDM) and XPO to handle an increasingly complex optical stack.

2. Arista Networks: XPO and the Last Line of Defense for Pluggables

Andy Bechtolsheim made a striking forecast: by 2028, AI will need 1 billion 1600G-equivalent optical units, driven mainly by scale-up applications whose bandwidth demand is ten times that of scale-out.

• XPO's extreme packaging: XPO (Extra-dense Pluggable Optics) packs the functionality of 8 OSFP modules into a single liquid-cooled module, achieving ultra-high density of 12.8T.

  
  

• Structural cost optimization: XPO can cut rack, power busbar and plumbing requirements by 75%, and even shrink AI data center floor space by 50%.

  
  

• A leap in reliability: Liquid cooling lowers laser operating temperature by 20-25°C, and the failure rate (FIT) of a single 12.8T module is expected to be below 20.

3. Marvell: Photonic Fabric and the Silicon Photonics CPO Roadmap

Marvell showed a full product line optimized for scale-up networks, aimed at breaking the physical barrier of copper reach (< 2 m).

• COUPE CPO technology: Built on TSMC N3P, it integrates the EIC and PIC via hybrid bonding to achieve 1 Tbps/fiber DWDM transmission.

  
  

• Photonic Fabric chipset: Gen 1 delivers 16 Tbps of bandwidth (exceeding HBM3E); Gen 2 targets 64 Tbps (matching HBM4), with energy efficiency below 3.1 pJ/b.

  
  

• A diversified portfolio: Beyond CPO, Marvell also offers a 6.4T Light Engine supporting NPO and liquid-cooled system designs.

4. AttoTude: Terahertz (THz) Dielectric Waveguides as a Disruptive Leap

Dave Welch put forward a highly disruptive idea: instead of relying on lasers, use transistors to generate terahertz photons.

• THz ASIC approach: 800 GHz ASICs are produced in standard electronics fabs and transmit over dielectric waveguides rather than conventional fiber or copper.

  
  

• Extreme performance: At 448G PAM4, energy efficiency is 2.8 pJ/b, reach is 20-40 m, and reliability matches copper at a 1-billion-hour MTBF.

  
  

• Cost structure: Because no precision optical alignment is needed (alignment tolerance relaxes to 10 um), its cost structure will match that of copper links.

5. nEye Systems: 2D OCS Chips Drive Compute Pooling

Ming Wu explained how optical circuit switching (OCS) is evolving from Google's 3D MEMS to 2D silicon photonics chips.

• 2D OCS advantages: Compared with traditional 3D OCS, the 2D architecture uses under 5 W of power (20x lower), has latency under 5 ns (10x lower), and switches in under 5 us.

  
  

• Key to compute pooling: Fast switching lets OCS flexibly expand high-bandwidth domains within the scale-up domain and improves flexibility for inference workloads.

Expert Consensus and Points of Divergence

Dimension	Industry Consensus	Key Disagreements
Transmission medium	Copper's reach limit at 224G/448G (< 2 m) means it cannot meet cross-rack scale-up needs.	AttoTude argues THz waveguides should replace fiber; Arista still sees active copper as a strong option within 5 m.
Packaging form	CPO is the ultimate long-term path to the highest bandwidth density and lowest power.	Arista believes the CPO supply chain is maturing too slowly and argues XPO is the answer for the 1-billion-unit market in 2028.
Cooling	Liquid cooling is essential for high-density racks and very large optical modules.	How to achieve modular serviceability in CPO systems remains an open question.

Simple Tech Trend's Take:

OFC 2026 sent a clear signal: interconnect technology is being restructured along the boundary between "electronic" and "optical."

1. XPO will be the mainstay for the next three years: CPO may be perfect on technical metrics, but the "time to volume" problem Andy Bechtolsheim pointed out is its Achilles' heel. By solving the thermal pain point of pluggables with liquid cooling, XPO will significantly delay the point at which CPO goes mainstream.

   
   

2. Scale-up networks are becoming proprietary: Discussions from Microsoft and Marvell suggest scale-up interconnect will no longer chase general-purpose Ethernet standards, but shift toward protocol-adaptive architectures like Photonic Fabric in pursuit of minimal latency.

   
   

3. THz dielectric waveguides are the biggest wildcard: If AttoTude's approach reaches HVM, it would fundamentally change the optical communications industry's reliance on lasers, a long-term systemic risk for traditional laser chip makers.