800G still dominates AI training clusters instead of 1.6T

In recent years, AI training clusters have become the most demanding battlefield for high-speed interconnects. As model parameters scale from billions to trillions, bandwidth requirements rise sharply. From the outside, it may seem logical that 1.6T should quickly replace 800G. Yet in real AI training clusters, 800G remains the mainstream choice. This is not a technology lag, but a rational engineering decision.

In an AI training cluster, network performance is not defined by a single link speed. It is defined by system balance, include compute, memory, switching capacity, power, cooling, and cost. Today’s AI training cluster architectures are already well-aligned with 800G. GPU nodes, leaf–spine fabrics, and optical interconnects are designed around 800G lanes, enabling predictable performance scaling. Moving directly to 1.6T often disrupts this balance rather than improving it. From a deployment perspective, 800G sits at a sweet spot:

Ecosystem maturity: DSPs, optical engines, connectors, and testing standards for 800G are well established.

Manufacturing yield: Compared with 1.6T, 800G modules deliver higher yield and better consistency.

Interoperability: AI training clusters require massive port counts, and 800G integrates smoothly with existing switching silicon.

In contrast, 1.6T is still in an early adoption phase. While technically impressive, it introduces higher risk in large-scale AI training cluster rollouts.

Power efficiency is a silent constraint in every AI training cluster. A 1.6T optical module does not simply double bandwidth, it often increases power density disproportionately. This creates challenges in airflow design, thermal budgets, and rack-level planning. By comparison, 800G delivers a more controllable power profile, and makes it easier to scale AI training clusters without redesigning cooling infrastructure.

Most AI training clusters today rely on Clos or Dragonfly+ topologies optimized for 800G lane aggregation. Switching to 1.6T would require new switch ASIC generations, higher-risk optical packaging, revalidation of loss budgets and fiber management. For many operators, upgrading 800G density is simply more efficient than rushing into 1.6T. 1.6T will absolutely have its moment, especially for next-generation AI training clusters beyond 2026. But until power efficiency, ecosystem maturity, and cost curves align, 800G remains the most practical backbone for AI training clusters worldwide.