Power Drives the AI Data Center Boom, but Connectivity Cannot be Overlooked

extract

Insight and analysis on the data center space from industry thought leaders. Power Drives the AI Data Center Boom, but Connectivity Cannot be Overlooked AI’s power demands are reshaping data centers, but without robust optical connectivity, these mega facilities risk becoming isolated islands, writes Sterling Perrin. Data centers are at the epicenter of the AI super cycle. The infrastructure required for training and running AI models has fueled a historic demand for power, cooling, floor space, high-speed networking, and specialized compute. At data center-focused trade shows, such as Metro Connect in Florida, power, cooling, and real estate (including financing) are front and center at the conference and in the corridors. Omdia forecasts that IT load power capacity will reach 314 GW globally in 2030 – a 2.9x increase in power compared today and driven predominantly by AI. Without reliable power, there can be no AI, so building and delivering power to these mega data centers is crucial. Networking and specialized compute are also essential components in the AI revolution. To date, these topics are discussed and debated in different forums, such as at Nvidia’s GTC or at OFC. With this separation, it is easy to miss how these essential components are all inter-related and often inter-dependent. Increasingly, data center operators will need to understand how networking and connectivity technologies will help – and hinder – their AI plans. Scale Across: A New Paradigm for AI Connectivity One example is the emerging use case called “scale across.” Coined in 2025 by Nvidia’s CEO Jensen Huang, scale across describes high-bandwidth connectivity between data centers specifically for AI backend workloads, including training as well as inference. This new DCI variant was a key theme at this year’s OFC as well as at the annual Optica Executive Forum held during that same week in March. Why is scale across needed? It goes back to the enormous power demands of the AI GPUs. Power requirements are becoming too large to be contained within a single facility. For example, while a large co-location data center may support 50-100 megawatts of IT load, next generation hyperscaler training clusters are being built and planned for a gigawatt of power, and even higher in some cases. These will be distributed, multi-facility campuses, sometimes spanning multiple electrical grids by necessity. But to operate as a single cluster, high capacity (800 Gbps+), low latency, low jitter, and ultra-high availability optical connectivity between buildings will be required. This is what scale across aims to deliver. Power Demands Are Reshaping Data Center Infrastructure In addition to scale across, access to power is driving a second trend that impacts connectivity: remote builds. In the US, historical hyperscaler and large colocation data center hubs are in places including Northern Virginia, Dallas/Ft. Worth, Northern California, and the Pacific Northwest. But additional power in these regions is in short supply, making it costly and with wait times stretching from two to five years. As a result, the industry is witnessing an explosion of massive data center projects in new locations – often rural – that have large tracts of land and timely access to reliable power. These include places like Lancaster, Pennsylvania, Memphis, Columbus, Ohia, rural Georgia, New Mexico, and many others. All these locations require massive fiber-optic capacity to hub locations – otherwise the mega data centers will be stranded islands. Significantly, because these locations are chosen primarily for their access to power, fiber capacity rarely exists and distances to major metros run hundreds of kilometers, and sometimes up to 1,000km. There is no question that the power demands of AI are the primary driver for data center builds today – whether it is building multi-site AI clusters for training or locating a gigawatt campus in remote Wyoming. These trends will surely continue, but as they do, fiber-optic connectivity must increasingly become a part of the planning equation. Earlier, I wrote that without power there can be no AI. But this statement needs to be amended: without power and optical connectivity, there can be no AI.