Internet exchange points are becoming the new chokepoint for AI data center traffic
The internet’s traffic jams used to form at submarine cable landing stations and tier-1 backbone links. In 2026, the new chokepoint is the internet exchange point (IXP) — the physical facilities, like DE-CIX Frankfurt or Equinix’s Ashburn campus, where hundreds of networks plug into shared switching fabric to exchange traffic directly instead of paying a transit provider to carry it. AI data center buildouts are now pushing egress and inter-region replication traffic through these facilities at volumes the peering model was never designed to absorb cheaply, and the result is a quiet but real capacity and pricing squeeze playing out across major exchanges.
The core issue is that large language model training and inference were built around massive east-west traffic inside a single data center campus, but inference serving increasingly isn’t. A hyperscaler training a frontier model in one region and serving inference to users worldwide needs to move model weights, embeddings, and request/response traffic across metro and even continental links, and a growing share of that traffic now transits IXPs rather than staying inside a single provider’s private backbone, because even the biggest clouds don’t have dedicated fiber to every market they serve.
Why Peering Economics Break Under AI Load
Traditional peering works because traffic between two networks at an IXP is roughly symmetric — a consumer ISP and a content network exchange video and web traffic in volumes that balance out over time, so neither side bills the other for transit. AI inference traffic breaks that symmetry. A network serving inference responses to millions of API calls sends far more data out than it receives in requests, and a growing number of networks are now seeing inference-driven egress that looks structurally like a CDN’s traffic profile, but without the decade of negotiated settlement-free peering agreements that CDNs built up specifically to handle that imbalance.
DE-CIX and AMS-IX have both reported record peak traffic repeatedly through 2025 and into 2026, and both have cited AI and cloud-to-cloud traffic as a specific driver of growth beyond the video-streaming baseline that dominated the previous decade. The practical consequence for the networks sitting at these exchanges is that port capacity — the physical 100G, 400G, or now 800G ports that connect a network to the exchange fabric — is filling up faster than exchanges can provision new switching capacity, and upgrade lead times for the newest port speeds routinely run months because the optics and switching hardware are themselves supply-constrained by the same AI buildout driving the demand.
The Private Interconnect Workaround
Hyperscalers have responded by increasingly bypassing public peering fabric altogether in favor of private network interconnects (PNIs) — dedicated point-to-point links directly between two networks’ routers, colocated in the same facility but not touching the shared exchange switch. A PNI avoids the congestion of a shared fabric and gives both sides a dedicated, billable, capacity-guaranteed link, which is exactly what you want when a single AI workload can saturate a shared port on its own. The tradeoff is that PNIs don’t scale the way public peering does: every new PNI is a separate physical cross-connect and a separate negotiated agreement, so a network that wants dedicated links to fifty AI-heavy counterparties needs fifty separate interconnects rather than one shared fabric port.
This is reshaping who has leverage in the interconnection market. Facilities operators like Equinix and Digital Realty, who sell the physical colocation space where both public peering and private cross-connects happen, are seeing demand shift toward higher-margin cross-connect and PNI revenue rather than public peering port fees, and smaller regional IXPs that can’t offer the same density of cloud and AI-compute tenants are losing relevance to the hyperscaler-dense flagship facilities in Ashburn, Frankfurt, and Singapore.
What This Means If You Run Infrastructure
For network engineers planning capacity, the actionable signal is that port-speed upgrade lead times at major exchanges should now be budgeted in quarters, not weeks, and multi-homing strategies that assumed cheap public peering as a fallback need a second look if a meaningful share of traffic is inference-driven and asymmetric. For anyone selecting a cloud region or colocation facility for a latency-sensitive AI workload, proximity to a tier-1 IXP with available port capacity is now a real selection criterion, not just a nice-to-have, because the alternative is discovering mid-contract that the exchange you’re connected to has no spare capacity at the port speed you need.
The broader pattern worth watching is whether IXPs start explicitly pricing for traffic asymmetry the way CDNs eventually had to, rather than relying on the settlement-free norms built for a more balanced traffic era. If that shift happens, AI companies currently treating interconnection as a fixed, negligible cost will see it become a line item worth negotiating directly, the same way cloud egress fees became one five years ago.