Next-generation AI clusters are being planned around up to 1 GW of power and 800 Gbps+ optical links between facilities. That changes the site-selection math immediately. Cheap land and a cooperative county office are nice. If the campus cannot move model traffic fast enough to make separated buildings behave like one machine, it is a real-estate project pretending to be infrastructure.
AI Campus Growth Depends on Interconnect Design
Data Center Knowledge’s framing is the useful correction to the usual AI buildout chatter. The market keeps treating expansion as a hunt for megawatts and acreage, when the newer requirement is “scale across” architecture: multiple buildings or sites tied together tightly enough to function as one cluster.
That means DCI stops being a support layer and becomes part of the product. A modern data center site may clear the power hurdle and still fail the workload if latency, jitter, route diversity, and optical upgrade paths were treated as somebody else’s problem.
Remote Markets Need More Than Cheap Land
Power scarcity in primary hubs is pushing AI campuses into secondary metros and rural corridors because utility timelines in major U.S. markets can stretch 2 to 5 years. That sounds like a win for cheaper land until you remember that remote markets often lack dense metro fiber, diverse long-haul routes, and easy access to cloud on-ramps.
So the industry gets a second bottleneck right after the first one. A site with available power but weak interconnect is not strategic capacity. It is stranded compute with better press photos.
A Powered Site Can Still Be the Wrong Site
This is where the internet infrastructure piece gets practical. Distributed AI campuses need disciplined routing policy, clean traffic engineering, and failure domains that do not collapse when one path gets cut. The conversation belongs as much in network connectivity planning and BGP security as it does in utility negotiations.
There is also an address-management angle, but not the cartoon version where every story becomes an IPv4 sermon. Operators still need sane public addressing, geolocation hygiene, and IPAM discipline for external services, management planes, and routed edges. The bigger point is that optical transport and routing architecture now decide whether the expensive GPUs on both ends can act coordinated instead of merely adjacent.
The Numbers Already Point to Fiber
The transport market is not being subtle about this. Zayo says another 200 million fiber-network miles will be needed by 2030 to avoid a bandwidth shortage, and metro dark fiber purchases jumped 268% from 2023 to 2024. That is not decorative demand. That is operators buying the physical path before the AI load arrives in full.
The rest of the stack says the same thing in different language. Omdia sees global IT load power capacity reaching 314 GW by 2030, Ciena says DCI bandwidth demand is expected to rise at least 6x over the next 5 years, and TeleGeography reports international internet bandwidth rose 23% in 2025. The market is building more power, more optics, and more backbone at the same time because AI capacity only counts when those layers show up together.
FAQ
Why is AI data-center expansion becoming a fiber and DCI story?
Because large AI clusters are increasingly distributed across multiple facilities, and those sites need high-capacity, low-latency links to operate like one system rather than isolated buildings.
Is power still the main constraint for AI campuses?
Power is still a major constraint, but it is no longer the whole story. A campus can secure megawatts and still underperform if its metro and long-haul interconnect design is weak.
What does “scale across” mean for AI infrastructure?
It means operators are connecting multiple facilities so one training or inference environment can span them. That raises the importance of DCI, optical upgrades, route diversity, and latency control.
How does this affect routing and internet infrastructure operators?
It pushes routing policy, path diversity, and backbone engineering closer to the center of AI planning. Fiber builders, carriers, and operators with resilient interconnect are no longer downstream vendors. They are part of the core infrastructure equation.
