NTT Global Data Centers says optical technology will not eliminate the need for additional fiber as AI traffic grows.
Bruno Berti, SVP of global product management at NTT Global Data Centers, framed the shift around AI traffic inside data centers rather than classic cloud patterns. Traditional cloud traffic is relatively predictable; AI training and large-scale inference create heavy east-west flows as GPUs talk to each other.
How AI changes the network fabric
Berti said AI network requirements need to become flatter, more deterministic, and more resilient. That is pushing higher-capacity fabrics, advanced optical interconnects, and tighter integration between compute, storage, and networking.
He also said the challenge now extends beyond networking itself. Power density, cooling, fiber density, fiber routing, campus architecture, and links to edge and cloud ecosystems all need to be designed together.
What higher-speed optics do and do not do
Higher-speed optics can extend the life of existing fiber at the data-center and metro levels, but optics alone will not remove the need for new fiber as AI traffic scales. The article specifically mentions 400G, 800G, and 1.6T as current higher-speed technologies.
Berti said network infrastructure is largely keeping pace with compute expansion, but that depends on deliberate planning and scaling alongside compute. AI workloads are sensitive to latency, bandwidth, and packet loss within GPU clusters and between connected facilities.
AI traffic is forcing physical rebuilds inside facilities
Large-scale AI training depends on synchronized communication between GPUs, which creates massive east-west traffic and favors architectures that minimize hops, reduce latency, and keep performance predictable. Optical networking is taking a larger role deeper inside AI facilities and across campuses.
Berti said network design is becoming more closely integrated with the physical data-center environment, including power density, cooling strategy, rack layout, and fiber routing. More space is being dedicated to bringing additional fiber and conduits into data centers.
Read for data-center operators and AI infrastructure planners
The practical read for data-center operators is that optics can stretch existing plant, but they cannot replace new fiber once AI traffic keeps scaling. Planning now has to cover compute, networking, cooling, and floor space together, or the network becomes the bottleneck.
That also matters for AI infrastructure teams trying to scale from hundreds of gigabits to terabits without disruptive redesigns. Berti said those requirements will keep evolving as workloads move from AI training toward inference and as agent AI takes over more infrastructure demand.
Photonics and the longer horizon
Berti pointed to emerging photonic approaches such as IOWN as a longer-term path toward lower latency and better energy efficiency across large-scale AI environments. The direction of travel is clear enough: optical systems are moving deeper into the facility, not just between buildings.