The Data Center Boom Is About to Change Its Address

For twenty years, "data center" mostly meant a nondescript building on the edge of a city — close to fiber backbones, close to customers, close to the people who'd need to drive out and fix something. That geography made sense when the bottleneck was latency to users. It's starting to make a lot less sense now that the bottleneck is electricity. The numbers on this shift are already stark. Roughly 87% of existing data centers sit in urban areas. But look at what's actually being planned rather than what already exists, and the picture flips: about two-thirds of new data center capacity is heading somewhere else entirely — rural counties, transmission corridors, land next to power plants. This isn't a niche trend. It's becoming the default mode of expansion for the industry. Why location logic is inverting The reason is simple once you see it: grid interconnection queues now stretch up to eight years in some regions, and wholesale power prices near existing hyperscale clusters have reportedly surged over 260% as demand collides with limited supply. When you're trying to stand up a facility that needs several hundred megawatts — sometimes a full gigawatt — waiting eight years for a grid connection isn't a rounding error, it's a dealbreaker. So the industry is doing the obvious thing: going to where the power already is, instead of waiting for power to come to where the industry already is. New rural sites cluster tightly around high-voltage transmission lines — the median distance is around four miles — because that's the fastest path to megawatts. And increasingly, "where the power is" means next to a nuclear plant, or backed by a long-term power purchase agreement with one. Meta's January 2026 deal to secure up to 6.6 gigawatts of nuclear power by 2035 — combining a 20-year agreement with an existing plant operator alongside forward commitments to small modular reactor developers — is the clearest signal yet that this isn't experimental. Microsoft and Amazon have made similar moves, in some cases backing restarts of previously shuttered reactors. The industry has a phrase for the logic behind all of this, and it's a good one: it's easier to move gigabits than it is to move electrons. Data can travel at the speed of light through a fiber-optic cable. Electricity, by contrast, is expensive and slow to deliver at scale to a new location — you need transmission buildout, permitting, interconnection studies, years of lead time. So instead of bringing power to the data center, the industry is bringing the data center to the power, and asking the network to close the resulting distance. The part that's easy to miss Here's where it gets interesting, and where I think the obvious version of this story undersells what's actually happening. It would be easy to read all of this as simply "data centers are moving to rural areas" and stop there. But the more consequential shift is what happens to the connections between facilities once they're spread out like this. AI training at the frontier has outgrown what a single building — even a very large one — can hold. OpenAI and Microsoft are reportedly working toward interconnecting multiple gigawatt-scale campuses into a single distributed training system, with individual sites potentially hundreds or even thousands of kilometers apart. That only works if the network connecting those sites can move enormous volumes of data with very low, very consistent latency — because if data can't move between locations fast enough, the expensive compute at each site sits idle waiting for it, and idle GPUs are about the most expensive thing you can own right now. This is a genuinely different infrastructure problem than the one the industry solved for the last decade. Inside a single data center, you're moving data over relatively short optical links. Once you're stitching together campuses across a state or a country, you need long-haul coherent optical transport — the same category of technology that undersea cables and continental backbones use, except now it's being pressed into service for something much more latency-sensitive than routing web traffic. Who actually builds that This is the adjacent-but-not-obvious opportunity. The companies everyone talks about when they talk about the AI infrastructure buildout are the chipmakers and the switch vendors — Nvidia, Cisco, Arista, Broadcom. That's the layer closest to the GPU, and it's the layer with the most attention on it already, which means it's also the layer where a lot of the obvious upside is already reflected in how those companies are valued. The optical transport layer is quieter. It's less discussed, harder to picture, and it doesn't show up in a keynote demo the way a new chip does. But if the trend toward power-driven, geographically dispersed data center siting continues — and the economics described above suggest it will, because the alternative is an eight-year wait for grid capacity — then the amount of long-haul optical bandwidth the industry needs doesn't grow linearly with compute demand. It grows faster, because every mile of separation between campuses is a mile that has to be crossed by fiber, not by concrete. Companies with real capability in coherent optical transport — the technology that makes high-bandwidth, low-latency, long-distance data movement possible — sit in a small club. It includes some familiar telecom-equipment names that have spent the last few years quietly retooling for exactly this use case, largely through acquisitions that gave them the optical scale they didn't have organically. It's a smaller, more specialized market than switching, which means less competition crowding in, but also means it's easy for the broader market to underweight relative to flashier parts of the AI buildout story. Still theoretical, worth watching None of this is a sure thing yet. Distributed, multi-campus AI training at gigawatt scale is still more roadmap than reality — the technical challenges of keeping training runs synchronized across sites separated by hundreds of miles are real, and there's a version of the future where the industry solves the power problem some other way (small modular reactors sited directly at existing campuses, for instance) that reduces the pressure toward geographic dispersion rather than accelerating it. But the direction of travel is hard to ignore: power is now the binding constraint on AI infrastructure, not land or chips, and the industry's answer to a binding power constraint is to spread out and lean harder on the network to hold it all together. That's not a story about any single company. It's a story about a layer of infrastructure that's about to matter a lot more than it used to, well before most of the market is pricing it that way.