While digital innovation is typically conceived as an enabler, it can just as easily become a source of complexity. Too often it creates new barriers – not only for those behind the curve, but also for the engineers and technicians responsible for implementing and troubleshooting the technologies on which it depends..
The compulsion to build ever more powerful AI infrastructure has made GPUs the headline of almost every conversation concerning the data center, with attention focused on the next generation of compute and processing power. But AI performance is just as much a connectivity story..
The GPU is the star of the AI data center narrative, and rightfully so. But a GPU in isolation accomplishes nothing Ed Gastle, Viavi. “The GPU is the star of the AI data center narrative, and rightfully so.
But a GPU in isolation accomplishes nothing,” says Ed Gastle, global product line lead at Viavi.. A new era of fiber. If you were working in telecommunications in the 80s, chances are your first commercial fiber optic deployments centered on single-channel, point-to-point trunk links.
By the early 90s, the industry had largely transitioned to duplex fiber, back then considered state-of-the-art for high-speed enterprise networking.. That picture has continued to evolve, and today’s hyperscale AI data centers are built around multifiber connectivity and optical arrays, designed to manage the extraordinary volumes of data required by modern compute clusters..
AI has fundamentally reshaped high-speed networking, overturning decades of established practice at a pace the industry has never experienced before.. In an AI data center, constant cross-communication is required to update thousands of nodes during model training, coordinate inference tasks across distributed clusters, and synchronize in real time.
A delay of just a few microseconds between GPU nodes can be enough to stall an entire training workload, turning a seemingly insignificant network issue into a costly bottleneck.. Multifiber at hyperscale. Modern AI infrastructure scales along three distinct dimensions – scale up, scale out, and scale across – each placing its own demands on the physical network.
Whether adding more GPUs within a server, expanding distributed compute clusters, or connecting multiple data centers into a single AI fabric, every step increases the complexity of the underlying fiber infrastructure.. Hyperscalers may do all three at scales of tens of megawatts per building, millions of fiber connections, and sub-millisecond latencies across global clusters..
Supporting these workloads means moving from 400G and 800G toward 1.6T and beyond by aggregating multiple optical lanes into a single link. Every one of those lanes depends on pristine fiber performance.. At this scale, engineers are also looking to next-generation optical media, such as hollow core fiber (HCF), which is rapidly moving from field trials into broad operational deployments due to its ability to offer unmatched speed and performance, as Gastle explains:.
“This is precisely because of its ability to lower latency by approximately 30 percent, reduce chromatic dispersion by around 70 percent, and deliver attenuation below 0.1 dB/km in specific wavelength bands – properties that are profoundly relevant to cloud providers scaling across long distances.”. The fiber infrastructure carrying all of that traffic is not peripheral to AI performance – it is the precondition for it Ed Gastle, Viavi.
Gastle has witnessed the transformation of physical media throughout his career. From his early days as a technician to leading product strategy and technician training programs, he has seen the resilience of the underlying fiber network grow increasingly critical.. He points to data that shows this is far more than a theoretical concern.
A 2023 analysis of AI training tasks found that fewer than 60 percent completed successfully on the first attempt, with nearly half of network-related failures originating at or below the transport layer. The GPUs were functioning as intended; the fiber was not.. “The fiber infrastructure carrying all of that traffic is not peripheral to AI performance – it is the precondition for it,” says Gastle..
The only way to ensure that the network behaves as intended is to perform comprehensive and rigorous testing. And as network complexity increases, testing and validation methodologies must evolve alongside it.. Challenges of increasingly dense fiber environments.
The scale of AI data center deployments is unprecedented, but the operational burden of validating them is growing even faster. Testing does not scale linearly with fiber count – it multiplies with every additional lane, connector, and optical path.. Thus, the playbook that served previous generations of networking no longer applies, forcing technicians to solve problems that simply did not exist a decade ago..
Multifiber testing is more challenging because of the sheer scale. MPO and MMC connectors – now the connective tissue of modern AI data centers – carry 12, 16, or even 24 fibers in a single interface. Each connector exposes many end-faces at once, exponentially increasing the risk of cross-core contamination, incorrect pin and polarity alignments, and installation damage..
Connector cleanliness has always been fundamental, but multifiber environments change the scale of exposure and the mathematics of probability. As fiber counts per connector increase, the probability that every end face is simultaneously clean drops dramatically. A single contaminated connector on one multifiber link can create enough insertion loss to compromise an entire 1.6T link, introducing intermittent errors that trigger timeout failures cascading across the cluster..
The only reliable approach is to inspect before you connect – every time, across every fiber in the connector Ed Gastle, Viavi. “Cleaning alone does not guarantee success. In some cases, it can redistribute contaminants rather than remove them,” explains Gastle, adding:.
“The only reliable approach is to inspect before you connect – every time, across every fiber in the connector – using purpose-built multifiber inspection tools that deliver panoramic imaging, automated pass/fail analysis, and results aligned to industry standards.”. This sequence matters.
Once a contaminated connector is mated, the opportunity to verify its original condition is lost. Contamination also risks causing physical damage to alignment pins or fiber end-faces – damage that cannot be undone by cleaning. Inspection before connection is therefore the only point in the workflow where defects can be identified and corrected before they become embedded in the network..
Cleanliness, however, is only one part of the challenge. Unlike duplex fiber, which relies on alignment sleeves, MPO and MMC interfaces use physical alignment pins. Installers must therefore verify that both pinned and unpinned connectors are correctly matched and that polarity is consistent across every link in the chain..
The risk of getting this wrong ranges from a failed test – which is recoverable – to a misleading pass result – which is not.. “If polarity is mismatched between the reference and test setup, the tester may report a passing result for a link that is not actually compliant.
That becomes a latent fault – one that sits in the infrastructure undetected until a high-load scenario reveals it in the worst possible way,” explains Gastle.. As networks become denser, engineers must also contend with physical layer challenges that were far less significant in previous generations.
Higher fiber densities increase susceptibility to crosstalk, while tightly packed connectors are more vulnerable to contamination and installation damage. These impairments can lead to intermittent faults that may only appear under high load as devices come under thermal stress.. What makes them so dangerous is their intermittent nature – they do not announce themselves under light load, and by the time they surface at full utilization, they are difficult and expensive to isolate Ed Gastle, Viavi.
“What makes them so dangerous is their intermittent nature – they do not announce themselves under light load, and by the time they surface at full utilization, they are difficult and expensive to isolate,” says Gastle.. The transition to 1.6T also marks the point at which traditional throughput tests are no longer sufficient to guarantee a network can handle synchronized AI compute efficiently, reliably, and at scale.
Instead, engineers must validate deterministic network behavior across thousands of synchronized endpoints.. Testing at scale as an operational discipline. Hyperscale AI data center projects now routinely involve multifiber link quantities from 30,000 to over 100,000.
Every single one of those links must be inspected, certified, documented, and reported before the infrastructure can be commissioned.. One of the most persistent misconceptions is that testing is a final checkpoint – a task to complete just before the network goes live. At hyperscale, this is backwards.
By the time you discover a systematic problem at the physical layer after tens of thousands of multifiber links have been installed, the rework cost is enormous, and the deployment timeline is already in jeopardy.. A second misconception is that scaling is simply a matter of having more of the same tools.
However, Gastle explains:. “Multifiber testing is not duplex testing multiplied. The workflows, reference setups, inspection protocols, data management, and reporting requirements are fundamentally different.”.
Teams approaching MPO and MMC deployments with traditional fiber practices quickly encounter the limits of manual processes. Without purpose-built tooling and process automation, the scale and rate of opportunities for human error increase massively.. Deployments at hyperscale level require automation.
Manual visual inspection and disconnected testing workflows cannot deliver the consistency or throughput required to certify AI infrastructure at the speed, scale or level of granularity hyperscale environments require. Instead, testing needs to be embedded throughout the deployment lifecycle, with validation performed continuously as the network is built, not deferred until the final audit..
“At hyperscale, where project timelines are tight and rework is expensive, there is always pressure to move faster. But the cost of discovering a contamination-driven failure after a link is certified and active – requiring deactivation, re-testing, re-inspection, and documentation updates – is far higher than the time cost of systematic pre-connection inspection,” says Gastle..
“The discipline of “Inspect before you connect” pays for itself many times over at the scale of a modern AI data center build.”. Three priorities for hyperscale validation. According to Gastle, staying ahead requires organizations to focus on three priorities..
1. Invest in purpose-built test tools. Testing multifiber infrastructure requires tools designed specifically for high-density optical environments, rather than adaptations of legacy duplex platforms..
Automated inspection systems, like the Viavi INX 700, capture panoramic imaging of every fiber within an MPO or MMC connector, applying automated pass/fail analysis against industry standards and removing the subjectivity of manual inspection.. Executing repeated test procedures, especially with large teams working simultaneously on the same deployment, is one of the most common causes of inaccurate certification results.
Platforms like the Viavi DCX 700 guide technicians through reference conditions and repeatable test procedures, helping standardize testing across large deployment teams.. 2. Automate the workflow, not just the measurement.
Beyond the measurements themselves, hyperscale testing shifts the challenge from collecting data to managing it.. A deployment containing tens to hundreds of thousands of multifiber links generates an enormous volume of inspection images, certification results, and compliance records.
Every dataset must be associated with the correct location, connector, and link, while remaining searchable, traceable, and available for producing compliant closeout reports.. Automation therefore needs to encompass the entire workflow, from field execution to reporting. Platforms such as Viavi Test Process Automation (TPA) integrate test instruments, technician workflows, and project documentation into a closed-loop system, ensuring every result is captured, attributed, and traceable across the entire project lifecycle..
3. Testing must be validated in its own right. Certification should not end when construction does.
AI data centers are living environments, continuously evolving as GPU clusters expand, workloads shift, and capacity is added. The physical infrastructure changes with them, meaning certification records must also evolve to support incremental updates and re-certification, not just point-in-time records..
Accountability is shared, but the obligation to establish testable, verifiable standards belongs to the testing industry, and we take that responsibility seriously Ed Gastle, Viavi. When a physical media fault eventually surfaces, the questions are always the same: was it detectable?
Was the appropriate test performed? Was the result documented?. If the answer to any of those is no, diagnosing the fault becomes significantly more difficult, and accountability becomes far less clear..
“What Viavi believes is that these questions should drive the industry toward higher standards of testing documentation and traceability – not just during construction, but throughout the operational lifecycle,” says Gastle. He adds:. “Accountability is shared, but the obligation to establish testable, verifiable standards belongs to the testing industry, and we take that responsibility seriously.”.
Future preparation. The physical media is where AI data center performance is ultimately won or lost. As AI infrastructure continues to grow in scale, density, and speed, validating the fiber plant has become a strategic engineering discipline – not just a routine commissioning task..
The industry cannot afford for testing capabilities to lag behind network innovation. While many operators are only beginning to deploy 1.6T infrastructure, Viavi is already preparing for the next generation of 3.2T networking, where new optical architectures, higher lane densities, and increasingly complex interconnects will raise the bar for system validation once again..
“The preparation required at each step – from 800G to 1.6T, and from 1.6T toward 3.2T – is not a one-time certification activity. It requires continuous investment in testing capability. And it must begin at Layer 0 – the physical media – before any higher-layer validation can be trusted.”.
The fiber story in AI data centers is not a footnote. It deserves to be a headline Ed Gastle, Viavi. Gastle emphasizes that testing must be deliberate, systematic, and continuous.
That means adopting “Inspect before you connect” as an operational discipline, investing in tools and automated workflows that are purpose-built for multifiber and emerging fiber types, and treating test data with the same level of governance applied to the compute.. The GPUs will keep getting faster.
The AI will keep getting smarter. The networks will keep scaling. But none of it matters if the fiber is not right.
Gastle concludes:. “The fiber story in AI data centers is not a footnote. It deserves to be a headline.”.
For more information, please visit Data Center Interconnect (DCI) | VIAVI Solutions Inc.. More from Viavi. 10 Oct 2025
