AI Data Centers Need a Network Plan, Not Just a Campus Plan
A campus can have land, electricity, cooling, and servers. Without resilient paths to the rest of the digital world, it still is not an AI-ready system.
LegacyGrid AI Editorial · August 11, 2026 · 8 min read
Original LegacyGrid AI editorial illustration — a conceptual view of network dependencies, not a company cable map
The AI infrastructure conversation usually starts with electricity. That makes sense: no reliable power, no useful compute. But another dependency is moving quietly from the back of the plan to the front: the network that lets a campus exchange data with users, cloud regions, research partners, and other compute locations.
A building full of accelerators is not valuable simply because it can process information. It must be reachable. It must have enough throughput to move training data and model outputs. It must have a credible recovery path when a fiber cut, carrier outage, landing-station event, construction accident, or regional disruption affects one route. Those requirements make network architecture an infrastructure decision—not an IT afterthought.
From “Fiber Available” to Network Readiness
In site-selection conversations, “fiber available” is often treated like a binary answer. A provider has a line nearby, so the box is checked. That is not enough for a high-value AI facility. The useful questions are more specific: How many physically independent routes can reach the site? Which carriers control them? Do they share conduit, bridges, central offices, or substations? Is the capacity lit today, contractable in a reasonable delivery window, or merely possible after future construction?
Those questions matter because outages are rarely caused by a single abstract “internet failure.” They arise from physical dependencies. A backhoe can cut a conduit. A storm can affect several segments with a common landing point. A carrier can have capacity but not a path that avoids the same corridor as its competitor. A campus that has two logos on a network diagram may still have only one real path.
“Network diversity is not a count of vendor contracts. It is an engineering question about whether a disruption can take both paths down at the same time.”
The Four Layers Leaders Should See
A practical network plan separates four layers. First is the campus layer: duct banks, meet-me rooms, internal fiber, switching, and enough space to expand. Second is the metro layer: how the site reaches carrier hotels, exchange points, and regional cloud on-ramps. Third is the long-haul layer: the paths that connect the region to other U.S. markets. Fourth is the global layer: subsea and international routes that shape cross-border traffic, cloud reach, and application resilience.
Most institutions will not build every layer themselves. That is exactly why they need to understand the dependency chain. A university, city, or landowner should know where its risk is being carried by a utility, a carrier, a developer, an operator, or a third-party network consortium. If no one can explain the dependency chain in plain language, the project is not ready for a confident public claim of connectivity resilience.
AI capacity is useful only when its connections are planned as deliberately as its buildings and electrical systems.
Why Project Waterworth Matters Beyond One Company
Meta's announced Project Waterworth is a useful reference point because it makes an invisible dependency visible. The company has described a multi-year subsea program intended to span more than 50,000 kilometers across five major continents, using a 24-fiber-pair architecture. Separately, Meta and Corning announced an agreement of up to $6 billion for optical fiber, cable, and connectivity solutions for advanced U.S. data centers.
The lesson is not that every campus needs a private subsea system. It is that serious AI builders are treating transport capacity, cable manufacturing, and route resilience as strategic inputs. Institutions considering smaller or regional projects should translate that idea to their own scale: carrier diversity, conduit access, network handoffs, equipment lead times, fiber delivery risk, and the contractual rights that determine who can use capacity when the campus grows.
Five Questions Before a Site Is Called AI-Ready
The New Standard Is a Systems Conversation
Network readiness does not replace power diligence, cooling strategy, or community review. It joins them. A sophisticated AI infrastructure decision considers whether the site has all of its critical systems advancing together: energy, water, land, fiber, equipment, workforce, permits, and governance. A failure in any one of those systems can delay the value of every other investment.
The institutions that treat connectivity as a strategic asset will negotiate stronger agreements and avoid more avoidable surprises. They will ask for actual route maps, defined diversity standards, restoration terms, and transparent expansion assumptions. That is not overengineering. It is what it means to build compute that can be used when it matters.