Hubs have often followed networks, not compute, but DE-CIX expects distributed AI inference to push more interconnection towards users, while Equinix says the heaviest traffic still runs over private fiber
In sum – what to know:
- Hubs are spreading, and inference may pull more towards users – DE-CIX is expanding in markets such as Richmond, Houston and Wichita, and expects AI to bring more distributed inference, closer to customers and connected through regional hubs.
- Private fiber carries the bulk – One estimate suggests somewhere in the region of 100 times more traffic crosses private fiber than public exchanges, with some customers starting new facilities at around 30 terabits of bandwidth and with bundles of 288 fiber strands.
- Neoclouds are still learning – While hyperscalers understand networking, neoclouds often have little networking experience, although more mature neoclouds will choose a combination of transit, peering and private links depending on the business need.
For years, interconnection followed a simple rule: networks went where the other networks were. The story of how one of the biggest hubs started is as much about neutrality as proximity. MAE-East, the first non-governmental exchange point that spread across Northern Virginia, was originally owned and operated by companies that were “picking and choosing” who could join, says Stefan Raab, senior director of business development at Equinix, a data center and colocation provider. A group of internet service providers subsequently put out a request for a neutral facility, and the team that became Equinix won it, choosing to build the company’s first data center in Ashburn in 1998, transforming the area into a major interconnection hub. Raab says its distinguishing feature was letting customers bring their own cables straight into their space, unlike rivals that sent them through a shared “meet-me” room. Ashburn remains the best-known hub in the US, with some estimates suggesting the area handles as much as 70% of global internet traffic, but it isn’t an Equinix monopoly. Some of the world’s largest wholesale colocation providers, hyperscalers, and cloud platforms have a footprint in the area.
New hubs follow reach
Since then, interconnection hubs have opened in other US markets, and now AI is giving them new reasons to expand. DE-CIX, an internet exchange operator, manages exchanges in New York, Dallas, Chicago, Richmond, Houston and Phoenix. According to Dr. Thomas King, CTO at DE-CIX, Dallas and Phoenix are “already substantial markets”, with the company marking ten years at the Dallas site this year. Each new market brings its own logic, with Richmond connecting to subsea cables landing at Virginia Beach, while Houston gives networks across East Texas and southern Louisiana access to services through Dallas. In Wichita, Kansas, a new exchange at Wichita State University will allow local networks to peer locally and reach DE-CIX’s Dallas ecosystem.
Where data centers can be built has, to some degree, been shaped by power requirements, but AI changes the dynamic. “You can put a training facility somewhere relatively remote if you have sufficient power and can move the training data there,” says King. But with interactive inference, an AI agent might go back and forth several times between sites to complete one task, such that the milliseconds it might take to complete a task start to quickly add up. As a result, King expects large training campuses alongside more distributed inference, the latter closer to customers and connected through regional hubs.
For Raab, the map is broader. The big US peering markets, he says, are Ashburn, Chicago and Dallas, together with Seattle and Silicon Valley on the West Coast. Tier-two markets, meanwhile, are starting to grow as traffic distributes out. But data centers moving to new sites doesn’t necessarily mean interconnection follows. Ohio’s data center growth, for instance, did not produce major peering points, says Raab, with much of its traffic returning to Ashburn or Chicago. Growth in Ashburn’s network hub, he adds, is moving “from hyper growth to sustained growth” while compute moves further out.
What this suggests is that hubs often follow networks, rather than compute. Ashburn took off because carriers wanted a neutral place to meet, and Ohio got the data centers without the peering. That makes Wichita an interesting test. The State of Kansas awarded Connected Nation a $5 million grant to build the exchange, with DE-CIX as platform operator. Building a hub ahead of demand is a different bet from the one that created Ashburn, and whether networks come will show if hubs can be planned, not just discovered.
Carrying the load
Raab cautions against reading exchange numbers as traffic, as public sources such as PeeringDB are self-reported and show what has been built, not what is used. Public exchanges mainly serve the “long tail” of partners, he says, while cross-connects are for the high end of traffic. What tends to happen, he adds, is networks start on an exchange and move to a direct link once, for example, 80% of their traffic goes to one provider. Estimates around interconnect traffic are difficult to come by, but Raab guesses roughly 100 times more traffic crosses private fiber than public exchanges.
What is now a clear and recognizable shift is in scale. One cross-connect “isn’t enough anymore”, Raab says, with some customers starting a new facility at 30 terabits of bandwidth and pulling bundles of 288 fiber strands between them, lighting pairs as needed. Perhaps most critically, he affirms, “nothing replaces physical fiber,” in terms of its ability to deliver critical bandwidth capacity.
If Raab’s estimate is right, and private fiber does indeed carry about 100 times the traffic, exchanges aren’t where the volume sits, which fits a company that sells cross-connects. But his own logic cuts the other way for the long tail, as more participants means more traffic on public peering. Both can be true. The AI build-out looks to be stretching the head of the market, with 30-terabit starts and 288-strand bundles, and widening the tail, with neoclouds and inference needing many partners.
Learning curve
How companies approach interconnection needs depends on the nature of their business. Where hyperscalers understand internet networking well, many neoclouds, Raab says, have very little networking knowledge or experience and mostly buy direct internet access, essentially an enterprise internet connection, although he expects that to change as they scale and take more control of their networking infrastructure and peering. DE-CIX takes a measured view of the same shift. A neocloud selling inference needs to think about the whole journey from request to response, King says, with a “mature” neocloud usually employing a combination of IP transit, peering and private links depending on the business need.
There is, King says, no universal threshold where a particular number of GPUs, or a particular traffic volume, suddenly makes peering essential. He argues that neoclouds looking to move into enterprise services need to pay attention to connectivity from the outset, as that will help determine what performance they can offer their customer base. Hyperscalers, for their part, use exchanges to meet the networks carrying their customers’ traffic, King says. Google’s Verified Peering Provider programme is an example. DE-CIX holds Gold status, which means a connected network can reach Google’s public services, including Workspace and public Google Cloud resources, without qualifying for its own direct peering. That gives a smaller network a shortcut to a hyperscaler it might otherwise struggle to reach directly, although King is careful to distinguish it from private cloud connectivity, which links an enterprise to resources inside its own cloud environment.
Pressure points
As RCR noted in August, AI is moving the networking bottleneck outward. Having moved from the rack to the campus and then to the routes between campuses, where carriers are locking in fiber supply, it is now reaching the places where networks meet. For interconnection, the open question is the balance between public peering and user-centric interconnection. Raab expects neoclouds to take more control of their networking as they scale, which he says would add to public peering, while King expects distributed inference to push more interconnection towards users. Whether that happens, and whether regional hubs attract traffic as well as networks, will shape how well interconnect infrastructure copes with demand over the coming years. That said, the pressure may simply move further out. Raab estimates most internet traffic comes from caches, so the middle mile is built mainly to fill them. If even a small share of that shifts to real-time inference at the edge, that could pose “a crushing challenge for the middle mile”.
