What would an AI Peering Exchange look like?
At the Global Peering Forum 2026 in Panama City, Matt Griswold joined a panel asking "Is an AI Peering Exchange Possible?"
Josh Snowhorn of Quantum Loophole moderated the discussion, with Thomas King of DE-CIX, Richard Petrie of LINX, Jeff Budney of Verizon and Matt "Grizz" Griswold of Open Pipes taking part.
This isn't a recap of their discussion or an account of the panellists' views. The title itself got me thinking: what do we actually mean by an AI Peering Exchange?
The phrase could easily sound like an Internet Exchange operated by AI. But perhaps the more useful interpretation is an exchange designed to connect the organisations and infrastructure involved in delivering AI services.
That might include cloud and GPU providers, model platforms, data centres, research networks, content providers, enterprises and the networks carrying traffic between them.
New exchange, or evolution?
Many of the basic requirements sound familiar: high-capacity connections, resilience, predictable performance and the ability to connect directly with the networks and infrastructure they need.
Internet Exchanges already provide neutral places where different networks can connect and exchange traffic. So the question may not be whether we need a completely new type of exchange, but whether AI creates different participants, traffic patterns or services that existing exchanges need to support.
An AI-focused exchange might need to make it easier to connect to specialist compute infrastructure or move very large datasets between particular locations. It might offer more automation, faster provisioning or private connectivity alongside traditional public peering.
There is another question too: at what layer would an AI exchange actually add value? Traditional Internet Exchanges provide the underlying connectivity and routing environment, while many of the activities associated with AI — accessing data, interacting with APIs and crawling web content, for example — happen higher up the stack. If the challenges are at the application layer rather than the network layer, would a different kind of exchange help solve them, or would the role of an IXP remain largely the same?
Perhaps that is the real distinction to explore: does AI require something fundamentally different from an Internet Exchange, or is this simply another evolution in the networks, services and communities that IXPs support?
What are we actually trying to solve?
For me, the interesting question is where an AI exchange could make a difference.
Who needs to connect, and what is difficult about doing that today? Is it capacity, location, cost or access to particular infrastructure? Could existing IXPs meet those needs, or is there something missing?
At Open Pipes, we're interested in exploring how shared and community-focused infrastructure might help test ideas like these. The answer may look different in each market, and we should avoid creating a new label before we understand the need behind it.
Perhaps the starting point isn't whether an AI Peering Exchange is possible, but whether the organisations building and using AI infrastructure need something from interconnection that they cannot easily get today.
What would an AI Peering Exchange mean to you?