Why Traditional Anycast Is Dead in 2026

Traditional anycast routing has served the internet well for two decades. But in 2026, its fundamental limitations have become impossible to architect around. It's time to have the honest conversation.

What Anycast Actually Does

For those unfamiliar: anycast is a network addressing method where the same IP address is announced from multiple geographic locations simultaneously. A client connecting to that IP is routed by BGP to whichever announcement its upstream provider considers "closest" — usually measured in AS hops, not actual latency.

This was a clever solution to a real problem. Before anycast, serving a global audience meant either a single origin (with terrible latency for distant users) or complex DNS-based geographic load balancing with its own pathologies. Anycast made CDNs, DNS resolvers, and DDoS scrubbing practical at scale.

The problem is that BGP's definition of "closest" hasn't meaningfully changed since RFC 4271 in 2006. The network has changed enormously. BGP's route selection criteria have not.

The Convergence Time Problem

When a link fails or degrades, BGP convergence — the process of all routers updating their routing tables — takes minutes to tens of minutes in the real world. During that window, anycast traffic is either blackholed, traversing a degraded path, or oscillating between multiple bad paths as different parts of the internet converge at different speeds.

For the January 29 SEA-ME-WE 6 cable event we described in our AI routing post, traditional anycast customers on the affected segment experienced elevated latency for 14–22 minutes. RouteKey's Neural Pathfinder had rerouted 96% of affected traffic within 900 milliseconds of the first signal. That's not a marginal improvement — it's a different class of reliability.

The Latency Topology Mismatch

BGP route selection is based on AS-path length, local preference, MED, and a handful of other attributes — none of which are actual latency measurements. A route with two AS hops through poorly-peered networks can be dramatically slower than a four-hop route over well-connected transit.

This mismatch is well understood in the networking community and has spawned various workarounds: traffic engineering with BGP communities, custom route reflectors, SD-WAN overlays, and so on. Each workaround is a tacit acknowledgment that BGP alone can't make latency-optimal decisions.

The hyperscalers solved this by building proprietary SDN control planes on top of BGP, treating BGP as a signaling mechanism rather than an actual routing decision engine. That approach works if you have Google's or Meta's infrastructure budget. For everyone else, the cost has historically been prohibitive.

What Comes After Anycast

We're not arguing that anycast should be abandoned overnight. For certain workloads — particularly DNS and DDoS absorption — it remains the right tool. The packet-level anycast behavior still delivers useful geographic steering even if the convergence properties are poor.

What we're arguing is that anycast should not be the decision layer. It should be a fallback. The real routing intelligence should live in an overlay that can:

• Measure actual latency between every PoP pair continuously, not infer it from topology
• Detect path degradation in sub-second timeframes, not minutes
• Reroute individual flows without waiting for global BGP convergence
• Learn from application-layer signals, not just network-layer metrics

That's what RouteKey's architecture delivers. BGP is still running underneath — we haven't invented a new internet — but it's not making the decisions that matter.

The Migration Path

The good news is that migrating away from anycast-first architectures is increasingly straightforward. RouteKey functions as a transparent overlay: you keep your existing anycast announcements, add RouteKey as your routing layer, and traffic gradually shifts to the optimal path as the system learns your topology.

Most customers see measurable latency improvements within 48 hours of onboarding, without any changes to their origin infrastructure. The anycast remains as a fallback — useful precisely because it doesn't require any coordination — while the intelligent layer handles the workloads that actually matter.

The era of "BGP is close enough" ended sometime in 2024. The tooling to replace it is now mature, affordable, and running at scale. There's no longer a good reason to route production traffic the way we did in 2004.