Table of Contents
Rather than treating AI as an extension of conventional cloud computing, AWS plans and operates AI infrastructure independently from its traditional cloud network
In sum – what to know:
Traffic patterns – AI workloads generate synchronized, latency-sensitive traffic that AWS treats separately from traditional cloud workloads, requiring dedicated network fabrics and independent capacity planning.
Deployment timelines – AWS says the biggest challenge is coordinating compute, power, and networking simultaneously, rather than overcoming shortages of any single resource.
Optical scaling – The company says 800G optics, coherent networking, and larger DCI fabrics are helping support AI growth, but new fiber deployment remains essential.
Artificial intelligence is fundamentally changing how hyperscale networks are scaled and operated, according to Stephen Callaghan, senior principal technologist, network-core at AWS, who told RCRTech that AI infrastructure requires a different networking model from traditional cloud environments because of its highly synchronized traffic patterns and unprecedented bandwidth requirements.
Rather than treating AI as an extension of conventional cloud computing, AWS plans and operates AI infrastructure independently from its traditional cloud network. According to Callaghan, AI training and inference workloads generate highly correlated traffic across large accelerator clusters, creating demands that conventional cloud architectures were never designed to support.
“Traditional cloud traffic is distributed and bursty. It’s made up of millions of independent customers, each generating modest, uncorrelated demand. AI workloads are the opposite: a single AI training cluster could represent a single customer, involve over 100,000 accelerators communicating simultaneously, generating petabits of tightly coupled traffic across a handful of datacenters,” the executive said.
To prevent AI demand from affecting traditional cloud services, AWS operates separate network fabrics and capacity planning processes for each environment.
The company is seeing pressure across metro, long-haul, and subsea networks, although Callaghan noted that each presents different challenges. Metro capacity is being consumed far faster than historical planning assumptions, while subsea expansion remains constrained by the physical time required to deploy new cable systems.
According to AWS, AI is also driving rapid expansion of data center interconnect (DCI) infrastructure. The company continues to scale both the number of links and the bandwidth of each connection as AI clusters grow larger and require predictable communications between facilities.
“We’re scaling both the number and per-link bandwidth of our DCI connections, while defining explicit bandwidth specifications at the cluster level so customers have predictable, guaranteed capacity for their inter-datacenter flows,” he added.
Callaghan said advances in optical networking are helping support this growth, with AWS deploying 800G optics across its data center fabrics while continuing to rely on coherent optical technologies for metro and long-haul transport. However, he stressed that higher-speed optics alone will not eliminate the need for additional fiber infrastructure.
“New optical technologies absolutely help, but they can’t keep up on their own… Upgrading existing fiber to higher speeds and building new fiber routes are necessary. It’s not one or the other,” Callaghan added.
According to the executive, the industry’s biggest constraint is no longer the availability of fiber or power themselves, but the time required to bring every element of infrastructure online together. AWS has responded by integrating network, power, and facility planning from the earliest stages of deployment.
Callaghan said this approach reduced network build time at one machine learning site by 48%, cutting deployment from 16 weeks to just over eight weeks.
AWS also believes its vertically integrated networking strategy has become increasingly important as AI infrastructure scales. The company owns and operates every layer of its networking stack, from fiber infrastructure and optical components to hardware, firmware, operating systems, and management software.
“The network should never be the reason a GPU sits idle… we own every layer of our network: from fiber routes and physical connectors through optical transceivers, network hardware, firmware, operating system, and observability stack.”
Looking ahead, Callaghan argued that future AI infrastructure growth will depend less on overcoming individual resource constraints than on coordinating them effectively.
“The limiting factor is coordination, not any single resource. Compute, power, and network all have multi-year supply chains, and any one of them can gate your ability to serve customers if it falls behind… The companies that treat these as sequential will be the ones that hit bottlenecks.”
The interview with AWS’s Stephen Callaghan is part of a recent 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.
