Vladislav Zaimov has spent years navigating the complex intersections of enterprise telecommunications and the high-stakes world of risk management for vulnerable networks. As the industry pivots toward an era dominated by artificial intelligence, he offers a unique perspective on the physical and logical shifts required to keep global communications from buckling under the weight of massive GPU clusters. With AI models growing increasingly complex, Vladislav understands that the bottleneck is no longer just the compute power within a single room, but the optical “highways” that connect these brains across entire regions.
Our discussion explores the emergence of distributed AI training environments that stretch across campuses and cities, effectively turning the network into a seamless extension of the data center fabric. We examine the specific bottlenecks forming within metro data center interconnects and the stark reality that current optical advancements, while impressive, may still fall short of the astronomical bandwidth demands. From the transition to multimodal AI to the deployment of hundreds of fiber pairs operating in parallel, this conversation highlights why traditional scaling methods are being pushed to their absolute breaking point.
The industry is seeing a major shift toward “scale-across” architectures for AI training. How is this changing the fundamental way we design networks to support these massive GPU clusters?
The shift to scale-across architectures is born out of a very practical problem: we simply cannot pack enough power and cooling into a single building to handle the energy-hungry GPUs required for modern AI. Because of these stifling power density constraints, organizations are forced to distribute their GPU clusters across multiple data centers, campuses, and even entire regions. This means the network is no longer just a peripheral service; it has become a critical extension of the data center fabric itself. To make this work, one AI model might be trained across hardware located hundreds of miles apart, which requires a single logical environment built on top of physically separated infrastructure. This architectural transition demands a level of reliability and low-latency connectivity that we’ve never seen before, as any hiccup in the network can bring the entire training process to a grinding halt.
With AI workloads expanding beyond individual facilities, where is the most immediate pressure being felt in the current network infrastructure?
While AI is impacting every corner of the grid, the most immediate and intense pressure is building up in metro networks, specifically within Data Center Interconnect (DCI) environments and multi-cloud access points. Initially, the demand was neatly tucked away within specialized AI clusters in large data center campuses, but as training models expanded across distributed sites, that pressure began to bleed into metro and long-haul routes. We are seeing hyperscalers scramble to deploy routes with up to hundreds of fiber pairs operating in parallel to manage the load. The goal is to deliver tens of petabits per second, a staggering amount of data that turns traditional metro planning on its head. There is a real sense of urgency here because these multi-million dollar AI facilities only start generating value for the company once that optical connectivity is fully operational and stable.
We’ve transitioned from text-based AI to multimodal applications that handle video and images. How has this specific evolution changed the bandwidth math for network operators?
The move to multimodal AI has been a complete game-changer for network traffic dynamics, moving us far beyond the relatively light load of text-based queries. When an AI is processing and generating high-resolution images, video, and complex audio in real-time, the bandwidth requirements don’t just grow; they explode. We are looking at an order of magnitude more capacity than what was required for traditional cloud applications or even early-stage AI. Operators are now seeing routes where traffic increases by 10 to 100 times or more, which is an incredible jump to handle in a short timeframe. This hunger for data is what is driving the need for tens of petabits of throughput, making the “digital plumbing” of the past look like garden hoses compared to the massive industrial pipes we need today.
There is a lot of excitement around 800G and 1.6T optical technologies. To what extent can these innovations actually solve the capacity crisis created by AI?
Innovations like 800G and 1.6T coherent optical technologies are absolutely vital, but they are not a silver bullet. These technologies allow us to extract roughly 30% to 50% more capacity from the fiber we already have in the ground, which helps extend the life of existing assets and buys us some breathing room. However, when you look at the raw numbers, there is a clear mismatch: we are gaining 50% more capacity while AI is driving 10 to 100 times more traffic on certain routes. As experts at Ciena have pointed out, while these upgrades are slowing the need for new fiber, they don’t come close to meeting the total connectivity requirements of the near future. We are reaching a point where optical tricks and better transponders aren’t enough; we need to rethink the architecture entirely or commit to a massive physical expansion of fiber footprints.
As we look toward future-proofing these networks, what are the most promising technologies on the horizon that could help bridge this massive capacity gap?
We are entering a very creative phase in optical innovation where everything from the fiber itself to how we plug into the switch is being redesigned. Hyperscalers are looking very closely at hyper-rail photonic line systems and 800Gb/s C&L-band coherent pluggables to maximize fiber density and efficiency. There is also a lot of buzz around co-packaged optics and even hollow-core fiber, which could potentially offer lower latency and higher speeds than traditional glass. These technologies, alongside full-spectrum transponders, are designed to squeeze every possible bit of data out of the spectrum. The ultimate goal is to create a network that is as fluid and responsive as the AI models it supports, ensuring that the physical layer never becomes the bottleneck for human innovation.
What is your forecast for the future of optical networking in the age of AI?
I believe we are about to see a massive wave of fiber expansion that will dwarf the build-outs of the last decade. While we will continue to innovate with coherent optics and photonic systems to get that 30% to 50% efficiency boost, the sheer volume of AI traffic—growing at that 100x pace—will eventually demand more “glass in the ground.” We will see a shift toward more automated, self-healing photonic layers that can dynamically reroute petabits of data to avoid congestion. Ultimately, the network will stop being seen as a separate utility and will be integrated directly into the compute stack, where the line between a data center’s internal bus and a metro fiber link becomes almost indistinguishable. If we don’t scale the physical layer at the same speed we scale the GPU clusters, we risk hitting a hard ceiling on what AI can actually achieve for society.
