The Real Bottleneck in AI Infrastructure Is Not What You Think

Every conversation about AI infrastructure in2025 started with the same word: GPUs. Who has the most? Who can get access? When will supply meet demand? It was a reasonable framing for 2023. Back then, the scramble for high-performance compute chips was genuinely the defining constraint. Data centers could be built. Power could be secured. But the chips needed to fill them were backordered for six to nine months, and secondary-market prices were twice the list price.

In 2026, that story has changed, and mostpeople have not caught up with the change.

A March 2026 survey by Flexential of 350-plusenterprise IT leaders found that 96% had experienced at least one network-related performance issue affecting their AI workloads in the past year, and 91% said that fiber availability had limited their ability to choose their preferred AI deployment site. These are not chip problems. These areconnectivity problems.

More striking still: according to a Global Data Center Hub analysis published in April 2026, hundreds of billions of dollars in planned U.S. data centre capacity were delayed or cancelled in 2025,not because chips were unavailable, but because the fiber to connect those datacentres to the internet either did not exist nearby or could not be delivered in time.

The bottleneck is no longer inside the chip. Itis between chips. Interconnect is no longer a supporting component, it is becoming core infrastructure for AI systems. OFC 2026 Analysis, April 2026

This is the story that is not being toldloudly enough. The global AI buildout has created a fiber crisis, quiet, structural, and entirely predictable in hindsight.

The Fiber Crisis in AI Infrastructure: What theNumbers Actually Say

The scale of the fiber supply problem is best understood through a set of numbers that have received almost no mainstream coverage.

The 36X figure deserves particular attention. A conventional cloud data centre. The kind that has been built in large numbers over the past fifteen years uses a relatively modest amount of fiber internally, mostly to connect servers to storage systems and the outside network. An AI data centre is built around fundamentally different physics.

Training a large AI model requires thousands of GPUs to exchange data with each other continuously, at extraordinary speeds, for weeks or months at a time. This creates a dense web of connections inside the facility that simply does not exist in conventional computing. Each GPU chip must communicate with dozens of neighbouring chips simultaneously. The only way to carry that traffic without introducing the latency that would cripple AI performance is optical fiber.

This transition from copper-dominated to fiber-dominated intra-facility networking happened faster than the supply chain could respond. To cope with demand, manufacturers have shifted production from standard telecom-grade fiber (G.652D) to the higher-performance G.657A fiber suited to AI data centers, creating secondary shortages in conventional fiber and driving price increases across the board.

Why AI Infrastructure Needs So Much More Fiber Than Anything Before It

To understand why AI infrastructure has such an extreme fiber appetite, it helps to understand what happens inside an AI data centre at the physical level. A decade ago, a typical enterprise data centre processed tens of terabytes of daily traffic. A single AI training run for a large language model today generates petabytes of inter-chip communication every day. The difference is not incremental. It is a change of kind, not degree.

Inside the Rack: The I/O Wall

NVIDIA’s CEO Jensen Huang has described what engineers call the “I/O wall”: compute power can be doubled by stacking chips and improving manufacturing processes, but the connections between chips do not improve at the same rate. As AI clusters scale to hundreds of thousands of GPUs, the data that needs to move between those chips grows faster than the connections can carry it. The solution is optics, moving data using light rather than electricity.

At 200 gigabits per second per lane, the speed at which AI clusters now routinely operate, copper cable reaches its physical limits. Signal integrity degrades. Power consumption spikes. In many configurations, copper cannot even reliably span a single server rack at these speeds. The switch to optical fiber inside the data centre is not a preference. It is a physics requirement.

The point is that AI does not create a new fiber requirement. It creates three simultaneously, and at a scale that has overwhelmed the industry’s ability to respond quickly.

The Bottleneck Has Shifted Within AI Infrastructure: From Chips to Cables

There is a pattern in every major technology transition: the initial constraint is always the most visible component, and the secondary constraint, the one that actually determines the pace of real-world deployment, takes longer for the market to price in.

In the early Internet era, the primary constraint was bandwidth. The secondary constraint, which took years to surface, was the organizational capacity to use that bandwidth productively. In the early cloud era, the constraint was compute. The secondary constraint that slowed enterprise adoption was governance and security architecture.

In the AI infrastructure era, the initial constraint was compute, GPUs. The secondary constraint, now fully visible in 2026, is connectivity. Specifically: fiber optic cable, the skilled labor to install it, the municipal permits to route it, and the supply chains to manufacture it at the required scale.

“95% of generative AI pilot programs have failed to produce measurable financial impact, according to a 2025 study from MIT’s NANDA initiative. The failures, the researchers found, stem not from model quality but from poor infrastructure integration, systems that could not get data to and from the AI fast enough, reliably enough, or cheaply enough to justify deployment. Connectivity is not a detail. It is the foundation on which everything else stands or falls.”

The organizations that are succeeding with AI infrastructure in 2026 share a specific characteristic. They treated fiber as a first-class planning decision, not an afterthought. They committed to fiber supply agreements early, in some cases years before their compute arrived. They designed their data center floor plans around optical interconnect architecture from the start, specifying bend-insensitive fiber for dense rack environments before the racks were even ordered.

The organizations struggling are those that designed for compute first and assumed connectivity would be available when needed. In 2026, that assumption is no longer safe.

What This Means for India and the Global Buildout

The global fiber supply crisis has a geographic concentration problem. The majority of optical fiber manufacturing capacity sits in a small number of countries. When demand spikes simultaneously across every major market, as it has with AI, lead times stretch globally, and prices rise everywhere. Countries with domestic manufacturing capacity are insulated from the worst of this. Countries entirely dependent on imports are not.

India’s investment in domestic optical fiber manufacturing, built over decades, not in response to the current crisis, is proving to be a strategic asset at exactly the right moment. The combination of manufacturing capacity, an active 5G rollout creating immediate domestic demand, a BharatNet program laying fiber foundation across hundreds of thousands of village councils, and growing data centre investment in Mumbai, Chennai, Hyderabad, and Pune creates a compounding opportunity.

The AI infrastructure buildout does not end at the data centre gate. It runs all the way to the user. And in a country the size of India, where the distance between the data centre and the last user is measured in thousands of kilometers, the fiber connecting those two points is not a commodity input. It is the essential infrastructure on which the entire AI economy runs.

AI infrastructure is only as capable as its least capable layer. Right now, in 2026, that layer is fiber. Understanding that changes how you plan, where you build, and who you choose as your connectivity partner.