Table of Contents
Bruno Berti, SVP of global product management at NTT Global Data Centers, told RCRTech that AI is changing not only the scale of data center infrastructure but also when and how operators need to invest in it
In sum – what to know
Upfront spending rises – AI is driving more front-loaded investment in high-capacity fabrics, optical interconnects and resilience because network capacity cannot simply be added later.
Speed-to-power matters – Power availability and delivery timelines are emerging as bigger physical constraints than fiber for AI deployments.
Systems converge – High-density AI environments increasingly require interconnects, power, and cooling to be engineered together rather than separately.
AI is changing not only the scale of data center infrastructure but also when and how operators need to invest in it, according to Bruno Berti, SVP of global product management at NTT Global Data Centers.
Berti said network capital expenditure historically scaled relatively linearly with traffic growth. AI workloads are changing that model by requiring more front-loaded investment in high-capacity fabrics, optical interconnects, and resilience. “You can’t incrementally ‘add network later’ once large AI clusters are running,” Berti said.
The reason is the sensitivity of AI workloads to latency, loss, and jitter. Berti said higher-speed optics, flatter architectures, and denser designs can allow operators to move more data over the same physical footprint, improving cost per bit even as overall spending rises.
“In many cases, spending more upfront actually lowers total cost of ownership over the life of the deployment,” he said.
Berti also described a shift toward systems-level optimization, with compute, network, power, and cooling increasingly planned together. Rather than making those investments separately, he said they are being coordinated so that they reinforce one another and avoid downstream inefficiencies.
The physical constraints of AI infrastructure reinforce that approach. Berti said the biggest physical constraint today is not fiber but power, followed closely by deployment timelines.
Fiber capacity in mature data center markets is generally manageable, according to Berti. Higher-speed optics and dense fiber builds can provide connectivity capacity, while the larger challenges are associated with power availability and delivery.
AI workloads require substantially higher power density than traditional compute, Berti said, while grid interconnections, substations, transformers, and permitting in many regions do not move quickly enough to keep pace.
Long lead times for grid upgrades, electrical equipment, and approvals can further extend projects even after land and capital have been secured. “That’s why speed-to-power has become just as important as speed-to-market,” Berti said.
Berti said the industry is responding through earlier utility engagement, onsite generation, modular designs, and strategic market selection. He said the challenge is less about technology limits than infrastructure coordination at scale.
Inside the data center, high-density GPU clusters are also changing physical design. AI training environments require continuous east-west communication between GPUs, increasing fiber density inside halls and campuses and driving flatter network layouts and greater attention to signal integrity and routing.
Power distribution is changing as well. Berti said GPU-dense racks draw many times the power of traditional compute, requiring redesigned electrical architectures, higher-capacity busways and shorter power paths to reduce losses. “Power is no longer something you scale gradually — it has to be engineered for peak AI loads from day one,” he said.
AI is also changing data center interconnect requirements between facilities. Berti said AI is driving higher sustained bandwidth, tighter latency tolerances, and greater sensitivity to consistency and jitter, particularly when AI training, inference and data pipelines are distributed across multiple facilities.
He identified three requirements: higher bandwidth expectations, greater importance for latency and predictability, and scalability and flexibility.
Over the longer term, Berti said photonic networking initiatives such as IOWN could become relevant for improving energy efficiency and latency over distance as AI systems become more distributed across campuses and regions.
The interview with NTT Global Data Centers’ Bruno Berti is part of a report published by RCR Wireless News and RCRTech, titled Scaling Optical Networks for the Hyperscale and AI Era, which can be accessed by clicking here.
