EDGE INFRASTRUCTURE
Edge inference sites: interconnection before distance
A nearby building can still have a long network path to its users. Site diligence needs to establish who you can reach, how you reach them and what happens when a connection fails.
Start with the users
A map makes an edge data center look straightforward: put compute close to demand. The network introduces a less convenient question. Close to whom, through which provider, and over which route?
An inference service reaches people through their access networks. If those networks exchange traffic with the proposed site through a distant hub, the physical distance between the user and the building can be misleading. At Digital Midstream, my interest is in locations that can support useful connections to the networks serving the intended users. Our 1 MW site under contract in Macon, Georgia, gives this work a local context. The same diligence applies there: establish which networks the site can reach, verify usable power and identify the work required before operation.
Ask what the fiber actually reaches
“Fiber nearby” is an opening question. It does not tell you whether a carrier can deliver service into the building, what construction is required, which facilities the route reaches or what capacity can be ordered. A line on a map is not a service commitment.
Ask for the delivery point, route, construction scope, lead time and commercial terms. Establish whether exchange access is local or carried over a transport service. An Internet exchange can provide a place to connect with other networks, but membership alone does not guarantee that a particular provider will peer with you. Confirm the arrangements that the service actually needs.
Test a market, not just an address
Suppose a building is twenty miles from a group of users, but their provider carries traffic to a hub several hundred miles away before exchanging it. A more distant facility with a direct connection to that provider could produce a shorter round trip. This is a hypothetical example; the result depends on the actual routing in both directions.
Test from representative access networks at different times of day. Measure round-trip delay, loss and variation under load. Use traceroute as supporting evidence rather than a complete map: return paths can differ, and some routers rate-limit diagnostic replies. Where a site has no service yet, measurements to a nearby facility are provisional. They do not replace acceptance testing after delivery.
Check the failure path
Two carrier names do not prove physical diversity. Services can share a duct, building entrance, transport segment or upstream facility. Ask where the paths separate and what remains shared.
Then consider the surviving connection. Can it carry the intended traffic if the other one fails? Does the alternate route still reach the important access networks without a large detour? The failure case belongs in the initial site comparison, because fixing it later may require construction or a different transport contract.
Treat the building as another engineering system
An older telecom facility may offer valuable electrical infrastructure and established fiber routes. It still needs an assessment for the proposed workload. Utility service capacity is not the same as usable IT capacity: cooling, electrical losses, redundancy and other building loads all consume part of the supply.
Verify utility availability, distribution equipment, cooling design, floor loading, fire protection and the permissions needed for the intended use. Date the evidence. Separate what exists from what is proposed, and distinguish an available service from capacity that has been reserved. A credible conversion budget includes the work needed to bridge those differences.
Make the decision auditable
For each candidate, I want a clear account of the users it could serve, the networks it could connect to, the power it could support and the work required before operation. Unknowns should have an owner and a next step.
Interconnection deserves attention early because it can change whether a location makes sense for the workload at all. A promising property becomes a useful inference site only when its network, power and building constraints work together.