What Data Centers Are Missing: The Optical Layer of AI

By Nino Shaptoshvili, VP Sales, PacketLight Networks
Industry expert with over 20 years of experience in data transport and optical networking

TL;DR

AI has changed the rules of data center design. GPU clusters now behave like distributed supercomputers, generating massive east–west traffic across buildings, campuses, and metro regions. Power and cooling still matter, but optical scale has become the real limiter. As AI infrastructure extends beyond a single site, the optical layer defines not only bandwidth and performance, but also isolation and security. DWDM is no longer “telecom infrastructure”; it is core AI infrastructure. Data centers that plan for optical scale from day one will attract serious AI workloads. Those that do not will encounter hard physical and architectural limits.

The future of AI is not built only in racks. It is built on wavelengths.

Insights from DCI Netherlands

For years, data center strategy has revolved around familiar metrics: rack density, megawatts, cooling efficiency, and square footage. Those dimensions still matter, but in the age of AI, they are no longer sufficient. AI has quietly shifted the center of gravity. The limiting factor is no longer how many GPUs fit in a room. It is how effectively those GPUs can communicate, at scale, across distance, and without disruption.

Modern AI infrastructure no longer behaves like a traditional data center. Built around NVIDIA GPUs and tightly coupled accelerators, today’s environments increasingly resemble distributed supercomputers.

Inside a node, communication is handled by NVLink. Between nodes and racks, InfiniBand or Ethernet (including RoCE and Spectrum‑X) takes over. But once GPU clusters extend beyond a single room—across halls, buildings, or multiple sites, the network stops being supporting infrastructure and becomes the system itself.

This is where many traditional data center designs begin to break down.

NVIDIA Is Scaling Faster Than Fiber

A single DGX H100 or B200 pod can generate multiple terabits per second of sustained east–west traffic, particularly during distributed training and frequent checkpointing. Multiply that across multiple pods, buildings, or data centers, and conventional interconnect strategies quickly reach their physical and operational limits.

Pulling new dark fiber is slow, expensive, and often constrained by rights‑of‑way. Cable density and fiber management become increasingly complex, while the power and space consumed by pluggable optics escalate rapidly.

At AI scale, fiber availability, not rack space, becomes the primary constraint.

DWDM: The Layer Nobody Talks About

Dense Wavelength Division Multiplexing (DWDM) has existed for decades, yet it remains oddly absent from many AI infrastructure discussions. Often labeled as “telecom gear,” it is still treated as something separate from data center architecture.

That distinction no longer holds. DWDM enables:

• Deterministic, low‑jitter transport suitable for RDMA‑based workloads
• Fast scaling from 100G to 400G and 800G wavelengths without forklift upgrades
• Campus, metro, and regional GPU clusters over limited fiber infrastructure

In practice, DWDM is no longer an optimization layer. It has become foundational AI infrastructure.

AI Doesn’t Live in One Building

NVIDIA SuperPOD reference architectures already assume physical distribution:

• GPU pods in different halls or buildings
• Storage systems separated from compute
• Control, management, and orchestration planes located elsewhere

Neocloud and GPU‑as‑a‑Service providers are extending this model even further, distributing capacity across multiple data centers within a metro region to improve utilization, availability, and cost efficiency.

Without an optical transport layer designed for scale, this model simply does not work.

InfiniBand, Ethernet, and the Optical Reality

Whether operators standardize on InfiniBand for ultra‑low‑latency training clusters or Ethernet and RoCE for multi‑tenant training and inference, the underlying requirements remain the same:

• Lossless transport
• Predictable, deterministic performance
• Massive bandwidth over limited fiber infrastructure

DWDM does not replace NVIDIA networking. It enables it, extending GPU fabrics beyond a single room or building without compromising performance.

Storage Is the Silent Bandwidth Killer

AI training is as much about data movement as it is about compute cycles. Dataset streaming, checkpointing, replication, and recovery increasingly dominate network flows.

In many AI data centers, storage traffic already exceeds north–south internet traffic. Without a high‑capacity optical layer, storage becomes the hidden choke point that limits training efficiency and overall cluster utilization.

Security: A Consequence of AI Scale

As AI infrastructure expands across buildings, campuses, and metro regions, security stops being a perimeter concern and becomes an architectural one.

AI workloads move some of the most valuable data in the data center: proprietary training datasets, model weights and checkpoints, customer inference data, and control traffic. This data moves east–west, often continuously, and increasingly between sites. Every interconnect becomes part of the attack surface.

DWDM fundamentally changes this security model. Each wavelength is physically isolated at Layer 1, reducing exposure to packet‑level attacks and lateral movement common in shared IP fabrics. Deterministic optical paths also reduce the operational risk introduced by congestion, misconfiguration, or software‑based controls.

Modern DWDM platforms further support line‑rate optical encryption, allowing sensitive AI traffic to remain encrypted in flight without adding latency or compute overhead. At AI scale, security is no longer something that can be bolted on later, it is an outcome of architectural choices made early.

What Data Centers Are Actually Missing

It is not GPUs.
It is not power.
It is not cooling.

What many data centers are missing is optical thinking.

AI‑ready facilities must be designed as architecture‑first systems:

• Multi‑building platforms, not isolated rooms
• Metro‑scale systems, not single‑site assets
• Optical fabrics first, packet networks second

DWDM is no longer a future upgrade or a nice‑to‑have. In the AI era, it belongs in the core architecture from day one.

Designing for the AI Reality

AI infrastructure has crossed a threshold where compute, networking, optics, and security are inseparable. Data centers that continue to treat optical transport as an afterthought will struggle to attract serious AI workloads.

Those that design deliberately for optical scale will quietly become the backbone of the AI economy.